Machine learning based remote sensing

Abstract

System (SYS, DS) and related method for processing sonar imagery. The system comprises an input interface (IN) for receiving sonar imagery obtainable by a sonar device (SD) in a body of water. An in- image search component (ISC) capable of searching, based on the sonar imagery, for a candidate in- image location in the sonar imagery. The candidate in-image location, when found, is potentially representative of a respective discharge event of fluid dischargeable into the body of water. A characterizer component (CC) is capable of processing sonar image values of the sonar imagery at the found candidate in-image location into a characterization result that characterises the discharge event. An output interface (OUT) provides the characterization result for processing such as for navigation, or other. The system may be used for underwater monitoring for discharge of CO2, or oil, or other fluids, into ambient waters, such as ocean, sea, lake, river system, etc.

Description

[0001] MACHINE LEARNING BASED REMOTE SENSING

[0002] Cross-Reference to Related Applications

[0003] This Patent Application claims priority from Italian Patent Application No. 102024000021254 filed on September 24, 2024, the entire disclosure of which is incorporated herein by reference.

[0004] Technical Field

[0005] The invention relates to a System for processing sonar imagery, to a related method, to an arrangement, in particular a remote sensing arrangement, including such a system, a training system of training a machine learning model on training data for use in such a system, a related method, a training data provider system for providing training data of training such a model, to a related method, to a use of horizontal field of view sonar for acquiring sonar data, to a computer program element, to a computer readable medium, to a use of a sonar device.

[0006] Background of the Invention

[0007] The IPCC (Intergovernmental Panel on Climate Change), a United Nations body, states in their 2023 synthesis report (doi: 10.59327 / IPCC / AR6-9789291691647.001) that emissions of greenhouse gases caused global warming, with global surface temperature reaching 1.1 °C above preindustrial levels. This has adverse impacts on ecosystems, human settlements, agriculture, etc.

[0008] Various climate change mitigation measures are being pursued, including decarbonization pathways, put in place by the Oil & Gas sector, and others. One such pathway is CCS (carbon capture and storage). In CCS and related technologies, a greenhouse gas such as CO2 is captured at source or from atmosphere, with subterranean injection, thus removing the gas from contributing to the greenhouse effect. In great many of cases, the gas is injected into geological reservoirs beneath the seabed, such as in depleted oil fields. Thus, such sub-seabed reservoirs are intended for permanent storage of the damaging gas. It has been estimated that CCS may help achieve an up to 13% reduction in global emissions. See of example A. Lichtschlag et al, in “Suitability analysis and revised strategies for marine environmental carbon capture and storage (CCS) monitoring," International Journal of Greenhouse Gas Control, vol. 112, Dec. 2021 .

[0009] However, CCS is only as good as it holds the gas safe, which is not always the case. Due to geological activities and because of inherent structures of the seabed at places, some gas may escape and discharges back into the surrounding wates and from there into the atmosphere.

[0010] Underwater monitoring for such leaks is a challenge. Gas may be discharged into water at unexpected places, over a huge area. Whilst some underwater remote sensing modalities, in particular sonar, could be deployed for such monitoring task, such endeavours are fraught with challenges: fixing such leaks may involve deployment of considerable expenses, and this combines unfavourably with current sonar-based leak detection efforts that often produce false positive results (that is, they indicate there is a leak, although, in reality, there isn’t any).

[0011] Another challenge is posed by the sheer volume of sonar acquired data that needs analysis. Hitherto, this is time consuming. The mentioned underwater monitoring for leaks is also of interest in other fields of endeavour, such as asset integrity monitoring, in particular in Oil & Gas, environmental monitoring & protection, underwater exploration, and others.

[0012] Summary of the Invention

[0013] Thus, there may be a need for improved remote sensing for detection of underwater discharge events where fluid discharges into a body of water.

[0014] An object of the present invention is achieved by the subject matter of the independent claims where further embodiments are incorporated in the dependent claims. It should be noted that the following described aspect of the invention equally applies to the to the said related method, to the arrangement, in particular to the remote sensing arrangement, including such as system, to the training system of training the machine learning model on training data for use in such a system, to the related method, to the training data provider system for providing training data of training such a model, to the related method, to the computer program element, to the computer readable medium, and to the use of the sonar device.

[0015] According to a first aspect of the invention there is provided a system for processing sonar imagery, comprising: input interface for receiving sonar imagery obtainable by a sonar device in a body of water; an in-image search component capable of searching, based on the sonar imagery, for at least one candidate in-image location in the sonar imagery, the at least one such candidate in-image location, when found, potentially representative of a respective discharge event of fluid dischargeable into the body of water; a characterizer component capable of processing sonar image values of the sonar imagery at the at least one found candidate in-image location into a characterization result that characterises the discharge event, and an output interface for providing the characterization result.

[0016] In embodiments, the characterizer component is based on a trained machine learning model (M).

[0017] In embodiments, the sonar imagery comprises a time series of frames, wherein the in-image search component, in searching for the at least one candidate in-image location, is capable of over-time processing of the frames to obtain the at least one candidate in-image location as one that is indicative of over-time changes of image values at the said at least one location.

[0018] The over-times changes may manifest as flickering signal across frames, caused by bubbles (or liquid) that pass through the sonar’s FOV, in particular the sonar’s scanning plane, which may be preferably horizontal.

[0019] In embodiments, the said over-time processing includes applying a moving window variance filter.

[0020] In embodiments, the said over-time changes of image values manifest as a pulsation.

[0021] In embodiments, the sonar imagery comprises a time series of frames, and wherein the in-image search component is capable of spatio-and / or-temporal analysis of the said frames in obtaining the at least one candidate in-image location.

[0022] In embodiments, the in-image search component is capable of searching for multiple such candidate in-image locations in multiple instances of sonar imagery acquirable at different times and as receivable at the input interface, wherein the in-image search component is capable of tracking an evolution of the said candidate in-image locations over time, and selecting, based on such tracking, a selection of one of more such candidate in-image locations for processing by the characterizer component. The tracking of the respective candidate in-image location may be based on one or more descriptors that describe spatio and / or temporal aspect(s) of respective subset of pixels (in general, a cluster) at the respective candidate in-image location.

[0023] In embodiments, the system comprises a geolocator facility capable of associating the at least one candidate in-image location and / or the respective provided characterization result with a respective geographical location of the respective discharge event.

[0024] In embodiments, the system includes navigation control circuitry capable of causing a watercraft, carrying the sonar device, to navigate to the geographical location.

[0025] In embodiments, the sonar device is capable of obtaining the sonar imagery based on a speed of sound measurement providable by a speed of sound probe co-locatable with the sonar device in the said body of water.

[0026] In embodiments, at least a part of the system is arrangeable as a processing unit, interfaceable with the sonar device.

[0027] In embodiments, the processing unit is mountable i) in the, or a watercraft that is capable of moving the sonar, during its operation, in the said body of water, or ii) in or at a fixed underwater structure. In embodiments, the watercraft is any one of a vessel or submersible.

[0028] In embodiments, the watercraft is an autonomous underwater vehicle, AUV, or a remotely operated vehicle, ROV.

[0029] In embodiments, a field of view, FOV, of the sonar device is forward relative to a motion of the sonar device, and parallel to an average profile of a floor of the body of water. In embodiments, the FOV, of the sonar device is forward relative to the motion of the sonar device, and wherein a plane of the field of view is substantially horizontal. Thus, the plane is such that is it approximately perpendicular to the upwards propagation direction of the plume of bubbles that for example, that form the discharge event. The thickness of the plane (the area insonified by the sonar) in the propagation direction (generally vertical, parallel to gravity field lines, upwards towards water surface) is negligible. The sonar device is preferably a high frequency, HF, sonar device, with operational acoustic pulse frequencies of about 900 kHz or higher.

[0030] In embodiments, the system comprises a visualizer for visualizing on a display device any one of: the characterization result and the candidate in-image location.

[0031] In embodiments, the characterization result is any one of: i) type of fluid that is released in the discharge, ii) a quantity of the / a fluid released in the discharge.

[0032] In embodiments, the discharge event is a leak event.

[0033] In embodiments, the fluid is a liquid other than water, or is a gas.

[0034] In embodiments, the discharge event includes bubbles formed in the body of water, the said bubbles being in a field of view of the sonar device.

[0035] In embodiments, the sonar device is a high frequency, HF, sonar device.

[0036] In embodiments, the sonar device is of the active type.

[0037] In another aspect there is provided an arrangement including the system of any one of the preceding claims, and further including any one or more of: i) at least a part of the sonar device, ii) a display device on which is displayable the characterization result and / or the candidate location, iii) a craft in or on or which the sonar device is mounted for moving same in the body of water, iv) interface circuitry through which the system and any one sonar device and craft are couplable, v) a power source through which powerable any one of: the system, the sonar device, the craft.

[0038] In embodiments, the system is at least partly integrated in the sonar device and / or in the craft. In another aspect there is provided a training system for training, based on training data, the machine learning model.

[0039] In another aspect there is provided a system of providing training data fortraining the machine learning model.

[0040] In another aspect there is provided a method of processing sonar imagery, comprising: receiving (S1330) sonar imagery obtainable by a sonar device in a body of water; searching (S1340), based on the sonar imagery, for at least one candidate in-image location in the sonar imagery, the at least one such candidate in-image location, when found, potentially representative of a respective discharge event of fluid dischargeable into the body of water; processing (S1350) sonar image values of the sonar imagery at the at least one found candidate in-image location into a characterization result that characterises the discharge event, and providing (S1360) the characterization result.

[0041] In another aspect there is provided a method of training, based on training data, the machine learning model.

[0042] In another aspect there is provided a method of providing training data fortraining the machine learning model.

[0043] In another aspect there is provided a computer program element, which, when being executed by at least one processing unit, is adapted to cause a processing unit or system to perform any one of the said methods.

[0044] In another aspect there is provided at least one computer readable medium having stored thereon the program element, and / or having stored thereon the trained machine learning model, in particular its trained parameters.

[0045] In another aspect there is provided a use of a higher frequency sonar device having a horizontal imaging geometry for collecting sensor imagery for processing into characterization of a discharge event.

[0046] In another aspect there is provided a sonar device capable of operation as forward-looking, and capable of causing a planar insonified area to be formed underwater that has a plane, so that fluid dischargeable into the ambient water propagates through the water and perpendicularly through the said plane whilst sonar imagery is acquirable for processing into a characterization result to characterize the discharge event. In embodiments the said plane in horizontal.

[0047] In preferred embodiments, the sonar device is a HF sonar device. In another aspect there is provided a method of operating a forward-looking sonar device, comprising: controlling the sonar device to cause a planar insonified area to be formed underwater that has a plane, so that fluid dischargeable into the ambient water propagates through the water and perpendicularly through the said plane; acquiring sonar imagery with said sonar whilst the said fluid is so propagating, and providing the acquired sonar imagery for processing into a characterization result to characterize the discharge event.

[0048] The processing may include filtering to isolate acoustic backscatter signatures in the acquired imagery, and passing image data from the sonar imagery that is associable with the said signatures for said characterization into a trained machine learning model.

[0049] The sonar imagery is spatial data. It may be raw data as acquired by the sonar device. It may be conditioned, amplified, etc. Sonar imagery may refer to the acoustic backscatter imagery obtained by operating sonar device. This involves capturing a horizontal portion of water, with a wide field of view horizontally and narrow vertically. This raw data may be combined with other data such as time of acquisition, navigation and environmental parameters for complete and accurate processing. The navigation data may pertain to the geographical location of the discharge event that caused signal to be represented the sonar imagery that the system was able to identify (to “find”), upon the proposed processing, as the said candidate location(s). More than one candidate locations may be so found.

[0050] The system is capable of searching of hitherto unknown discharge events and characterize same. It can do this efficiently thanks to a 2-stage setup: first there is search in image domain for promising candidate locations. Only image pixels from those candidates are processed by the characterizer in the 2ndstage. This allows for better results and responsiveness. Characterizer needs only process subset, not whole of sonar imagery, thus the system is highly responsive and capable of real time processing. This is useful in particular in sonar, where the acquired imagery can be very large, with a pixel density of 1 pixel per centimetre.

[0051] In candidate location(s) is in general not a point (pixel, but multiple pixels forming a portion / region (a subset) of the whole sonar image. Some or each such subset per candidate location may form a cluster of pixels. The localization of the event therefore corresponds to a portion of the sonar image, which thanks to navigation compensation (which allows comparing images taken at different times relative to the geographical position in the inertial system in the context of the mobile variance) can be precisely localized in a specific geographic location.

[0052] The characterization result might include details such as fluid type, release amount, and may be enriched with associated geographic location of the discharge event, etc. The proposed system and method may be understood as plume search methodology in a body of water (oceans, sea, lakes, rivers, etc), preferably based on forward looking horizontal sonar data, as opposed to vertical sonar data such as may be used traditionally in water column inspection (“WCI”). The use of such horizontally oriented sonar has an impact on the proactivity of the investigation and the accuracy of the detection. The proposed system is capable of real-time, pro-active search, with low false positive rates, as compared with current methodology. The computing capability is preferably integrated into the watercraft or fixed station that carries the sonar device. The system implements a cascaded data processing chain for processing the acquired sonar imagery, with gradual narrowing down of an initial set of candidate locations. The so found candidate location(s) may then be used for navigation to the physical geographical site that corresponds to a given such candidate location(s), and / or may be used by the characterizer component for remote characterization in terms of type of fluid (other than water) that is likely to have been discharged into the waters, and / or how much of such fluid was / is so discharged. Such cascaded narrowing down reduces the risk of futile remedial efforts being spent on false positive detections.

[0053] In addition, improved performance (robustness, accuracy) of the characterizer component has been observed, which, it is thought, is partly due to the prior processing of the sonar imagery by the inimage search component into the candidate location, and the feeding of data in relation to candidate locations found by the in-image search component into the, preferably ML based, characterizer component. Having the characterizer component produce accurate, reliable, characterization results allows better judging remotely, whether it is worthwhile actually traveling to the corresponding site, such as with a vessel or other watercraft (such as ROV, AUV, submarine, or whichever) for example, and perform remedial work (such as leak fixing) there, or perform further analysis on-site, etc.

[0054] The proposed system is proactive in that, thanks to at least in parts of real-time processing and the forward-looking sonar imaging geometry, the detection can be done before passage of the sonar carrying craft through the (partly) insonified forwardly located portion of the water body.

[0055] By leveraging a forward-looking horizontal sonar, the proposed system offers a targeted and optimized approach to underwater monitoring, significantly improving detection accuracy and operational efficiency. Thus, in preferred embodiments, horizontal sonar is used, using high-frequency sonar, which has a thin vertical aperture. This means that the observation is not volumetric as with traditional forward-looking sonars or acoustic cameras, but is instead two-dimensional, focusing on a cross- sectional view of the bubble plume. The orientation of the sonar fan is in most cases horizontal. What is of benefit herein is that the (mean) propagation vector of the fluid, eg a plume of bubbles, is perpendicular to the plane defined by the sonar fan. As said, in most cases, such orientation will be horizontal. The horizontal sonar fan is positioned in such a way that the mean propagation vector (in general parallel to gravity field lines) of the bubbles' plume is perpendicular to this plane. This orientation ensures the detection system DS can effectively observe a pulsating backscatter signature of the bubbles as they travel upwards is has been found herein. However, whilst use of such horizontal sonar beam / FOV sonar device is preferred, the proposed setup is not necessarily tied to such horizontal sonar imaging geometry. Other sonar imaging geometries may be used instead.

[0056] Whilst mounting of the sonar is envisaged mainly in submersibles, such as ROVs. and AUVs and others, such use is not at the exclusion of other uses such as mounting the sonar device on the hull of the support vessel for example where the propulsion is then provided by the vessel itself. However, even mounting sonar on a fixed station on the sea floor for example is also envisaged herein, for continuous discharge event monitoring of a fixed area, as opposed to when using the proposed system in an AUV that is free to roam the waters to pick up discharge events clues wherever the event falls within the sonar’s field of view. Such fixed station may be placed in proximity to a known or critical point for long-term monitoring.

[0057] The proposed system may be installed in an underwater vehicle or fixed station, developed for the automatic and real-time detection, in the field of view of a high-frequency sonar, of fluid discharges (in-water emission) with a different density to water, or gaseous discharges. The system consists of one or more high-frequency sonars with a wide field of view to the front and extended substantially parallel to the seabed over which the sonar carrying watercraft is travelling during sonar image acquisition. Sonar with a frequency of over 900 kHz is preferred herein, as such allow identification of even minute density variations, with a spatial and temporal resolution that achieves location accuracy to the nearest centimetre. The use of high-frequency sonar obtaining high-resolution sonar / acoustic imagery of acoustic signatures associated with the discharge event being investigated, thus allowing the extraction of specific properties associated with the discharge event.

[0058] The in-image search component, which is preferably part of the overall detection system, receives the acoustic images generated by the sonar. The processing unit analyses a sequence of acoustic images using computer vision algorithms, thus automatically identifying the responses associated with the passage of liquid or gaseous fluids in the sonar's field of view. If more than one sonar is present, the unit merges the images in order to obtain a single image that extends the field of view of an individual sonar. Sonar image processing allows a fluid discharge event to be detected and located using the system's geolocation information. The event is located by analysing the acoustic image, obtained by interpreting the time delay of the echoes in relation to the velocity of sound in water. The latter is a factor of consideration in calculating the exact distance of the event from sonar and is therefore preferably precisely measured using a dedicated probe. This can be coupled directly to the data processing unit or, alternatively, the sound velocity data can be acquired from the craft’s host system. The data may be provided as a data string in NMEA (National Marine Electronics Association) format, or any other format. This approach offers flexibility and ease of integration with a variety of platforms, ensuring that the system can operate favourably in different operational contexts. The presence of the probe for directly measuring the velocity of sound in water ensures that the system can adapt to changes in the environment, guaranteeing reliable results regardless of marine conditions. The combination of this data, possibly together with navigation data as provided by the watercraft, allows the underwater discharge event to be located accurately, minimizing estimation errors to a few centimetres over a range of tens of meters. Geolocation may be achieved using positioning information in the case of a fixed station or navigation information, including position and attitude, provided by the propelled underwater craft’s inertial navigation unit.

[0059] The in-image search component is preferably designed to be coupled via the interface to a data network of the underwater vehicle or fixed station and to efficiently analyse acoustic imagery, taking preferably advantage of the computing power of a graphics processor, in particular to achieve (quasi- ) real time processing. In some embodiments, in order to further narrow-down the found candidate locations, and to thereby reduce false positive detection errors, the in-image candidate location is tracked as it evolves over time, thus allowing for the discrimination of detected events based on a number of characteristics, such as speed and shape, or other descriptor(s).

[0060] To ensure effective system integration with both underwater craft or fixed station, the data interfacing may rely on a simple, but efficient architecture. Preferably, this is achieved by integration of the sonar and detection system, both of which may be interfaced to the watercraft or station via respective ports, preferably standardized, with the same technical interface. This configuration may not only facilitate power supply to the system, but also facilitates ease of connection to the host system’s data network.

[0061] For installations on a fixed station, the sonar may be movable by a panning unit (e.g., pan & tilt), which, by rotating on a vertical axis, gives up to a 360° view of the monitoring area surrounding the station.

[0062] The system, once activated, may perform regular scans, capturing a series of acoustic images. By analysing time series of sonar images (frames), it may determine the evolution of the detected emission, providing clear indications of the expansion or contraction of the identified spot. This continuous monitoring process makes it possible to observe the evolution of the emission over time and act promptly in the event of significant changes.

[0063] If installed on an underwater craft, the precise location of the identified event may be communicated to the watercraft, enabling it, if autonomous, to recalculate the navigation course and automatically head towards the event. This allows the vehicle to effectively collect all the data needed for the complete characterization of the liquid or gaseous emission. Specially, when mounted on an AUV or ROV, the system can be navigated through the water to help identify and investigate potential discharge events, making it particularly suited for exploratory and dynamic monitoring. Fixed stations, on the other hand, provide continuous and long-term monitoring of specific areas, such as known emission sites or critical points / areas, ensuring ongoing surveillance and data collection in critical locations.

[0064] Once the discharge event is detected and located, in the course of monitoring, the system proceeds to the quantification and classification stage by exploiting the raw acoustic backscatter data in the area of the identified candidate location(s), possibly suitably narrowed down to a single such location or a smaller plurality is initially found. This methodology is based on the principle that acoustic backscatter, or the reflection of the sound emitted by the high-frequency sonar, contains valuable information about the nature and quantity of the emitted fluid.

[0065] The proposed approach uses machine learning by the characterizes component, to analyse acoustic signature(s) as recorded in the acoustic images corresponding to candidate location / event, allowing for discrimination of type of fluid discharged, and / or estimation of amount of fluid discharged. That there is, as a matter of theoretical principle, a relation between acoustic backscatter and different fluids is suspected herein, and pertinent scientific literature appear to include indications that this may indeed be so, such as in Phelps and Leighton in their paper “High-resolution bubble sizing through detection of the subharmonic response with a two-frequency excitation technique", published in J. Acoust. Soc. Am, vol 99 (4), Pt. 1 , April 1996, pp 1985-1992.

[0066] This proposed ML based characterization (classification and / or quantification) opens up new perspectives in underwater environmental monitoring, allowing the nature of emissions to be accurately identified and their impact assessed.

[0067] The application of ML in this context in the interpretation complex patterns / signatures of acoustic backscatter signals, allows correlating specific acoustic response profiles with determined physical emission characteristics, the proposed setup allows quantitative estimation of the amount of fluid emitted, thus providing critical data for environmental assessment and mitigation planning.

[0068] The proposed setup has a wide range of uses in various applications, offering advanced solutions for the detection, location and tracking of liquid and gaseous underwater discharges. Some of the main fields of application envisaged herein includes oil & gas, environmental monitoring, CCS monitoring, underwater exploration, or others still.

[0069] For example, the proposed system may be used in oil and gas industry, enabling occasional or continuous monitoring of emissions during underwater operations, relating to the extraction, production, and transportation of hydrocarbons. The early identification of leaks or spills contributes significantly to environmental and operational safety.

[0070] In the field of environmental research and monitoring, the proposed system may prove a valuable tool for monitoring underwater areas. The ability to detect and track liquid or gaseous discharges enables rapid responses to potential environmental impacts, facilitating preventive interventions and mitigation measures.

[0071] The proposed setup is of particular benefit for monitoring areas dedicated to the offshore storage of carbon dioxide, eg in the context of CCS. The ability to detect in real time and locate any gas or liquid leaks from an underwater infrastructure or storage reservoir associated with CCS projects improves the operational effectiveness and safety of these facilities. It is of benefit that discharges are tracked and identified accurately to safeguard integrity of the carbon storage process, and to avoid or at least mitigate undesirable environmental impacts.

[0072] As to underwater explorations, the proposed system or method can be used in underwater exploration to detect and study natural or anthropogenic phenomena. The ability to detect and monitor underwater fluids helps advance scientific knowledge and marine exploration activities.

[0073] In general, the term “machine learning” includes a computerized arrangement (or module) that implements a machine learning (“ML”) algorithm. Some such ML algorithms operate to adjust a machine learning model, thus configuring same to perform (“learn”) a task. Other ML operate direct on training data, not necessarily using such an explicit model, thus in such cases the training data may form the model, or may be part of such model. This adjusting or updating of the model is called “training”. In general, task performance by the ML model may improve measurably with training experience. Training experience may include suitable training data, and exposure of the model to such data. Task performance may improve, the better the data represents the task to be learned. Training experience helps improve performance if the training data well represents a distribution of examples over which the final system performance is measured. See for example, T. M. Mitchell, “Machine Learning”, page 2, section 1.1 , page 5-6, section 1.2.1 , McGraw-Hill, 1997. The performance may be measured by objective tests based on output produced by the model in response to feeding the model with test data. The performance may be defined in terms of a certain error rate to be achieved for the given test data.

[0074] Brief Description of the Drawings

[0075] Exemplary embodiments of the invention will now be described with reference to the following drawings, which, unless stated otherwise, are not to scale, wherein:

[0076] Figure 1 shows a block diagram of a remote sensing arrangement for underwater environments as envisaged herein;

[0077] Figure 2 shows details of the remote sensing arrangement including a submersible watercraft that carries a remote sensing device, such as a sonar device;

[0078] Figure 3 indicates an imaging geometry of a sonar device as may be used herein in embodiments;

[0079] Figure 4 shows a block diagram of a detection system that may be used herein to detect an underwater discharge event;

[0080] Figure 5 shows a more detailed block diagram of the detection system of Figure 4;

[0081] Figure 6 shows a block diagram of an acoustic image processor as may be used in some embodiments of the detection system as envisaged herein;

[0082] Figure 7 illustrates acoustic imagery as processed by the detection system as envisaged herein; Figure 8 illustrates further details of acoustic imagery as processed by the detection system as envisaged herein;

[0083] Figure 9 shows a block diagram of a feature extractor of the detection system according to some embodiments;

[0084] Figure 10 shows a flow chart illustrating operation of an event tracker as may be used herein in embodiments of the detection system according to some embodiments;

[0085] Figure 11 shows a block diagram of a characterizer component as may be used herein in embodiments of the detection system;

[0086] Figure 12 illustrates acoustic patterns as may be found by the proposed detection system;

[0087] Figure 13 shows a flow chart of a method of processing sonar imagery for detection of a discharge event;

[0088] Figure 14 shows a block diagram of an architecture of a machine learning model as may be used herein;

[0089] Figure 15 shows a block diagram of a training system as may be used herein to train a machine learning model based on training data;

[0090] Figure 16 shows a flow chart of a method for generating training data;

[0091] Figure 17 shows a flow chart of a method of training a machine learning model; and

[0092] Figure 18 shows a block diagram of a training data provider system as may be used herein in embodiments to provide training data for training a machine learning model.

[0093] Detailed Description of the Invention

[0094] Reference is now made first to the diagram of Figure 1 which illustrates components of a remote sensing arrangement RSA, which may be used in an aquatic, in particular in a marine, setting.

[0095] The remote sensing arrangement RSA may comprise remote sensing device RSD such as a sonar device SD that operates underwater. The (one or more) sonar device SD may be mounted on or in a vehicle V such as a watercraft CR. That watercraft CR may be a submersible which may be unmanned or manned, or it may be pole-mounted or mounted to hull of vessel V, etc. Unmanned watercraft CR are preferred, however. Examples of such watercraft CR which may carry the sensor device SD include any one of a ROV (remotely operated vehicle), or an autonomous underwater vehicle (“AUV”), or any other vehicle capable of underwater operation. Mounting such sonar device SD on or at a stationary underwater station UWS is also envisaged in alternative embodiments, to serve different purposes such as long-term monitoring, etc.

[0096] Broadly, the sensor device SD is capable of collecting data d in relation to a discharge event DE. The discharge event DE may be a leak event (“leakage”) where fluid, such as gas or liquid is released from a reservoir RSV into a body of water BW. For example, the released fluid may be methane, CO2 or crude oil, or any other substance. In general, the discharge event DE may be undesirable, and may happen anywhere in a large area underwater. Thus, whether or not there is such a leak detention event is as such unknown and the remote sensing arrangement envisaged herein is capable of detecting whether there is such a discharge event, so that remedial action may be taken to stop the leakage.

[0097] The sonar device SD may be of the high frequency type, capable of operating at 900 kHz or higher. A transducer of the sonar device is operable to send out, as an interrogating signal IS, one or more acoustic wave pulses that propagates through the water, interacts with matter other than water, and is then bounced back as a backscatter signal RS which is then detected at a receiver of the sonar device SD. In this manner, acoustic backscattering signals are collected by sonar device SD. The collected acoustic backscattering signals form sonar imagery. As will be explained in more detail herein, the sonar imagery may be processable by a computing system SYS to so detect, based on the received sonar imagery, whether or not there is such a detection event DE within the sonar SD’s range.

[0098] In addition to a mere detection of such an event DE, which may be of interest in its own right, the system SYS may be further capable of localization of the detected event DE, such as in terms of geographical co-ordinates in 3D space, or other. Thus, the localization may yield a location in 3D space of this discharge event DE. This location may be relative to a current location of the sonar device SD. In some embodiments preferred herein, the received sonar imagery is further analysed by computing system SYS to characterize the discharge event DE. Such characterization may include determining the amount of fluid that escaped over a given period of time or in terms of flow rate, or the type of fluid that so escaped in the discharge event. Thus, the system may be capable of processing the received sonar imagery into data d indicative of any one or more of i) detection, iii) localization and iii) characterization of discharge event DE. For simplicity and brevity, the computing system SYS operable to process the sonar imagery into such data d for any one or more of i)- iii) will be referred herein as the detection system DS, with the understanding that in addition to the mere detection of discharge event DE, it is also localization and / or said characterization that is intended, and the system is so configured. Having said that, in some simpler settings or application scenario, a mere detection of event DE may suffice. Preferably however, in addition to such detection, system SYS, DS further delivers localization and / or characterization. In preferred embodiments, triadic processing is delivered, including all three of detection, localization and characterization. Also, and in the same vein, reference will be made herein on occasion to “detection data” d computable by detection system DS, with the understanding that such detection data d may be enriched detection data, not only indicative of detection of event DE, but also to any one or both of its localization and characterization.

[0099] The proposed remote sensing arrangement RSA may be used in coastal waters CA, in estuaries, in open oceans, seas, in river systems and lakes, as the need may be. The discharge event may be one where CO2 is escaping from an offshore carbon storage site, with such site forming the said reservoir RSV. Such may be relevant in carbon-capture storage (“CCS”) contexts. In CCS, captured CO2gas is injected into reservoir RSV under the seabed SB, usually a depleted oil or gas field for example, or other porous sub-seabed rock, which is intended for permanent CO2 storage. However, such intention for permanent storage may be undermined by adverse phenomena, such as seismic activity or other. Cracks may emerge in the seabed SB that allows unintended release of the CO2. Such event and others can be quickly detected by the proposed remote sensing arrangement. Remedial action may be taken to seal off the leakage site where the detection event DE occurred.

[0100] Thus, broadly, the proposed remote sensing arrangement RSA, in particular detection system DS, may be used in oil & gas industries, in environmental monitoring, oceanographic or other aquatic research endeavours, and others still. The reservoir RSV from which the fluid escapes to form the detection event may be under the seabed SB as indeed mentioned above in the CCS context. But the reservoir RSV may not necessarily be so, as it may be on the seabed SB, or it may be elsewhere in the water column, such as fuel tanks or pipelines which leak gas or crude. Other examples may include oil rig machinery, such as blowout preventers, manifolds, or other rig related machinery at wellheads where leakage of the crude or gas may occur, or any other machinery, abandoned or in use. In any of the above-mentioned examples, the leak / discharge event is undesirable, sometimes even outright dangerous, and can be detected quickly with the proposed remote sensing arrangement, thus helping to increase safety and / or environmental sustainability, or other objectives. The reservoir RSV may be manmade or natural, or partly both.

[0101] The data processing by detection system DS as proposed herein is preferably done on-board the watercraft CR that carries the sensor device SD as depicted in Figure 1. Such on-board processing, where the detection system DS is at least partly, preferably wholly, integrated into the sensor device or sensor-carrying structure (watercraft CR or fixed station) is indeed preferred herein for real-time, responsive processing. Thus, principles of edge computing, positing processing close to data source, may be heeded herein for good real-time experience. However, the computing may be done instead at least partly or fully away from the sonar device or watercraft WC, such as onshore, or on a support vessel VS, as the need may be. Whilst such support vessels VS may be used herein, such as with ROVs or other, such support vessel SV is not a necessity herein. Some embodiments can do without such vessels VS, or they may rarely come in onto play, such as AUVs or fixed underwater station UWS is used for example.

[0102] Operation in coastal areas in waters over the continental shelf CSH is one main application field envisaged herein, however this does not exclude deep water DW operations beyond the continental drop-off DO. As said, the sonar-carrying watercraft CR,V that carries the sensor device SD is preferably an AUV. However, in some embodiments, a ROV is used instead, remotely operated by a human user. In such or other embodiments, the sonar-carrying watercraft CR may be linked by a tether TH to the support vessel VS.

[0103] The tetherTH (essentially a cable) is used to physically secure against loss of the potentially expensive watercraft CR, and may also serve as a data link where data d output by the on-board detection system DS can be transmitted from watercraft CR to the support vessel VS, or the tether TH may allow transmitting power from support vessel VS to ROV watercraft CR. A reel system RL may be used for cable management of the tether to allow as unimpeded movement of the watercraft CR as is possible to carry out its operation. The vessel VS may be of the research vessel type / class (“RV”), such as Germanischer Lloyd (GL) classification “100 A5 E or any other. Double-hulled Vessels VS may be used, capable to withstanding more challenging water environments at high seas, or in heavy coastal swells. The vessel VS may include a control centre CRS with computer equipment that is capable of receiving sonar imagery and / or detection data d transmitted up the link TH by the submersible watercraft CR. In some scenarios, the data d and / or sonar imagery may be analysed by users onboard the vessel VS, by using suitable computing and periphery equipment, such as monitors for visualization, and / or / further) data processing, as the case may be. The detection data d can be used to navigate the vessel VS closer to the leak. Crew may then launch equipment to fix the leak. In some cases, the ROV itself may include equipment, such as manipulators, through which user can effect such fix, such as may be the case of for work-class ROVs. However, in other cases the ROV is merely of observation-class and any fix of leak may need to be done through other equipment that can be launched off vessel VS thanks to the leak DE detected by the watercraft CR’s detection system DS. Preferably however, the proposed system and discharge event data it produces is consumed autonomously by a AUV in which the system is mounted. Thus, no human user for analysing the data, or a support vessel VS, etc, may be needed for the most part, if not ever.

[0104] The vessel VS may include a communication device CD that allows communication over-air with a control centre CRS’ on shore. For example, the data d and / or sonar imagery collected by the watercraft may be transmitted by radio-communication device CD through radio waves to counterpart over-air onshore communication devices CD’ and can be processed there as needed. Satellite communication may be used.

[0105] Use of the detection system DS in a ROV is merely one embodiment. Other, indeed much preferred embodiments, envisage use in the earlier mentioned AUV. Thus, watercraft CR may be an AUV that carries the sonar device SD, instead of ROV. The AUV is capable of autonomous navigation through the waters. Thus, AUVs do not require, during their operation, tethering TH with support vessel VS. Such AUVs may use an onboard power supply, such as battery or fuel-cell pack. Such AUV may be launched from vessel VS at the site of interest. The AUV is then free to explore the area using autonomous navigation. It may or may not proceed on a pre-programmed routing, but may still do so in simple embodiments. The autonomous vehicle AUV, CR may include obstacle avoidance logic. The AUV may cover an area, either pre-defined or not, where the autonomous vehicle AUV can roam the coastal or deep waters in its quest to detect such discharge events anywhere where they may occur. In such AUV embodiments, it is the AUV itself that consumes the discharge event detection data d as computed by its onboard detection system DE, based on the sonar imagery collected by AUV’s sonar device SD. Thus, the AUV may use the detection data for navigation to the geographical location indicated by computed detection data. It should be noted that the proposed sensing arrangement herein, be it for ROV, AUV or other, does not pre-suppose any knowledge on existing or known discharge events. No such knowledge is needed for the proposed sensing arrangement. Much rather, the watercraft CR roams through the water, whilst at certain intervals or (quasi-) continuously sonar imagery is acquired by its sonar SD. The so acquired sonar imagery is then processed by the detection system DS onboard the craft, in order to detect any discharge events, including their geographic location. If so found, navigation control circuitry NCC (see Figure 2) of the craft may enable it to navigate to the geographic location of the assumed discharge event. The AUV, once there, may then collect additional data in relation to the discharge event, such as acquiring other imagery, such as optical imagery, of the event DE, taking of on-site chemical or sediment samples for further analysis, etc, for example to confirm the computed characterization result. Any one or more of the sonar imagery, the computed detection data and optionally the additional data in relation to the on-site data gathering may be stored on on-bord computer memory of the watercraft CR.

[0106] Such or other data, in particular detection data d, so stored on-board watercraft CR’s memory can then be recovered later by the support vessel VS. For example, in an AUV setup, the watercraft CR may be programmed for regularly scheduled rendezvous between craft CR and vessel VS. The watercraft CR surfaces, and detection data d, sonar imagery, etc. is then read out / uploaded through reader circuitry that is plugged by user into the AUV’s data exchange interface circuitry, such as connector(s), etc. A wired orwireless protocol (such as wifi, Bluetooth, etc) may be used forthis. After data read-out, the AUV may then continue its leak DE detection patrol until the next rendezvous, etc. The read-out data, in particular detection data d, can be then analysed by other computing systems or user as needed, in order, for example, to initiate suitable remedial action to fix the leak, or to launch further investigation / monitoring, etc.

[0107] In general, in AUVs or fixed underwater stations UWS, data is primarily stored This ensures that detailed and large datasets are preserved for in-depth analysis. In ROVS, when a direct tether TH to a support vessel VS is available, real-time data transmission can occur, allowing immediate analysis and response.

[0108] In AUVs, at the end of its mission, or on rendezvous, some or the complete datasets on surfacing, and when the AUV is recovered onboard at the end of a mission / battery recharge, the data may be retrieved.

[0109] Instead of, or in addition to over-the air or cabled data communication, acoustic communication underwater may be used by AUV to transmit the gathered data, some of it, or some of its findings to another watercraft, station, etc, whether under- or on water. Whilst such acoustic communication is indeed constrained by distance and bandwidth, it may still be effective for transmitting critical and succinct information. For instance, the AUV can send alerts such as “Discharge Event detected at LAT, LON, proceeding to investigate", etc, in natural language or in coded form, such as a string “D+LAT-LON", that indicates a positive detection at the locale. This shortrange and / or short-message acoustic communication allows the support vessel or other recipient to be aware of significant findings in near real-time, enabling a responsive operational strategy.

[0110] Fixed stations UWS store most data on-board. They can transmit limited, critical data via acoustic signals due to the same constraints faced by AUVs. In specific scenarios, such as shallow water, fixed stations might deploy a Wi-Fi antenna for periodic data uploads.

[0111] More detailed reference is now made to Figure 2 which shows a block diagram with more details of the proposed remote sensing arrangement, preferably sonar based. The craft CR, if arranged as an ROV, may transmit the sonar imagery m and / or the detection data d in respect of a discharge event through the data link DL to mission control MC. A human user may use a control console CRC with which operation, in particular motion, of the watercraft CR may be controlled, such as in the mentioned ROV setting. At mission command, computer equipment CP may be used to further analyse, or manipulate data d. For example, the data can be visualized by a visualizer IZ on a display device DD for view by the team onboard the support vessel VS for example, or once transmitted onshore, by onshore mission command. No such tethering TH is needed in AUVs, preferred herein.

[0112] Inset Figure 2A shows more details of the watercraft CR. In addition to the sonar device SD, shown at the bottom of Figure 2A, there may be other components to gather data. For example, such data gathering components may include the mentioned optical camera OC, configured for still or video imagery. In addition, or instead, other camera system may be used, such as infrared, etc. Operation of optical camera may be supported by a lighting system LS that provides the necessary illumination. In general, the components are arranged, attached or mounted in or on a frame of suitable rigidity.

[0113] An on-board computing system CS of the watercraft CR may control the operation of some or all of the mentioned components. The computing system CS may include a processing unit PU, such as microprocessor capable of data manipulation, eg a CPU (central processing unit). In addition, or instead, more dedicated data processing circuitry may be used, such as a GPU (graphical processing unit), as will be described in more detail below in relation to some embodiments of the detection system DS. The computing system CS may further include memory MEM on which data can be stored.

[0114] The processing unit PU (CPU / GPU, etc) allows fetching data and instructions from memory for processing to produce the desired data, in particular the detection data d in relation to the discharge event DE. Thus, the processing unit PU may be used to implement the detection system DS. The detection system DS may be so implemented in software and / or hardware. The computing system CS of the watercraft CR may be arranged on one or more printed circuit boards PCB. The computing system CS may be arranged in a watertight enclosure ECL that can be closed off by a lid LD equipped with suitable sealing SL to securely hold all the sensitive electronics described. The enclosure ECL secured to the watercraft’s frame. The enclosure ECL is capable of withstanding the ambient underwater pressure up to the watercraft’s depth rating. The processing unit PU implementing the detection system DS may be connectable one or two or more interfaces / connector(s) CN to the host system of watercraft CR by an ethernet and a power cable, for example. The connection via connector CN is preferably removable for better maintenance, and versatility, such as in “building block” system. Specifically, unit PU may be removed from one draft and disconnected there, and plugged into another one, such as may be of use when a fleet of crafts is maintained, with different capabilities, depth ratings, etc,

[0115] In particular when the watercraft is an AUV, it may include its own on-board power supply, such as battery or fuel cell pack. Otherwise, in ROV’s, or other, power may be supplied through tether TH.

[0116] Through a connector CN, the data d and / or sonar imagery is transmitted or received up or down the data link DL to the receiver such as the control centre CS on the research vehicle or any other destination.

[0117] Underwater propulsion system PL, such as a suitable number of thruster modules, may be controlled by navigation control circuitry NCC under instruction from the control system CS. In AUVs, the navigation circuitry NCC implements the autonomous navigation operation towards a target, such as the location of the computed detection data. The underwater propulsion system PL allows the watercraft CR to move in or across the water column to position itself at the right altitude, and to roam the waters to collect the sonar imagery, based on which the discharge event data d is computed. All of the above components may be found equally in ROVs, only that there the navigation circuitry NCC may receive its instructions remotely from user onboard vessel VS. Also, no on-board electrical power source PS may be needed in ROVs, as they may receive power through tether TH. AUVs may not need a tether TH, but may still have a connector CN / interface, to apply the read-out circuitry on surfacing of AUV, as mentioned earlier when reading-out the collected sonar imagery and / or the detection data d as provided by detection system DS.

[0118] In downward looking sonar (not preferred herein), their operational altitude over the seabed SB in general followed the 10% rule, where the preferred altitude is about 10% of the range of the sonar device. In sharp contrast, in the present HF (high frequency) forward looking horizontal sonar SD, there are no such confines. In the present case, and explored in more detail below at Figure 3, horizontal aperture is broad, but vertical aperture of the sonar SD should be “thin” vertically, typically in an angular vertical divergence range of about 1° up to 72°, so that neither the water surface, nor the seafloor enters sonar SD’s FOV. As high frequency (“HF”) sonar is envisaged herein, in the region of 900 kHz or more (although lower frequencies may also be considered), the range is in general (but not necessarily) less than 100 metres. Operation in continental shelf waters is mainly envisaged, as this is where most discharge events of interest can be expected to happen, but operation beyond continental shelf waters or other aquatic environments is also envisaged.

[0119] Whilst use of the detection system is mainly envisaged in self-propelled submersible watercraft CR, such as the said ROV, AUVs, etc, other uses are not excluded herein. For example, the sonar device SD may be included in a towfish device, that is towed by vessel VS to cover an area to be examined for discharge events. Other options include mounting the sonar device on a surface vessel beneath its waterline. For example, sonar device SD may be so mounted on the hull of support vessel VS. In both cases, propulsion of sonar SD is provided by the vessel itself. In still other embodiments, sonar device is mounted on bottom crawlers that are designed to sink to the sea floor on deployment as may use taction against sea floor such as wheels or tracks for propulsion. In this case, the crawler may use a gantry to mount sonar device at a suitable height above sea floor SB. In still other embodiments, no movable infrastructure may be needed at all. Instead, the sonar device SD is mounted to a fixed underwater station UWS for monitoring a given area for the emergence of discharge events. Such stationary structure / station etc, may include benthic landers, water samplers or profilers, or some such. In such stationary use, the sonar device SD may include a stepper motor or similar that allows 360° rotation of the sonar’s transducer / transmitter / transceiver around a vertical axis to provide the necessary field of view. Such stepper motor or similar equipment for rotation of sonar’s FOV may also be provided when using the sonar SD on-board the said watercrafts, whether self-propelled (AUV, ROV, etc) or not (hull-mounted or towfish, etc).

[0120] Reference is now made to Figure 3, which shows more details of the sonar device SD as envisaged herein in embodiments. A co-ordinate system frame used in Figure 3 is three dimensional (3D), with three axes X,Y,Z. Axes X, Yin general describe the imaging plane in which the acquired sonar imagery m can be thought to be conceptionally located in 3D space, whilst the third spatial dimension perpendicular to the image plane, referred to as axis Z, represents the altitude or depth of the sonar device SD as carried by the watercraft CR. Specially, and following marine convention, Z is positive downwards. X is positive bow, and Y completes right-handed backhoe coordinate system, with positive starboard side direction, in (A) axis Z points into the plane of the drawing, whilst in (B) axis Y points into the plane of the drawing.

[0121] Referring first to sub-Figure 3A), this shows, along depth direction Z, a plan view of the imaging geometry used by the sonar device SD envisaged herein, whilst inset Figure 3B) affords an elevation view of the imaging geometry, perpendicular to a normal of the imaging plane. Acoustic waves or pulses that are emitted by the sonar device in multiple beams in its data collection form an insonified area ISA, which the field of view FOV of the sonar device is based. The sonar device is of the active sonar type, in particular of the scanning sonar type. The sonar device may include one or more beamforming circuitry capable of generating a fan-shaped insonified area ISA, also referred to herein simply as “fan”. The fan / insonified area ISA is approximation planar, thus having plane that defines the fan’s orientation. The terms “fan” and “insonified area ISA” are used interchangeably herein. Thus, in general the sonar may be said to have a horizontal imaging geometry. Because the fan / insonified area ISA is approximately planar, either may be referred to herein as “scanning plane” (of the sonar device SD). Multiple acoustic beams that are emitted by sonar SD’s emitter make up the fan / insonified area ISA. This, in combination with the capabilities of the sonar SD’s receiver define the sonar’s actual field of view (FOV). In radical departure to traditional scanning sonar systems, which are reliant on vertical fan, the proposed imaging geometry is instead one of a horizontal fan geometry. This is illustrated in the said Figs 3A) , B) . It will be understood herein that the user herein of spatial qualifiers such as horizontal and vertical and similar, are with reference to the said coordinate system X,Y,Z, and in particular with reference to field lines of gravity.

[0122] Thus, the horizontal imaging geometry of the forward-looking scanning sonar SD used herein has a wider horizontal divergence, or fan angle a, than its vertical divergent angle p. Vertical divergence angle p is negligible as compared to horizontal fan angle a, p « a. The acoustical beam(s) thus forms a substantially planar horizontal fan. In most cases, such as when used over continental shelf waters, the fan’s plane is substantially parallel to an average seabed profile over which the sonar SD operates as the craft CR is navigating through the waters BW.

[0123] Whilst the use of such traditional vertical fans are not necessarily excluded herein, the proposed horizontal fan forming the sonar SD’s insonified area ISA, or its image plane, is in some cases, such as in continental shelf waters, configured to be substantially parallel to the seabed SB. It has been found herein that such horizontal imaging geometry is particularly suited for the detection of the discharge (eg, leakage or in-water emission) events DE, as will now be explained in more detail. Use of such horizonal fan makes the sonar sensitive to such discharge events DE of interest herein. As said, in many of application scenarios, the sonar is positioned and operable so that the acoustic scanning plane of the fan is parallel to the seabed. The sonar device SD should be securely attached to mitigate, if not avoid vibrations, eg, by using buffer elements and / or by firm integration of it within the structural frame of the craft CR. This and the horizontal orientation of the fan ensures that the water column to be detected crosses the sonar SD’s scanning plane transversely.

[0124] The exact alignment of the horizontal sonar fan can be established by using an optional second sonar system, or using the same sonar system in a different mode, namely for vertical bottom profiling, in which backscatter signals are collected data downwards, along Z direction, toward of the seafloor SB underneath the sonar device, to so establish a reference direction for the seabed SB. In addition, or instead, existing bathymetric maps that map out the ocean floor profile may be used instead to establish the average orientation of the sea floor for present purposes. Yet in other embodiments, accelerometers mounted in watercraft CR are used to establish the gravity field. The insonified area ISA, that is the horizontal fan, is then generated by the sonar device in the scanning plane perpendicular to the gravity field. A gravitometer may be used for example to establish the precise orientation of the fan’s horizontal orientation, as shown in Figures 3A), B). If more than one sonar device SD, SD’, etc is used, their respective images may be merged into a single acoustic image to improve coverage, in particular horizontal coverage.

[0125] A purpose of using such an untraditional horizontal sonar fan is illustrated in Figure 3C). Some or all discharge events DE of interest herein may manifest as a plume of bubbles or a jet of such bubbles that issue forth from the reservoir RSV into the ambient water BW. The bubbles traverse the water column towards the water surface, driven by buoyancy, and substantially following the opposite direction of the gravity field. As the sonar device passes through the water column at a suitable altitude, the bubbles are bound to traverse the insonified plane ISA of the horizontal sonar fan. As the bubbles pass through the plane over time, a unique pattern of acoustic reflectance cross sections o(t) will emerge: for a given bubble, this cross-section will increase in size to a maximum and then decrease again as the bubble is ingressing, and then egressing, the insonified area ISA, which in good approximation is in most cases part of horizontal plane with negligible vertical “thickness”, thus small angle / 3. This observation applies to each bubble, and this combines in total as shown in the graph of Figure 3D) to an oscillating or pulsating (flickering) backscatter signal. Broadly, this observation is harnessed herein thanks to the horizontal sonar fan angle (with very small angle P) to analyse the received sonar imagery for backscatter signatures that are indicative for such a pulsating / oscillating behaviour because of bubbles crossing the horizontal scanning plane of the sonar SD. Because the source of the leak is, in general, spatially stationary such, as a crack in the seabed SB through which gas, such as CO2 or Methane, is released, the discharge event of interest herein is not only characterized by the above-mentioned temporal, pulsating backscatter reflectance, but also by its relative stationarity in space. Thus, there is some spatial invariance that together with the temporal, pulsating acoustic reflectance, can be harnessed herein as an acoustic signature for which to search in the acquired imagery. Such an in-image search for such in-image structure enables the proposed detection system DS to find hitherto unknown leakages or other discharge events anywhere in the insonified portion of the water column.

[0126] A source of the leak (such as a crack in the seabed or a defect in a pipeline) may then be found by examining the water column along direction Z that passes through the geographical location of the found detection event as per the sonar imagery. Thus, the craft CR, once it reached the geographical location, may examine the water column up and / or down. Descending to the seabed underneath the geographical location may be an option to find leaks that are geological in origin. However, in general, the source of event DE may be anywhere in the water column along axis Z through the geographical location of the detected event, such as a leaking of man-made structures, a rogue or abandoned, jettisoned, tank, container, pipeline, etc. Examination in whichever way of vertical water column may be used to locate origin of discharge. Thus, for this purpose, sonar SD mare be switched from horizontal to now vertical geometry, or another watercraft with vertical sonar may be called in for this task.

[0127] It will be understood that the above-described horizontal sonar fan parallel to seabed SB underneath is merely one example. What matters herein is that the (mean) propagation vector of the plume of bubbles released in the discharge event is generally upwards towards the water surface of the body, and such mean propagation vector is preferably perpendicular to the plane of sonar SD’s fan. In particular, the spatial extension of the insonified area in the direction of the propagation vector of plume (in general, along Z) should be substantially less than the spatial extent of the fan / ISA in the other two spatial directions X,Y. Thus, fan should be “thinner” than it is wide as observed earlier, with the scanning plane in (X,Y) being perpendicular to propagation vector of the bubbles that make up the plume. Whilst said perpendicularity of fan vs said mean propagation vector is preferable herein, this is not a strict requirement: some angulation may be acceptable [ 90° + / - zf (zf > 0) ], as this may still allow the detection system DS to pick up the described pulsating feature at a suitable frequency. It can be seen then that a traditional vertical sonar fan may not be so suitable, as the frequency of the pulsation may be very low or may be drowned out altogether in the relatively long passage thorough the traditional vertical fan (preferably not used herein). Thus, the detection system DS may not be able to detect a suitably strong acoustic signature with such vertical fan. Thus, as can be seen, for best results, the spatial extent of the proposed sonar fan in the (mean) direction of the plume’s propagation direction should be very small (eg, in the order of the said 1 ° <p <12°, and preferably be in the order of, or less than, an average size of the bubbles in the plume. The thinner the fan, the smaller p, the better. The exact geometry of such substantially horizontal sonar fan can be readily established in an experimental setup, where average bubbles sizes of some fluids of interest can be obtained, eg via optical imagery and measurements of bubble features recordable in such imagery. Thus, the horizontal sonar fan is positioned in such a way that the (mean) propagation vector of the bubbles' plume is perpendicular to this plane. This orientation ensures the detection system DS can effectively observe the pulsating backscatter signature of the bubbles as they travel upwards.

[0128] The above observation on interaction of bubbles with the horizontal sonar fan, in particular, the i) temporal pulsation of acoustic reflectance, and, optionally, the ii) the (global) spatial stationarity of such pulsation, will be referred to herein, for the sake of brevity, as the “bubbling-off’ property. Thus, the detection system DS proposed herein is configured for detection of discharge events DE by taking advantage of the said bubbling-off property and how this property manifests in in-image structures as per the collected sonar imagery. A method for leveraging this property for effective discharge event detection will now be described in more detail below.

[0129] Reference is now made to Figure 4, which shows a block diagram of the discharge events detection system DS (referred to herein for brevity as the detection system DS) as may be used herein to process sonar imagery. Broadly, the detection system DS is configured to search for, in the acquired sonar imagery m, signals having the bubbling-off property, as discussed above in Figure 3. Thus, such signals are spatially stationary but temporally pulsating signals.

[0130] The detection system DS may be implemented by one or more computational entities, or nodes, possibly fully integrated into the sonar device or into the watercraft CR carrying the sonar device. General purpose CPUs may be used for implementation, but more dedicated special purpose processor circuitry, such as GPUs or others, may be used instead or in addition to general purpose CPUs, in particular to facilitate real-time processing. Preferably, the CPU / GPU is low-power for example, less than 10-5-25 W), and / or of multi-core design. A minimum data transfer rate in the GB / s range is preferred for better responsiveness. The input data processed by the detection system DS includes raw sonar imagery m as acquired by the high frequency sonar device SD using the horizontal fan imaging geometry as discussed above. Specifically, input data received at input port IN of detection system includes 2D acoustic (raw) imagery, also referred to herein as sonar imagery. Acoustic images are generally enormous in size (~ 1px / cm). One of the challenges that proposed solution addresses herein is the collection and analysis of such a large dataset that, with traditional techniques, must be blanket collected. The proposed detection setup allows optimizing the data collection (that is, the committing in long term memory), as data is retained in memory only where necessary, in particular when such event is actually detected).

[0131] The sonar device SD may provide the sonar imagery m in the form of acoustic backscatter imagery made up of a matrix of pixels. Each row and column of this matrix represents or is related to the acoustic intensity value of the backscatter for a specific distance and azimuth as measured from the sonar reference, such as respective position of the sonar device at the time the backscatter is received. The backscatter is caused by acoustic reflectance of material / matter (other than water) that happens to be within the insonified portion of the ambient water. Each pixel in the sonar image (the terms “sonar image(ry)” and “acoustic image(ry)” will be used interchangeably herein) may be provided in suitable coding format, such 16-bit grayscale backscatter intensity data, where the value “0” (renderable for example as black) corresponds to no backscatter return, and the value 65535 (renderable for example in white) corresponds to the maximum return. However, it should be noted that for present discharge event detection purposes, no actual displaying of the sonar imagery is needed, but can still be done if required. Metric distances may be calculated by the sonar SD using data measured on the velocity of sound provided by a specialized speed of sound probe SSP.

[0132] Specifically, pixel values m(x.y) of sonar imagery m as acquired by the HF (and thus high resolution) scanning sonar is preferably adjusted to account for prevalent environmental conditions, in particular for the actual speed of sound in the waters the sonar SD operates in. The speed of sound probe SSP may be used, co-located with the sensor device SD, or at least with the watercraft CR. Thus, the speed of sound probe SSP may be mounted on the watercraft CR, such as on the sonar device SD itself. The speed of sound probe SSP is capable of measuring the actual speed of sound c’ in the ambient waters the sonar device operates in. This is because the actual speed of sound c’ is known to vary with environmental parameters of the ambient waters, including pressure, water temperature, salinity, and other parameters. The distance from the sensor device SD to where a particular backscatter event occurs may then be computed so that pixel value m(x,y) ~ c’*T / 2, with pixel position (x,y) being related to a particular bearing / azimuth, and T referring to the echo time of the bounced off acoustic pulse as transmitted earlier by sonar SD’s transmitter-transducer. Thus, the raw imagery m= m(x,y) is adjusted for the actual speed of sound c’ so that the imagery m comports with the given prevalent environmental conditions of the ambient waters. Whether or not the imagery m is so adjusted for actual speed of sound c’, the same notation “m" will be used herein. Thus, any reference herein to sonar imagery m or similar may include a reference to so sound of speed adjusted sonar imagery. In the context of sonar imagery, "bearing" typically refers to the direction relative to the vessel's heading, while "azimuth" generally refers to the angle between the projected vector and a reference direction on the horizontal plane. For present purposes both terms may be used interchangeably to denote the direction from the sonar device to the location of the backscatter event. Specifically, "azimuth" refers to the angle in the sonar imagery. However, when considering the installation of the sonar on the watercraft and the associated offsets or lever arms, "bearing" becomes relevant as it relates to the vessel's heading. In navigation, "bearing" indicates the direction the craft CR needs to point to stay on course. Similarly, x and y coordinates in the sonar imagery correspond to geographical positions in terms of northing and easting for example.

[0133] The input data may further include local spatial calibration data, such as lever arm LA data. Such data may be understood as off-set values (expressed, eg, as length data (eg, in cm or mm) or in angular data. The lever arm LA data relates to position(s) and / or orientation(s) of the mounted sonar device SD on watercraft CR. Thus, collectively, lever arm LA data represents measurements on how / where the sonar SD is mounted on the watercraft CR. The lever arm LA data may be expressed with respect of a reference point on or in the watercraft CR. The lever arm LA data may be used to calibrate the acquired sonar imagery m. Again, the same notation “m" will be used herein for raw data calibrated based on lever arm data LA.

[0134] The sonar imagery is essentially spatial data so that each pixel value m(x,y) that make up the imagery pertains to a 3D position in the ambient waters BW. Geo-referencing may be used to locate the acquired imagery in the said 3D space. Thus, in embodiments, a geo-referencer facility GF is used to so geo-reference the acquired acoustic / sonar imagery m. Furthermore, each distance / azimuth data pixel is convertible to local metric x / y coordinate. Geo-referencer facility GF (which may also be referred to herein as a navigation compensation module) is operative to pair some or each acoustic image frame mt=m with a corresponding navigation data packet p, based on reference time (timestamp). In this way, geo-referencer facility GF recalculates the reference plane of the image by transforming some or each local tangent based x / y point into a north / east geographic coordinate (ENU (east, north, up) or NED (north, east, down)), based on projection geodetic datum (UTM - universal transverse Mercator projection). Instead, any other projection may be used, and so may be left-handed versions of the geographic coordinate, instead of the right-handed more commonly used in a navigation. The craft CR’s navigation data p is provided by an onboard navigation module NAV and includes position and attitude of the vehicle. Such data p may be provided in any suitable format, such as per an NMEA (National Marine Electronics Association) protocol, such as NMEA-0183 protocol, or any other equivalent data format in terms of information, i.e. , position of the craft or fixed station on which sonar device SD is arranged. Information encoded in data p may include one or more of all of temporal position indicator, latitude / longitude position, depth, elevation from seabed, latitude / longitude standard deviation, depth standard deviation, UTM, North / East projected position, velocity vs. seabed estimation from a DVL (Doppler Velocity Logger) (not shown), including misalignment and / or scale factor, velocity of sound, heading (or course or compass), roll, pitch, heave (waves), rotational speed, CoG (Course over Ground) terrain, horizontal speed (Speed over Ground), standard deviations, of equivalents of any one or more of the aforementioned. Any time synchronization may be implemented with a PPS signal and a ZDA message (such as per NMEA-0183), or in any other way. In case sonar SD is mounted on a fixed station, the location of the station installation and lever arms data (the geometry of the sonar installation in or on the station) will be sufficient. Time synchronizer TSY may time synchronize the various data streams involved, including the sonar imagery, speed of sound probe, navigation data, etc. This may allow to offset for latencies for example. PPS (Pulse Per Second) may be used. PPS provides a highly accurate time signal from the navigation system NAV (such as GPS or other), ensuring precise time alignment. In addition, or instead, ZDA (Time and Date Stamp) may be used. This may supplement the PPS signal with date and time information from the navigation system (in particular GPS, ensuring relevant data are synchronized to the same time reference. In addition NTP (Network Time Protocol) may be used. This may act as a secondary synchronization method. It adjusts the system time over a network, ensuring relevant devices remain synchronized. For example, NTP is applied to systems already synchronized with PPS+ZDA to maintain time accuracy and correct any minor drifts.

[0135] Thus, some or each instance of the acquired sonar imagery m is tagged with, or otherwise associated with, navigational data p that represents, for some or each pixel in the acquired imagery, a respective geographical position, preferably also including altitude over seabed SB, to which the said pixel pertains. The so possibly geo-referenced acoustic / sonar raw imagery m is received at the data processing system for processing. The per pixel m(x,y) geo-positional data p may however not necessarily be included in imagery m as such, such as meta-data or other, although this can still be done in some embodiments. Instead, it may be sufficient herein that an associative data structure, such as a LUT or table or other, is maintained in memory that associates pixel (x,y) with such geodata p. Preferably, in addition or instead to such geo-referencing, the sonar imagery is time- tagged / stamped by the time f at which it was acquired. The acoustic imagery m is preferably received as a time series of multiply frames m=mt, and preferably some of each frame mtinstance is georeferenced by time and geographical position, and may be tagged accordingly m = mt,P. Thus, the sonar device may be operable to acquire the sonar imagery m as multiple frames m= mt over time, each frame having its own pixels values. Reference herein to sonar / acoustic imagery should be this constructed as a reference to some or all such frames collectively, or as a reference to a given frame at a given time t. Again, the same notation “m" may be used herein for the geo- and / or time referenced sonar raw data, m = mt,P.

[0136] Thus, the sonar raw imagery may be assumed herein to be any one or more of: i) geo-refenced, ii) time stamped, iii) actual speed of sound adjusted, iv) calibrate for lever arm data. The time stamping allows synchronization, as mentioned above. The time stamping may include time of day t (hours, minute, second, etc), and date t’ (day / month / year). The geo-referencerGF is not only used for enabling navigation to a possible discharge event DE, but is also used to position (register) the sonar imagery in 3D space thereby allowing the in-image search component ISC to correctly compare the received sonar imagery m. In principle, for this registering operation, any other reference facility, different from the geo-referencer GF, may be used that allows spatially registering the received. Broadly, the detection system DS allows analysing the sonar imagery m to decide whether or not there is a discharge event DE somewhere in underwater 3D space that has been insonified by the sonar device SD. Thus, the detection system DS is capable of searching for hitherto unknown discharge event(s) DE and, preferably, locate and characterize same. For example, the detection system DS may be able to establish what kind of material (fluid) is released in the discharge event, and how much (amount) of material was so released into the ambient waters BW. The latter can be estimated based on estimating the flow rate of the observed discharge event. The detection system DS is capable of efficient processing, particularly in real-time as the sonar imagery is acquired by sonar SD. It can do this in parts thanks to a multi-stage set-up which will be discussed hereinbelow in more detail. The first stage is formed by an in-image search component ISC that is capable of searching image domain of the acquired sonar imagery for promising in-image candidate locations u (referred to herein simply as “candidate location(s)”) within the acoustic imagery. Only image pixels from those candidate locations u (discrete subsets of pixels in image domain) are then processed by a second stage downstream the first stage, which stage will be referred to herein as the characterizer component CC.

[0137] This multi-stage set-up allows for better results and more responsiveness. This is because operation the characterizer component CC, preferably machine learning based, is computationally rather expensive. By having the first stage administered by the in-image search component narrow down the whole of the acquired imagery to specific, discrete pixel sub-sets that are the candidate locations within the respective image planes of the acquired imagery, allows for more targeted, and thus responsive, processing.

[0138] The in-image search component ISC may proceed itself in multiple sub-stages to gradually narrow down, possibly across levels, the candidate locations. For example, at a first level a certain number of candidate locations are found, and these are then further narrowed down in multiple levels to form ever smaller sub-sets thereof, to arrive at the last level at a single or plural candidate locations that are thought to represent such discharge events DE. It is only raw image data that is collected from the original raw acoustic imagery, but only at the selected candidate locations after narrowing down, that is fed into the characterizer component CC for characterization of the detection event. This narrowing down from level to level into an ever smaller number of candidate locations is indicated in the Figure symbolically by the inverted striped triangle. Pixels outside the candidate locations are ignored, or at least down-weighted relative to pixel values within the one or more candidate locations u so found. Binary or weighted (“fuzzy”) masks may be used to implement such narrowing down processing.

[0139] In yet more detail, the in-image search component ISC may include sub-stages. One such sub-stage may include a digital acoustic images processor AIP that may process the raw imagery by filtering or otherwise conditioning in order to emphasize or amplify pixels values in the imagery whose distribution and variation may be consistent with the bubbling-off property. The acoustic images processor AIP may use one or more filter stages that are so geared to amplify, if not isolate from remaining pixel values, the ones that comport in suitable proximation to the bubbling-off property. The so conditioned imagery, eg filtered, sonar imagery f(m) is then passed on to the next stage for processing by a detection component DC. The detection component identifies locations in the so filtered imagery based on geometric properties of pixel value distributions that are thought to be caused by effects (eg, bubble formation) at play in the discharge event of interest. For example, if bubble formation is involved in the detection events, this may suggest using geometric shape filtering that configured to respond to shapes related to ellipsoidal structures.

[0140] The detection component stage may be operative to monitor / track an evolution of pixel values m(x,y) G u = t / / < in the detected candidate locations u. Based on such evolution (if any) over time, decision logic ET may be operable to decide whether a given candidate location should be considered a final / genuine candidate location for the discharge event of interest, worthy of processing by the next stage characterizer component CC. Other candidate locations are eliminated based on selection policy, and are not put before the characterizer component CC for processing.

[0141] Thus, as will be apparent from the above, with detection system DS’s process flow passing through one or more levels of narrowing down, an initial list of candidate locations may be narrowed down more and more, possibly in one or more iterations, to arrive at a final list of final candidate locations, and it is only raw pixel data form such final candidate locations that are processed by characterizer component CC. Raw pixel data m(x,y) demarked by those one or more (final) candidate locations in the originally received acoustic input imagery is then accessed and forwarded on for processing by the characterization component CC.

[0142] It should be understood that the operation of the detection system may be dynamic in nature. Thus, the in-image search and the characterisation may be updated, whenever a new acoustic frame mtis received.

[0143] Reference is now made to Figure 5 which shows in more details of operation of the in-image search component ISC. The acquired acoustic imagery m, optionally time t / date t’-stamped, and / or georeferenced p, is received at in-image search component ISC.

[0144] The in-image search component ISC may include the acoustic images processor AIP. Based on certain processing parameters w,b,s, to be described in more detail below, the acoustic images processor AIP filters the acquired acoustic imagery m to produce a transformed or filtered image f(m)=m’. For this purpose, the acoustic images processor AIP may implement a specifically configured transformation f to be applied to the received acoustic imagery m.

[0145] In preferred embodiments, transformation f may implement based on a moving average filtering and / or variance filtering, or in particular a moving window variance filtering. Moving window variance (“MWV”) filtering has been found to be particularly sensitive to imagery that includes structure indicative of the bubbling-off property as discussed above at Figure 3. In general, the acoustic images processor AIP produces the said candidate locations u. The candidate locations are sub-sets of the image plane of imagery m. There may be more than one such candidate locations u in the said imagery m. If the input imagery m comprises a times series of multiple frames mtacquired at different times t, the processing by acoustic image processor AIP may proceed per frame mt. There may be one or more such candidate locations u per frame mt, The candidate locations demark one or more regions of pixels in the original imagery m. The pixels so demarked have a particular pixel value pattern representative of an acoustic signature of the bubbling-off property.

[0146] The so filtered imagery f(m) may then be passed on to the detection component DC. The detection component DC narrows down the list of candidate locations found by the acoustic image processor AIP to a subset of such candidate locations. The detection component DC attempts to detect, in the pixels demarked by the candidate locations, patterns (acoustic signatures) that comply with a given geometric shape pattern. Candidate locations u whose pixels do so comply are then passed on to the next level, at which a further narrowing down operation may be applied, and so on, to so derive, at the final level, at a reduced list of candidate locations as compared to the initial list of candidate locations provided at the first, initial level, such as by acoustic image processor AIP.

[0147] For example, the detection component may work in tandem with an optional event tracker ET. The event tracker may track the evolution over time of the detected patterns at the candidate locations u provided by detection component DC, in order to so yet further narrow down the candidate locations at the next level. A tracker policy q may be used, as will be explained more fully below at Figure 5 later below. Broadly, the tracker policy q that may prescribe temporal characterizations, such as persistence, etc, expected of true / genuine discharge events.

[0148] Thus, as described above, in some embodiments there may be 3 narrowing down levels implemented, starting with the acoustic image processor AIP, followed by the detection component DC, and then by the optional event tracker ET. There may be more than 3 levels, or there may be merely two. The order of operation of the detection component DC and the event tracker ET may be reversed in some embodiments. Preferably, acoustic image processor AIP implements the first / initial level to provide an initial pool of candidate locations which may then be sequentially narrowed-down in the following one or more levels.

[0149] The final list of candidate locations may then be passed on to the characterizer CC component, although in some application scenarios a mere detection of discharge events may be sufficient. The characterizer component CC may fetch for some or each candidate location per frame mtthe set of pixels values (acoustic signatures) demarked by the respective candidate location in the respective frame mt, to obtain the characterization result x- The characterization result x may be understood as event descriptors that describe in more detail the discharge event DE, such as in terms of material released or the amount of material so released, or other. The proposed sequential narrowing down {uk} c {n of an initial list of candidate locations u ={ui} may increase responsiveness of detection system DS, not least because a computational load on the characterizer component CC may be reduced. In addition, improved reliability may be obtained this way, as the characterizer CC component is provided with "confirmed" events

[0150] Because of the geo-referencing, the geographic location p as pertains to the pixels demarked by the final in-image candidate locations can be looked-up in a geo-database or other suitable associative data structure and passed on to the navigation circuitry NCC. The navigation circuity may be coupled to the craft CR’s navigation module NAV, such as GPS (in case of ROV), or Ian inertial navigation system in case of AUV. The inertial navigation system calculates position based on both direct and indirect measurements. The direct measurements include GPS data when the AUV is on the surface, and Doppler Velocity Loggers (DVL) data while diving. Kalman filtering may be used. Any drift errors that may accrue over time can be corrected for by periodic surfacing to consume fresh GPS das, or based on received acoustic signals, using USBL (Ultra Short BaseLine) or LBL (Long BaseLine) systems. Such acoustic signals may be provided by support vessel occasionally, or by a navigation aid buoy, etc. In some cases, even ROVs may benefit from such acoustic data, such as USBL, or other.

[0151] Based on the current navigation data p as provided by the navigation module NAV, the navigation circuitry NCC may operate, in particular in an AUV setting, to control watercraft CR ‘s propulsion system PL, in order to cause watercraft CR to navigate to the geographical location of the suspected discharge event. Thus, the craft CR may so physically arrive at the site of the discharge event DE for further examination, if any. Thus, the geo-referencing allows navigation of the watercraft CR to the detection event site for further inspection, such as the taking of samples, the taking of further imagery, such as of optical imagery, or for other on-site analysis, etc, as mentioned earlier and as the case may be.

[0152] The following describes the above-mentioned components in yet more detail. In this connection, reference is made first to Figure 6 which shows the acoustic images processor AIP in more detail. The acoustic image processor AIP itself may include two stages, one upstream of the other. In the first stage TA a temporal analysis is done which is followed by the second stage SA where a spatial analysis may be done of the received input acoustic / sonar imagery as collected by sonar device SD.

[0153] Thus, the filter function fmay be understood as the functional concatenation f = l / l / ° B °T of operations l / l / , B, T, now described in more details. More specifically, in the temporal analyser stage TA, a transformer TF is used that enhances or amplifies a temporal characteristic as may be manifest over multiple frames mt, and in line with the sought after bubbling-off property as discussed above in Figure 3. In this connection, the transformer TA may be configured as a moving average filter, such as the said moving window variance filter l / l / , to produce a variance image W(m), using a window size w, as one of the operational parameters mentioned earlier. Thus, a moving windowed function l / l / may be implemented by transformer TF to produce variance image W(m). In the, optional, follow-up spatial analyser stage SA, artifacts or irrelevant in-image details are removed. For example, in particular, a blur filter BF may be used that implements a blur function B using a suitable kernel size b. This may be followed by a thresholding function TRF that implements a thresholding function T, based on the threshold s. Pixel values below a certain threshold are disregarded in this manner and so are possible artifacts that are “ironed out” by the blur filter B. At the same time, subset(s) / region(s) of pixels that are mimicking the bubbling-off property and may thus be considered tokens for genuine discharge events, are emphasized by the moving window variance filter. The blur filtering B and the thresholding T are optional.

[0154] In some embodiments, the acoustic images processor AIP may operate or combine multiple frames mtof the acquired acoustic imagery m in order to produce the transformed frame W(m) at a given instance t. Thus, a buffer or similar memory may be used, where a time series of acquired acoustic frames mtis accumulated. Once the buffer is full, that is, once the requisite number of frames mtare acquired and are so accumulated, these are then processed together to produce the filtered image W(m) for a given time instance t. This operation may then be repeated by considering the next time series that is formed by the next incoming frame(s) and that then fill up the buffer to produce the next filtered image for the next time instance, and so forth. This operation is illustrated in in the left part of Figure 7, which shows such a time series of prior acquired acoustic imagery mt, which can be used to produce the filtered image W(mt) at time t, based on earlier frames mt, t’ < t.

[0155] Specifically, the windowed filter W of acoustic image processor AIP is geared is to pick up acoustic signature of passing bubbles that tend to produce a flickering or pulsation in the backscatter response in a sequence of consecutively recorded images frames m , as mentioned above at Figure in 3 in relation to the bubbling-off property. Specially, the acoustic image processor AIP processes a sequence of acoustic images mtto filter the fixed parts of the image, and enhance the variable components, such as by using the said Moving Window Variance filtering. In order to reduce computational workload, the Moving Window Variance may be implemented in recursive form according to Welford's algorithm, or other. An unbiased sample version of this recursive approach may be used, such as:

[0156] In (1), mtis the acoustic image frame received at time / step f stored and handled as a matrix of integers l / l / x / 7. mtdenotes the sample mean of the previous A / samples (mt-N, — ,mt) as per window size, given recursively by m’t = W(mf) denotes the Moving Window Variance image at time / step t, of the last N samples, again stored and handled as a matrix of integers of the same size l / l / x / 7, and the “bar symbol " " denotes over time averaging of the samples of the images m taken at times t-1, t-2, ... t-(N-1). The setup (2) fosters numerical stability and efficiency in computation.

[0157] In other words, at each step t, the filter produces a matrix in which each element represents the Moving Window Variance of each pixel from the sequence of the previous N acoustic images mt-N, ■■■ ,mtas captured by sonar SD. Matrix (1) may be represented as a variance image W(m), in which each pixel may be coded at 16-bit value. Such values represent the Moving Window Variance corresponding to each pixel of given acoustic images at time mt, based on its N precursors frames mt, t’ < t received in the sonar imagery stream as captured by sonar SD. Each such value in W(m) represents backscatter variability at a given geographic location, assuming the imager is geo-referenced as is preferred herein. In this representation, the pixel values of variance image W(m) change with the amount of variability, depending on the exact coding. For example, in one coding scheme, smaller values indicating less variability than greater values. Variance image W(m) may be conceptualized graphically, eg, in greyscale, with white (zero variance) corresponding to pixels where backscatter remains unchanged in the sequence, whereas the darker the shading the greater the local variability at the respective pixel position. Again, reference to bitmaps and “greyscale” etc, is merely for illustration. Herein, for detection purposes, no visualization is needed, but can still be done if required. Instead of using moving window variance, other filtering schemes, including operation in frequency domain, and then inverse transforming back into spatial domain, is also envisaged herein in alternative embodiments. Thus, Wavelet transforms based filtering, or median filtering, are also envisaged herein. In short, any image filtering setup may be used that respects, picks up or delivers good response to in-image structure that is represents / is consistent with the bubbling of property (Fig 3).

[0158] The Gaussian blur filter B may help to reduce noise. This filter component has the effect of removing scattered pixels containing numerical variance residuals. The thresholder T applies threshold so that all pixels where the variance is less than a certain value (the Threshold s) are suppressed.

[0159] The output as produced by acoustic image processor AIP results in an intermediate image (matrix) rri, in which only the candidate locations emerge as nonzero pixels. Such pixels are those potentially associable with the events of interest DE. Referring back to Figure 7, its right portion is a visualization of variance image W(m) in monochrome bitmap, in which the dark rendering (black) represents the selected pixels that make up the candidate location. Figure 8 is a similar visualization, but now with blur filtering and thresholding applied.

[0160] The window size w of filter W is the number of samples to be considered in the Moving Window Variance filter. It can be adjusted to reduce acoustic artifacts caused by parasitic phenomena such as echoes and reverberations. A high value of w makes it possible to capture phenomena with slower variability. Window size w should be chosen by considering the ping (acoustic wave emission and reception) of the sonar SD (that is, its acoustic image acquisition frequency) and, therefore, the time window w corresponds preferably to the length of the acquired image sequence mt.

[0161] Blur or filter kernel size b is the size in pixels of the Blur filter B. Low values are effective for dealing with small emissions. High values reduce sensitivity when observing the event of interest.

[0162] The variance Threshold s is to discriminate events of interest from other artifacts. High values are effective for filtering out variations related to slow or sporadic changes in images, thus limiting the possibility of false positives.

[0163] As user interface may also user to adjust any one of more of parameters w,b,s of acoustic image processor AIP.

[0164] With continued referred to Figure 8, this illustrates such the blurred and thresholded variance image W(m) = m’ as may be produced by the acoustic images processor AIP discussed above in Figure 6. In more detail, Figure 8 represents the image plane of the so processed acoustic imagery / frame f(mt) for a given time t, with a candidate location u being identified, in fact isolated, from the remaining pixel values, by filters W,B and T. The pixel values m(x,y) in the original imagery m at the found location u have certain variation patterns that may be referred to herein as an acoustic signature. This acoustic signature is thought to represent an instance of a detection event, as the over-time behaviour across frames mto pixel values at the acoustic signature is consistent with the assumption of the bubbling- off property of Figure 3. Thus, the filtering yield the candidate location u. Once this is found, system fetches pixels from the original data m, and this goes into the characterizer CC for the actual discharge event characterization.

[0165] The output m’ of the in-image search component, using the bubbling-off property enhancing filter such as MWVF or other, optionally blur filtered and thresholded, may be represented as a binary mask, a binary image, where '1 ' indicates the presence of flickering (bubbles), and 'O' indicates its absence. The pixel values that map to “1” form the candidate location(s) u, The shape of the locations u (considered as subset in the respective image plane of the acquired sonar imagery m, may be detected / filtered for by the detection component DC A fuzzy mask with weights may be used instead of a binary mask.

[0166] Reference is now made to Figure 9 which shows details of the detection component DC in its function as a feature extractor FE. The variance imagery W(m)=m’, or otherwise filtered or pre-conditioned acoustic imagery m’ that indicates candidate locations u are further processed by structural spatial filtering, such as a Hough transformer H and / or a blob detector L, and / or more, as needed. This helps to further narrow down the detected locations as per m’, to produce more defined locations as subsets of pixels or regions. The feature extractor FE may be configured as a filter that responds to certain predefined geometric shapes. The geometric shapes may pertain to shapes of bubbles or group of bubbles that are formed in some discharge events as the lower density fluid is drawn in a plume to the water surface. Thus, the feature extractor FE may be configured to respond to ellipses or circles, such as is the case for the Hough transform H or blob transform L. The candidate locations which cause a response of the geometric shape filter FE at a sufficient strength (again, thresholding may be used) are flagged up as candite location for the next level. Thus, feature instructor FE looks up the pixel values in the original imagery m within each such flagged up candidate location(s) u\, and such pixels values are then made available for processing by the characterizer component CC, or are passed on the next narrowing down level, eg for processing by event tracker ET. Thus, in general the geometric shapes to which feature extractor responds are related to effects that are at play in the type of discharge events of interest.

[0167] In more detail, The Hough transform H, as one example envisaged herein, provides a list of circles detected in the, filtered image f(m)=m’, and providing, for some or each of them, the location and radius. A blob as may be found by blob detector L may be defined as a group of pixels that form a kind of colony, spot or a larger object, in general topologically connected, that is distinguishable from the background. The Blob detection algorithm L returns the in-image position and size of each blob detected in the filtered image rri, based on a definition of the shape of the blob (size, convexity, inertia, and circularity, or other). The two or more different detection techniques L,B may be run in sequence, or, preferably, in parallel / independently on the (filtered) variance image rri, to obtain for example, a list of circles and a list of blobs as potential, narrowed down, candidate locations. The geometric information so returned may be used as possible identifiers of the event of interest DE. Thus, the list of candidate (event) locations that may include a respective associated event descriptor. Each such descriptor per candidate location may describe a related candidate location event, eg in terms of a set of information, such as location, size, time of observation, and / or others, as deliverable by the detection component DC.

[0168] Based on such descriptors, candidate locations may be considered coincident if they are overlapping in location and size. If the same candidate locations is detected by both geometric feature detectors, L,H, that is, is detected as blob and as a circle, a higher weight may be assigned.

[0169] The list of candidate locations thus obtained may be passed to the Events T racker module ET, in order to further narrow down the list of candidate locations, to obtain reportable events, processable by characterizer component CC.

[0170] Instead of, or in addition to, using the Hough and / or Blob transform, Laplacian of Gaussian (LoG) filtering may be used instead.

[0171] Figure 10 shows more details of the operation of the event tracker ET that tracks evolution of the candidate locations, in particular of the evolution over time of pixel values that are demarked by the candidate locations of an earlier level (such as provided by acoustic image processor AIP or feature extractor FE, DC, or other), to so yet further narrow down the list of current level candidate locations and that are then put for processing before the characterizer component CC.

[0172] Tracker ET may Implement a policy q of tracking and updating candidate locations, detected by the detection component DC This allows further discriminating, and thus narrowing down, based on tracked evolution, the list of candidate locations. This allows distinguishing the event of interest relating to gaseous or liquid discharges / in-water emissions from other phenomena such as moving objects (eg, shoal of fish etc), or residual acoustic artifacts. An optional event classifier component (not shown) of the event tracker ET can further discriminate the list of events based on training, taking into account a varied and reasonably comprehensive set of parameters.

[0173] Thus, the tracking policy is used to define how the evolution of an event that follows a certain pattern can lead to the identification of a relevant event. In this level of the chain of narrowing down operations, the tracking policy q may be defined by any one of more or more of the following evolution parameters: margin, oblivion and persistence threshold.

[0174] The margin, or spatial resolution, defines a spatial distance, eg, a minimum metric in-image distance between two candidate locations (also referred to in the present evolution context as in-image “events”) to be considered distinguishable.

[0175] The oblivion parameter relates to a time interval, after which an event must be forgotten if it is no longer observed.

[0176] Persistence measures how frequently an event is observed relative to the total number of sonar pings at a candidate location. The persistence indicates and asks for a ratio. The ratio measures in how many of the observations the event was seen (detection in the transformed image f(m) as per detection component DC), compared to the total number of observations (sonar pings) at a candidate location. This helps to qualify events such as continuous bubble plumes, sporadic plumes with periodic small groups of bubbles, or single bubbles passing occasionally. The persistence value remains constant if the passage of bubbles occurs at the same geographical location (candidate location). Otherwise, the persistence value decreases until the event is removed from the list if it falls below a certain threshold. The persistence threshold distinguishes between persistent events or those that decrease asymptotically, and events associated with the passage of objects in the insonified area or other one- off events (e.g., artifacts) that are no longer present. This threshold helps to indicate when to forget an observed event if its persistence is not relevant.

[0177] In other words, events with persistence below the threshold are considered phenomena (in the sense of variability) of passage, which do not occur in the same position. Above the threshold, however, they are events that occur in the same position with increasing "intensity of persistence": from sporadic passages of bubbles to continuous plume, in this sense, while oblivion allows a banal temporal filtering of forgetfulness, persistence contains more complex information that includes frequency (time), geography (space) and thus intensity.

[0178] The event tracker ET may proceed iteratively. At some or each iteration, the event tacker may receive S1010 a list of current candidate events / locations as input and may process same to provide the reduced list ' of final candidate locations. The candidate locations may have associated with them respective one of more in-image geometric descriptors D, as may be supplied by the detector component, as explained above in terms of the example of Hough / - / and Blob L filter.

[0179] The event tracker ET may be implemented as follows way. A new candidate location is read in S1010. For some or each such read in candidate event ui, a condition is checked at S1020: if the candidate event is not in the current list, eg as may be measured in terms of the margin parameter A, it is added S1030a to the detected events list by initializing its descriptors D to a default value. The descriptors are initialized when a new candidate event is added to the list, and refer to the descriptors of the said event. Such descriptors may be a spatio-temporal descriptor. It may include any one or more of a center position or other reference position, size, timestamp, observation count, and persistence, etc. Otherwise, if the candidate event has already been marked as detected, its descriptors will be updated S1030b based on the candidate location being analysed. Thus, at S1040 an updated list+is obtained. Such updating may include any one or more of i) the center is updated using the new information (eg, descriptor related to new observation / candidate location), ii) the size is updated using the new information, iii) the timestamp t is updated to the new value (eg, time stamp of new candidate location), iv) the observation count is increased, v) persistence is recalculated. However, i)-v) are mere examples. The exact nature of the description update will depend on the type of descriptor at hand. At S1050 it is checked if all candidate locations / events are processed. If not, the above is repeated for the next event. If yes, process follow proceeds to a cleansing phase.

[0180] In other words, and in sum, the above processing S1010-S1050 is a cyclical processing, based on a list of candidate locations provided by module DT. For some or each such candidate location in the list: a given such u is read in at S1010. It is then check if it was already in the list «_t) from the previous step, by comparing the locations based on the "margin" criterion. The flow checks at S1020 whether the said u is already in the list. I not, u is added S1030a to the list. If yes, the descriptors are added S1030b. The list of candidate locations is updated S1040 accordingly. At S1050 it is checked whether all applicable u’s have been processed / completed.

[0181] In the cleansing phase, the updated list of detected events is read in S1060, and the list is cleansed at step S1080b, based on checking S10710 a condition. For example, a newly detected event can be eliminated S1080b if one of the following conditions occurs: i) the event was not seen for the time defined by the oblivion parameter A, ii) the persistence r is below the persistence threshold. If the event passes the check at S1070, it is maintained S1080a. At loop S1090 it is checked whether all detected events are cleansed. If not, the above is repeated for each newly detected event. If yes, the cleansed list ' of detected events is made available, eg, may be passed on to the characterizer component CC for processing.

[0182] Thus, the tracker ET may use variance images rri or otherwise suitably filtered imagery to identify events at each step. It then correlates these events with those observed in previous steps to track their evolution. For instance, if a plume is made up of small groups of occasional bubbles, the tracker can recognize that intermittent observations of variance at a specific location correspond to the same plume, estimating persistence parameters to distinguish between continuous and occasional plumes. If an event moves spatially across observations, the tracker distinguishes it based on movement patterns, such as a plume consistently emitted from the same area versus a moving object like fish, which shows greater spatial freedom.

[0183] The event tracker ET may be understood as a processing component that filters and “enriches” the groups of pixels, on the basis of information that at this processing stage begins to become more substantial. Thus, at this stage the groups of pixels associated with candidate locations u are transformed into groups of information (eg, descriptors D, any one or more of: position, size, shape, persistence, speed, image (for the characterizer CC), etc) by means of operational parameters (eg, any one or more of margin, oblivion, persistence threshold).

[0184] Figure 11 is an illustration of the characterizer component CC. It may include a classifier component CL that is operable on the received acoustic patterns as per the (final) candidate location(s) ui, to classify same into a material / fluid type that it is estimated to have been discharged into the water in the discharge event. In addition, or instead, a quantifier component QT may be used that estimates the amount of material / fluid so discharged, whichever fluid this may have been. Preferably, however, both the classifier CL and the quantifier QT are used in conjunction to obtain discharge event descriptors x- These may then be combined or aggregated by an aggregator Z to so release a fuller characterization of the alleged detection event DE. A function of the aggregator Z may be to combine the output from the classifier CL and the quantifier QT. It may integrate the type of material and the estimated quantity discharged into a comprehensive event descriptor. This process ensures that the detected event DE is fully characterized, providing a complete and accurate representation of the discharge event for subsequent analysis and action. For instance, the aggregator Z may correlate information from both modules to refine the overall estimation. For example, knowing the type of gas can improve the accuracy of the quantification process. The aggregator may combine the result from the characterizer component CC with the associated geographical location, eg, for navigation on-site.

[0185] Figure 12 illustrates histograms of sample acoustic patterns / acoustic signatures in candidate locations found by the detection system DE described above. In more detail, histograms A)-D) illustrate how different characteristics of the bubble plume (such as extent, density, intensity, etc.) result in distinct acoustic images. The histograms highlight how various plume characteristics lead to different intensity distributions in the acoustic images. These differences may be harnessed herein for ML based classification and / or regression into / for characterization results.

[0186] The characterizer component CC, in particular the classifier CL and / or the quantifier QT, are preferably arranged as a respective trained machine learning (“ML”) model Me, MQ. Any one or both of such models Me, MQ may be referred to herein as the model M = (Me, MQ) of characterizer component CC. It has been observed that the proposed two stage approach with the in-search image component first identifying the candidate locations and only having those processed by the machine learning based characterizer component CC allows securing certain advantages. Specifically, operation of the inimage search component ISC may be considered a type of pre-processing that is preferably run also on the training data on which.the model M is trained. Thus, it is preferred herein to access some same training sonar imagery, apply thereto the in-image search component ISC, and the output of the inimage search component ISC is then fed into the model M as input in training. In response to this training input, the model M outputs training output data, which is then compared to certain targets. Based on this comparison, training parameters 9 of the model M are adapted, in the training. The training procedure is usually formulated as an optimization process, driven by a cost function. The optimization may be gradient based. The pre-processing via the in-image search component ISC facilities the optimization process by nudging the optimization to better “optima” (usually minima) on the surface of the cost function. This avoids, or at least reduces the likelihood of, the optimization process becoming “trapped” in some local minima.

[0187] Thus, it is intended herein to use at least the in-image search component ISC, optionally with the detection component DC, FE and / or event tracker ET, also as a pre-processing stage in training data generation and / or in training the machine learning model, in addition to use during deployment as described above at Figures 1-12. Thus, training data, preferably collected in the field or lab, or synthetically generated, are first processed by the intended in-image search component ISC, and it is the so processed imagery that is then used to train the machine learning model, thus ensuring better performance and conditioning overall.

[0188] Reference is now made to the flow chart of Figure 13, which shows steps of a sonar / acoustic imagerybased detection method. Specifically, the method is suitable to detect possible discharge events, based on received acoustic imagery m. In some embodiments the imagery is made up of a time series of sonar / acoustic frames m=mt, and is provided by a sonar device operating underwater. The sonar is preferably of the HF type. The method affords real-time detection of hitherto unknow discharge event(s), anywhere in the ambient waters insonified by sonar’s fan.

[0189] The proposed method may be understood to implement the above-described discharge event detection system DS, but the below described steps may constitute a teaching in its own right. At step S1310 one or more high frequency sonar is operated, preferably of the scanning type, with a horizontal fan to acquire acoustic imagery. A submersible or floating watercraft may carry the sonar device during a patrol in an underwater area whilst the input sonar imagery is so acquired. This allows insonifying a large body of water in which to monitor for any such discharge event(s). The craft may be an AUV or ROV. Mounting such sonar device at a fixed point on the seabed, such as on a benthic lander or other underwater station / structure is also envisaged. The sonar’s transducer transmitter / receiver may be rotatable for up to 360° coverage.

[0190] At step S1330 the so acquired acoustic imagery is received, preferably geo-referenced at S1320 using GPS data or any other.

[0191] At step S1340 an in-image search is conducted in the received imagery to find candidate locations (subset(s), in particular clusters, of pixels) that may represent an acoustic signature (a pixel pattern) that may be representative of a discharge of interest. In general, such clusters or groups of pixels form contiguous pixels, thus representing homogeneity and / or spatial coherence of the detected acoustic signatures.

[0192] At step S1350 pixels in the received sonar imagery as demarked by the so found location(s) u are then processed to characterize the discharge event. Thus, in this step, pixels are fetched from within so found location(s) u (region, neighbourhood) in the sonar imagery m (as such raw imagery) as acquired / received. Thus, discharge event characterizer data may be made available as characterization result. Processing at step S1350 is preferably based on one or more trained machine learning models, trained on training data. The demarked pixels may be fed as input in the model to obtain such characterizing result. Such characterizing result may indicate any or more of amount of material / fluid being discharged per unit time or over given time period, or the type of material that is being discharged.

[0193] The characterizer event data is then made available for further processing such as for displaying, storing, or other processing. In particular, at an optional such processing step S1360, the so found and / or characterized candidate location(s) in the imagery are used to navigate the submersible or floating watercraft to the respective underwater location that corresponds to the one or more in-image candidate locations. Such correspondence can be established if the acquired sonar imagery is georeferenced S1320, as is indeed envisaged herein in preferred embodiments.

[0194] The in-image search step S1340 is now described in more detail. This step may include at (sub-) step S1340_10 of filtering the received imagery using a moving window filtering, such as moving window variance filtering or any other. This allows isolating sub-set(s) of pixels, whose over spatial and / or over-time behaviour is consistent with a modelling property of fluid discharge events. Such property may include the bubbling-off property as described above at Figure 3. At sub-step S1340_20 spatial processing may be done such as any one or both of blurring and thresholding.

[0195] At step S1340_30 a feature detection is done, based on imagery as processed in prior steps S1340_10-S1340_20, taking into account expected geometric properties of the said candidate locations. A given candidate location is in general a subset of pixels. The subset forms a 2D region, as said above cluster of pixels, and can thus said to have shape. In general, the subset is a cluster of adjacent pixels. The shape is thought to be caused by the discharge event of interest in general. For example, some discharge events manifest as a released plume of bubbles. Spatial geometries of such bubbles may be reflected in the spatial filtering done at step S1340_30. For example, bubbles may be assumed in approximation to have a spherical shape or more general ellipsoidal shape, thus having circular or ellipse shape in 2D, or more generally, convexity. As an example, spherical / ellipsoidal feature detectors / extractors may be used herein in step S1340_30, such as Hough filtering, or blob filtering, or any other geometric shape driven filtering that is apt for the given case at hand. In general, convex feature detectors can be effective in identifying the shapes of bubble plumes and other relevant discharge events it has been found.

[0196] At step S1340_40 an evolution over time in relation to the so detected locations can be conducted using a tracking policy q. This can be done in particular when the imagery is in the form of a time series of acoustic frames mt, with some or each frame received at step S1330.

[0197] Thus, it may be understood that steps S1340_10-S1410_40 may be understood as a cascaded sequence of narrowing down operations, where an initial set of candidate locations as may have been found at step S1340_10 is narrowed down or reduced, over one or multiple levels, to a final set of candidate locations. This final set comprises fewer candidate locations than the earlier set at earlier levels. It is then pixels for this reduced set of (final) candidate locations that are put before the one or more machine learning models at S1350 for processing into the characterization result.

[0198] In general, the candidate locations are formed from one or more subsets of pixels, located in the image plane and found in the search S1340. A given, so found, candidate location may be comprised of a topologically connected subset of pixels, but some fragmentation is not excluded herein such as may occur in case of backscatter signals received from divergent plume, fanned out over a wider area.

[0199] In addition, it should be understood that preferably, that the search step S1340 yields the candidate location as a subset of pixels that map out a region. The purpose of step S1340 is to find these (candidate locations). But the pixels put before the characterization step S1350 are then taken at the found candidate location from the original raw imagery as received at step S1330 and are in general not taken from the various filtered intermediate imagery produced in the narrowing down operations of step S1340. This can be done because there is in general native spatial registry between the original raw imagery and the various intermediate imagery, and thus of the events / candidate location descriptors D.

[0200] Reference is now made to Figure 14 which shows an architecture of the machine learning model M as may be used herein in embodiments. The reference “M” for the machine learning model as used herein, is a generic reference and refers either to the classifier model Me or to the quantifier model MQ, or to both. The model M is preferably of the neural network (“NN”) type, more particularly of the convolutional NN type (“CNN”). Such CNN per model MQ, Me may be used herein in order to account for the spatial correlations inherent in spatial data, such as is the acquired acoustic imagery.

[0201] As illustrated in Fig 14, the NN-type model may be made up a number of cascaded (hidden) layers Lj, in between input IL and output layer OL. More than one, in particular more than two such hidden layers may be used to form a deep learning architecture. The model includes nodes in the layers having parameters. The nodes are inter-connected such that nodes in one layer receive input from previous layers, and so on. Data flow may proceed from left to right with output from one layer passed on as input to the next layer, and so on, until the discharge event characterization x emerge at output layer OL as a regression or classification result.

[0202] The architecture and in particular the parameters 9 of the model obtained through training may be stored in the memory MEM’, and as may be used in deployment post-training, or may be used for in testing on testing data, as needed.

[0203] Preferably, some or all of the hidden layers are convolutional layers, that is, they include one or more convolutional filters CV which process an input feature map from an earlier layer into intermediate output, sometimes referred to as logits. Convolutional layers CV have been found to be beneficial in particular in upstream processing because they are good at recognizing spatial patterns in imagery. An optional bias term may be applied by addition for example. An activation layer processes in a nonlinear manner the logits into a next generation feature map which is then output and passed as input to the next layer, and so forth. The activation layer may be implemented as a rectified linear unit RELU as shown, or as a soft-max-function, a sigmoid-function, tanh-function or any other suitable non-linear function. Optionally, there may be other functional layers such as pooling layers P or drop-out layers (not shown) to foster more robust learning. The pooling layers P reduce dimension of output, whilst drop-out layer sever connections between nodes from different layers to combat overfitting effects.

[0204] In embodiments, downstream of the sequence of convolutional layers, and upstream the output layer, there may be one or more fully connected layers (not shown), in particular if a regression result is sought. The output layer ensures that the output has the correct size and / or dimension. In classification networks, the final layer may be a soft-max function layer. Broadly, once suitably trained, the input data such as the set of pixels within a given candidate location as produced by the in-image search component ISC, are received at input layer IL, and are propagated through the cascading layers thereby processed herein into feature maps as intermediate data inbetween layers, and are then provided as final output at output layer OL as a regression or classification result, such as type of fluid and / or amount of fluid discharged, or any other characterization, as needed. The input and output data may be provided in any suitable format. For example, fluid amount may be indicated as a scalar value, a flow rate, a volume, etc. The fluid type my correspond to an entry in classification vector, having a length equalling the number of types of fluids expected to be found in the body of water under investigation or monitoring. The input data u may be provided in in mark-up language (such as JSON or XML) and / or or bitmap patch. A bitmask may be used. The mark-up language may encode which pixel subset(s) are to be considered by the model M.

[0205] In general, in machine learning, four phases can be distinguished including, in this order, the training data provision phase, then training phase, then test phase and then deployment phase. The above described at Figs 1-13 related to deployment or training, thus the already trained model was used in practice or in test runs on data that is different from training or test data. Training in training phase may be a one-off operation, or can be done repeatedly once new training data emerges, using the currently trained model M from the earlier training cycle as pre-trained model and hence as an initial parameterization which may then be refined in the new training cycle.

[0206] Preferably, as a pre-processing stage in training, prior to feeding training input data x, such as sonar imagery, into the model M, the in-search component ISC is applied first to isolate candidate locations. It is then only pixel data from the original training sonar imagery, but within the found candidate location(s), that are fed into the ML model M as input for processing into training data output M(^). This is consistent with the manner in which trained model is used in deployment and as described above at Figs 4-13. The manner or processing of data by model M is in general the same for training and deployment / testing. The data (training input or real word data us applied to the model at input layer IL and is then cascaded trough the hidden layers by transforming into intermediate data, and is then output at output layer into a result M(x). This result may be classification into material fluid type or a regression into amount of material / fluid released. However, it is the amount estimate may not necessarily be a regression: in some cases a classification in intervals of flow rate / volume may be sufficient. Thus, classifier model MC may be classifier with a softmax layer as output layer. The quantifier model MQ may also be such as classifier model, or it may be arranged as regression model in which the output layer OL, and one or more hidden layers preceding same, may be fully connected, whilst layers further upstream, possibly including the input layer, are convolutional layers.

[0207] In more detail, in training phase, this includes adjusting parameters of the model M (either random or otherwise uninitialized with some random or otherwise parameters, or pretrained), based on training data. Once adjusted, based on pre-defined criteria, the model can be released for testing to check its performance. If satisfactory, the model can be released for deployment in real world applications, such as in oil & gas, or in any other field of endeavour to find discharge events. If training by a training system TS is done the, the in-image search component ISC may be used as mentioned above as preprocessor to first process accordingly the training data, and it is this pre-processed training data that is then used to train the model by training system TS.

[0208] The training data may be obtained synthetically or may be preferably obtained experimentally in a lab setup, or in field work. For example, in some embodiments, training data may be synthetically generated, for example by using suitably configured generative machine learning models, such as generative adversarial models (“GAN”), or other. In such a setup, from a relatively small pool of real sonar imagery data, the generative model can learn to generate realistic looking samples. The GAN approach was reported by / Goodfellow et al in their paper “Generative Adversarial Networks", published on preprint service ar:Xiv under document reference arXiv:1406.2661 , accessible under https: / / arxiv. org / abs / 1406.2661.

[0209] In addition, or instead to such as synthetical / computational approach, a lab set-up can be used to experimentally generate the training data. Preferably, a mix of data is used partly, from industrial (eg, asset integrity operations) or geological or oceanographic field work, in combination with synthetically or experimentally generated training data.

[0210] Figure 15 is a block diagram of a training system TS as may be used to train the model M based on training data. In Figure 15, and throughout herein, training data will generally be denoted by script letters x,^. x is indicative of training input data, such as (filtered) acoustic imagery, whilst^ represents the related target / label (ground truth), associated with training input data x. Such label / ground truth may indicate the material whose discharge is observed, or the amount of material so released, such as flow rate, volume, etc. Because of this association the paired notation x.^ may be used to denote training data generically herein, in particular in a supervised training setup as envisaged herein.

[0211] The training data {x,^ is in general different from the data seen by the trained model in deployment. The training data comprises x training input data x, and its related target ^ (“ground truth”). Such setup in pairs of training data either (x,^) are particularly suited for supervised learning, but other learning schemes such as sub-supervised or non-supervised, reinforced can be used instead.

[0212] In training phase, an architecture of a machine learning model M, such as the shown CNN network in Fig 4, is pre-populated with initial set of parameters 9, sometimes referred to as weights in NN contexts and others. The parameters 9 of the model represent a parameterization M9of the model (function) M. The parameters 9 may include filter parameters of the convolutional operators CV, and / or other parameters. It is the object of the training system TS to optimize and hence adapt the parameters 9 based on the training data. The training data may include the set of pairs {( x,^)k}, where k indicates the respective instance of such pair (^,-^)k: = (xk, ^k)k. In other words, the training can be formulized mathematically as an optimization scheme where a cost function F is minimized, although the dual formulation of maximizing a utility function may be used instead.

[0213] Assuming for now the paradigm of a cost function F, this measures the aggregated residue(s), that is, the summed error incurred between data estimated by the model and the targets as per some, or all, of the training data pairs in a batch or overall training data: argmineF =kD[ Mexk),^k\ (3)

[0214] In eq. (3), function / W() denotes the result of the model M applied to training input specimen xk. The result will differ in general from the associated targetfe. This difference, or the respective residuals for each training pair k, are measured by a distance measure D[-, ■]. Thus, the cost function F may be pixel -based, such as the Li or l_2-norm cost function, or any other norm Lp. Specifically, The Euclidean- type cost function in (3) (such as least squares or similar) may be used for regressing measurement xkinto an estimate of the object of interest position M9(xk). When the model is to act as classifier, the summation in (3) is formulated instead as one of cross-entropy or Kullback-Leibler divergence or similar. Binary or multi-class classification may be envisaged herein for some embodiments, where a discharge amount / flow rate estimation into intervals is sufficient.

[0215] The output training data M9(xk) is an estimate for target y,kassociated with the applied input training sensor measurement xk. As mentioned, in general there is an error between this output / We( and the associated target y,kfor each pair k. An optimization procedure such as backward / forward propagation or other gradient based method may then be used to adapt the parameters 9 of the model M so as to decrease the residue for the considered pair (a:fe,fe) / <or, preferably, a sum of residues for a batch (a subset) of such training pairs from the full training data set.

[0216] The optimization procedure may proceed iteratively. After one or more iterations in a first, inner, loop in which the parameters 9 of the model are updated by updater UP for the current batch of pairs the training system TS enters a second, an outer, loop where a next batch of training data pairs (^ / <+n,^ / <+n)k+n, (n=1,..,L) is processed accordingly. The structure of updater UP depends on the optimization procedure used. For example, the inner loop as administered by updater UP may be implemented by one or more forward and backward passes in a forward / backpropagation algorithm or other gradient based setup, based on the gradient grad(F) of F. While adapting the parameters, the aggregated, for example summed, residues of all the training pairs in the given batch are considered. The aggregated residue can be formed by configuring the objective function F as a sum of squared residues such as in eq. (3) of some or all considered residues for each batch. Other algebraic combinations instead of sums of squares are also envisaged. In general, the outer loop passes over batches (sets) of training data items. Each batch comprises plural training data items k, and the summation in (3) extends over the whole respective batch, rather than iterating one by one through the training pairs, although this latter option is not excluded herein. In training, as administered by training system TS, the machine learning model learns a latent mapping L-. X y that there is between the space X of acoustic pixel patterns found by in-image search component ISC o the one hand, and the space y of discharge event characterization on the other. Such mapping may be difficult to model analytically, but can be done by using the machine learning model and adapting its parameters based on the training data. The mapping is learned as a pattern in the product space Xx y, and such pattern is encoded in collection of parameters 9. In particular, neural network models are known to be able to provide good approximates of such mapping. It is suspected that there is a correlation between the type of discharged fluid (eg, nitrogen, air, methane and CO2) and their respective specific acoustic patterns that they cause on inson ification , as may be gleaned from Tomczyk et al. "Detection, localization and quantification of the emissions of gas from the seabed in fieldwork and experimental studies using active sonar systems" - Proceedings of the 11th European Conference on Underwater Acoustics, July 2012. Similarly, different flowrates of fluids may also cause a distinctive acoustic pattern, thanks to the particular way bubbles are accumulating and travelling,, acoustic resonance, where bubbles oscillate at specific frequencies when insonified. See also Phelps & Leighton (supra). The passage of bubbles through the flat transverse / horizontal sonar fan at different flow rates may affect the mentioned bubbling-off property, such as different frequencies of the pulsation / flickering caused.

[0217] The training data may be provided by a training data provider system TDPS. The training data includes, in particular, training input data x. The training input data includes examples or specimens of training candidate locations extracted by in-image search component ISC from training sonar imagery. The training data may further include related targets -y,, at least one such target -y, (known or measured discharge characterization) for each given input x. Thus, in particular in supervised training setups, the training data may be conceptualized, stored, and processed as a set of data pairs {( x,-y,)k}, where k indicates the respective instance / specimen of such pair (^, < / ,)k= (^k,-^k) / <.

[0218] The training data set {( x,^y)k} may be procured from existing sonar imagery, as may be held in databases. A human expert may annotate such historical imagery, alternatively, the data is produced by field measurements, or in controlled experiment in a laboratory (lab), or in any other way. This is described in more detail, below at Figs 16, 18. Data is collected preferably with the specific acquisition methodology of horizontal & HF sonar as earlier described. Automatic labelling for non-historical data may be used with benefit herein, as will be detailed below.

[0219] Whilst the above mainly related to supervised learning approach, which is robust and well-suited for the present setup, other additional / alternative learning techniques may also be used herein, such as reinforcement learning, semi-supervised learning, unsupervised learning, deep reinforcement learning, or others still. Reference is now made to Figure 16, which shows a flow chart of a method of providing training data, whilst Figure 17 shows a flow chart of a method of training the machine learning model(s), based on training data.

[0220] Referring first to Figure 16 in more detail, the method of providing training data may be realized in a manifold of different approaches. The method may include the general step S1610 of providing training data which may be done in any one of the following manners: i) generating synthetic sonar imagery with a priori known labels by using generative machine models such as GANs or other, ii) accessing existing sonar imagery in databases, and having such labelled by one or more human sonar experts, iii) acquiring sonar imagery in oceanographic, or more genera.l aquatic / underwater, field work for existing discharged events known to exist, such from geological-oceanographic maps and surveys, operational oil or gas extraction fields, etc. The type of fluid may be a priori known, such as from underwater geysers, hydrothermal vents, etc. Flow rate may be again estimated by a human expert for example. As a further option iv), computer simulations are run, based on hydrodynamic knowledge and properties of the fluid of interest. As a yet further option v), the step S1610 may include running lab experiments under controlled conditions. Optionally, training data imagery may be normalized to reduce variations due to environmental conditions and improve data consistency. Training for classification can be restricted to type(s) of fluid a priori of interest, optionally with a “dummy” class, indicative of neither such fluids of interest. The selection of which fluid(s) to train for may be driven by the intended operation, such as in CCS monitoring or in Oil & Gas asset integrity monitoring, or other.

[0221] For some oreach of the above embodiments i)-v) of step S1610, there may be a step S1620 of storing the respective type of training input sonar imagery x, obtained in whichever way, in association with its related label in a training data memory. The imagery x is suitably varied to build up a corpus of suitable number of such sonar imagery specimens x', each having its own label y , to so obtain a set of pairs {(^',^')}, which can be released for training of model M. Either when storing, or as a preprocessing step when passing on the training data to training system TS for training, the in-image search component ISC is applied. It is then only a subset in the image plane of each specimen of training imagery as demarked by the candidate locations(s) as provided by the in-image search component ISC that is passed on training data input for training model M. A binary of fuzzy mask may be used for this, for example to ensure only relevant portions of the image are used for training.

[0222] Referring now in particular to option v), the step S1610 may include causing a controlled discharge event in a pool of water and acquiring, in synchrony with such causing of discharge event, sonar imagery by operation of sonar device located in the pool of water and having discharge event’s plume insonified by the sonar device. Thus, the sonar device is operated so that the discharge event is within the sonar’s field of view. Preferably, sonar device uses horizontal sonar fan, whose plane is substantially perpendicular to plume’s upwardly propagation direction. As a refinement, parameters of the controlled discharge event may be varied to acquire such experimental sonar imagery under a number of different conditions and associated labels. Such parameters may include contextual parameters such as water temperature of the pool of water, pressure / depth, etc. Other contextual parameters may include density, salinity. Other parameters are label related and may include the type of fluid discharged into pool, and / or flow rate / volume of fluid so discharged, discharge / release pressure, vertical distance between the sonar and the discharge point, etc. Including the release / discharge pressure as a parameter may provide more accurate training data by accounting for variations in bubble formation due to different release pressures.

[0223] It may be preferable to combine options iii), v), thus, to use as training data lab data in combination with data collected in oceanographic (more generally, aquatic / underwater) / geological field work.

[0224] Referring now to first to Figure 17, this a shows a flow chart of a method of training a machine learning model based on training data.

[0225] Such training method may include a step S1710 where the training data is received. This may include historical data and / or the synthetically generated data, or data as generated by the method as per Figure 16, or in whichever way obtained.

[0226] Broadly, based on the training data, parameters of the model are adapted. This adaptation may be done in an iterative optimization procedure, such as (3) or similar. The procedure is driven by a cost function F. Once a stopping condition is fulfilled, the model is considered trained.

[0227] In more detail, and with particular reference to a supervised training setup, at step S1720, the training inputs xkin the current batch are applied to a machine learning model / W0having current parameters 9 to produce (respective) training output(s) M9(xk).

[0228] / deviation, or residue, of the training output / We(^;<)from the respective associated target ^k, is quantified at S1730 by cost function F. One or more parameters of the model are adapted at step S1740 in one or more iterations in an inner loop to improve the cost function. For instance, the model parameters are adapted to decrease the residues as measured by the cost function.

[0229] The training method then returns in an outer loop to step S1710 where the next batch of training data is fed in. In step S1720, the parameters of the model are adapted so that the aggregated residues, considered over the current and preferably over some or all previous batches, are decreased, in particular minimized. The cost function quantifies the aggregated residues. Forward- backward propagation or similar gradient-based techniques may be used in the inner loop. A dual formulation in terms of a maximization of a utility function is also envisaged.

[0230] Examples for gradient-based optimizations may include gradient descent, stochastic gradient, conjugate gradients, maximum likelihood methods, EM-maximization, Gauss-Newton, and others. Approaches other than gradient-based ones are also envisaged, such as Nelder-Mead, Bayesian optimization, simulated annealing, genetic algorithms, Monte Carlo methods, and others still. The training method in Fig 17 is an example. Any form of training, such as clustering or in whichever way is envisaged herein, where a model is adapted based on training data.

[0231] Reference is now made to Figure 18 which shows an embodiment of the training data provider system TDPS. The training data provider system TDPS is configured to generate training data in a lab setup with sonar device and a body of water, such as pool PE or in a real marine / underwater environment. The training data provider system TDPS may generate a controlled discharge event DE’ by releasing a known material into a controlled environment such as in a pool environment PE. The sonar device SD is preferably of the same type, as that used in deployment. Specially, it may be of the HF type, and may have beam forming circuitry capable of generating the horizontal sonar fan as described earlier herein at Figure 3.

[0232] The system TDPS may include a discharge event mechanism. This may include a pressurized fluid (eg, gas) reservoir PFR that provides compressed fluid of interest (i.e. CO2, CH4, air,...), in pressurized form into the pool PE. Inclusion of air or similar gases may be useful to allows system, in order to learn to better distinguish harmless (eg, air) discharge from harmful / undesirable ones. A valve system VS may be used, consisting of an intercept valve and a pressure regulator valve. A flow controller FCC then allows the pressurized gas to be released into the pool PE of water to so generate the experimental discharge event DE at various flow rates and / or for various released fluids. The flow controller FCC may include a reservoirs of different materials / fluids of interest, and these are selectively released into the pool environment PE, possibly at different flow rates. For liquid discharges, compressed air may act as propellant. The flow controller FCC measures some or all properties of the injected fluid. Such selection may be user controlled or automated via protocol. The respective selection of fluid and / or flow rate may be logged as control parameters that may serve as a respective label y,.

[0233] The fluid egresses a submerged nozzle NZ and is thus discharged into ambient pool PE water. The discharged fluid produces bubbles, which, in a plume, travels upwards to the surface. Whilst the plume propagates upwards, the plume is scanned by the sonar device SD as described earlier to produce input training data x (acquired sonar imagery). Because parameters of the released fluid are controlled, the labelling is inherent. The input training data x and the labelling (fluid type and / or flow rate / escaped volume or amount parameter in whichever form) is stored in a training data memory for later use in training the model(s) M by training system TS. Speed of sound probe SSP may be used as described above to acquire accurate sonar imagery. Experimental parameters, such as the type of fluids released and / or characteristics of the ambient water in the pool environment PE may be varied and measured, such as density, water salinity, temperatures, etc. In this manner, a training data set of sufficient variation is produced to so improve performance of model, such as its robustness and generalizing capability. Operation of the sonar device SD is synchronized with the flow controller FCC to ensure respective sonar imagery is taken of the plume DE’ at the set control parameter. Control of the flow controller and the sonar device is by a control management component CMG that transmits control signals through a communication network COM in which flow controller and sonar device are connected. The variation of the mentioned experimental parameters may be administered and / or caused by the control management component CMG.

[0234] The so acquired sonar imagery, which is naturally (in effect, automatically) labelled as the flow rate and the materials released are controlled by the management system CMG may be forwarded through the communication network to the image-search component ISC, to process the training input imagery as described above, and such data may then be stored in the training data memory MEM’. This data can be used later for training of either one or both ML models M= (MQ, Me), for example as explained above Figs 14,15 and 17. The control management component CMG may cause the various sonar frames x' to be stored in memory MEM’ with their associated label The said label includes the applicable control parameter (flow rate, type of fluid released, etc) used to cause the plume DE, whose sonar image was captured by sonar device thanks to the operational synchronization of sonar device and flow controller.

[0235] Whilst the discharge event DE was mainly explained above in terms of plume of bubbles, such is not a necessity herein. For example, the discharge event DE may be a plume or jet of a liquid released into the water WB that may not necessarily cause bubbles (but it may well do, such as micro-bubbles). So long as this liquid has different material properties (notably, flow property, viscosity, density, etc) from the ambient water WB, the above observations still hold true. Specifically, even there may not be bubbles, the “bubbling off’ property would still apply as the liquid passes preferably perpendicularly through the sonar’s fan. The difference in density, viscosity, etc, would still result in the flickering acoustic signature, akin to those caused by bubbles. The jet may gush out at an angle however, in which case the sonar SD’s beam forming circuitry may be operable to slant fan at such an angle, to ensure perpendicular passage of jet therethrough, if such source of the discharge event is suspected. Thus, sonar fan ISA would still be “thin”, but is angled relative to the vertical Z axis. The exact angle is in general not known, but could be “tuned” and the signatures found by the system DT could be observed to find the best such angle.

[0236] The components of the detection system DS may be implemented as one or more software modules, run on one or more general-purpose processing units PU integrated with the sonar device SD, or at least with the watercraft. Alternatively, some or all components of detection system DS may be arranged in hardware such as a suitably programmed microcontroller or microprocessor, such an FPGA (field-programmable-gate-array) or as a hardwired IC chip, an application specific integrated circuitry (ASIC), integrated into the craft CR or sonar SD. In a further embodiment still, the detection system DS may be implemented in both, partly in software and partly in hardware. The different components of the detection system DS may be implemented on a single data processing unit PU. Alternatively, some or more components are implemented on different processing units PU, possibly remotely arranged in a distributed architecture and connectable in a suitable communication network.

[0237] One or more features described herein can be configured or implemented as or with circuitry encoded within a computer-readable medium, and / or combinations thereof. Circuitry may include discrete and / or integrated circuitry, a system-on-a-chip (SOC), and combinations thereof, a machine, a computer system, a processor and memory, a computer program.

[0238] In another exemplary embodiment of the present invention, a computer program or a computer program element is provided that is characterized by being adapted to execute the method steps of the method according to one of the preceding embodiments, on an appropriate system.

[0239] The computer program element might therefore be stored on a computer unit, which might also be part of an embodiment of the present invention. This computing unit PU may be adapted to perform or induce a performing of the steps of the method described above. Moreover, it may be adapted to operate the components of the above-described apparatus. The computing unit PU can be adapted to operate automatically and / or to execute the orders of a user. A computer program may be loaded into a working memory of a data processor. The data processor may thus be equipped to carry out the method of the invention.

[0240] This exemplary embodiment of the invention covers both, a computer program that right from the beginning uses the invention and a computer program that by means of an up-date turns an existing program into a program that uses the invention.

[0241] Further on, the computer program element might be able to provide all necessary steps to fulfill the procedure of an exemplary embodiment of the method as described above.

[0242] According to a further exemplary embodiment of the present invention, a computer readable medium, such as a CD-ROM, is presented wherein the computer readable medium has a computer program element stored on it which computer program element is described by the preceding section.

[0243] A computer program may be stored and / or distributed on a suitable medium (in particular, but not necessarily, a non-transitory medium), such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the internet or other wired or wireless telecommunication systems.

[0244] However, the computer program may also be presented over a network like the World Wide Web and can be downloaded into the working memory of a data processor from such a network. According to a further exemplary embodiment of the present invention, a medium for making a computer program element available for downloading is provided, which computer program element is arranged to perform a method according to one of the previously described embodiments of the invention.

[0245] It has to be noted that embodiments of the invention are described with reference to different subject matters. In particular, some embodiments are described with reference to method type claims whereas other embodiments are described with reference to the device type claims. However, a person skilled in the art will gather from the above and the following description that, unless otherwise notified, in addition to any combination of features belonging to one type of subject matter also any combination between features relating to different subject matters is considered to be disclosed with this application. However, all features can be combined providing synergetic effects that are more than the simple summation of the features.

[0246] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. The invention is not limited to the disclosed embodiments. Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing a claimed invention, from a study of the drawings, the disclosure, and the dependent claims.

[0247] In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude a plurality. A single processor or other unit may fulfill the functions of several items re-cited in the claims. The mere fact that certain measures are re-cited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope.

Claims

CLAIMS1 . System (SYS, DS) for processing sonar imagery, comprising:- input interface (IN) for receiving sonar imagery obtainable by a sonar device (SD) in a body of water; an in-image search component (ISC) capable of searching, based on the sonar imagery, for at least one candidate in-image location in the sonar imagery, the at least one such candidate in-image location, when found, potentially representative of a respective discharge event of fluid dischargeable into the body of water; a characterizer component (CC) capable of processing sonar image values of the sonar imagery at the at least one found candidate in-image location into a characterization result that characterises the discharge event, and an output interface (OUT) for providing the characterization result.

2. System of claim 1 , wherein the characterizer component (CC) is based on a trained machine learning model (M).

3. System of claim 1 or 2, wherein the sonar imagery comprises a time series of frames, wherein the in-image search component (ISC), in searching for the at least one candidate in-image location, is capable of over-time processing of the frames to obtain the at least one candidate in-image location as one that is indicative of over-time changes of image values at the said at least one location.

4. System of claim 3, wherein the said over-time processing includes applying a moving window variance filter.

5. System of claim 3 or 4, wherein the said over-time changes of image values manifest as a pulsation.

6. System of any one of the preceding claims, wherein the sonar imagery comprises a time series of frames, and wherein the in-image search component (ISC) is capable of spatio-and / or- temporal analysis of the said frames in obtaining the at least one candidate in-image location.

7. System of any one of the preceding claims, comprising a geo-referencer facility (GF) capable of associating the at least one candidate in-image location and / or the respective provided characterization result with a respective geographical location of the respective dischargeevent.

8. System of any one of the preceding claims, including navigation control circuitry (NCC) capable of causing a watercraft (CR), carrying the sonar device (SD), to navigate to the geographical location.

9. System of any one of the preceding claims, wherein the system is arrangeable as a processing unit (PU) and this processing unit (PU) is mountable i) in the, or a, watercraft (CR) that is capable of moving the sonar, during its operation, in the said body of water (BW), or ii) in or at a fixed underwater structure (UWS).

10. System of claim 8 or 9, wherein the watercraft (CR) is an autonomous underwater vehicle, AUV, or a remotely operated vehicle, ROV.11 . System of any one of the preceding claims, wherein a field of view, FOV, of the sonar device is forward relative to a motion of the sonar device, and wherein a plane of the field of view is substantially horizontal, and / or wherein the plane is such that it is approximately perpendicular to a propagation direction of the fluid that forms the discharge event, and / or wherein the sonar device is a high frequency, HF, sonar device.

12. System of any one of the preceding claims, wherein the characterization result is any one of: i) type of fluid that is released in the discharge, ii) a quantity of the / a fluid released in the discharge.

13. System of any one of the preceding claims, wherein the fluid is a liquid other than water, or is a gas and / or wherein the discharge event includes bubbles formed in the body of water, the said bubbles being in a field of view of the sonar device14. Arrangement (RSA) including the system of any one of the preceding claims, and further including any one or more of: i) at least a part of the sonar device (SD), ii) a display device (DD) on which is displayable the characterization result and / or the candidate location, iii) a craft (CR) in or on or which the sonar device is mounted for moving same in the body of water, iv) interface (IF) circuitry through which the system and any one sonar device and craft are couplable, v) a power source (PS) through which powerable any one of: the system (SYS), the sonar device (SD), the craft (CR).

15. Method of processing sonar imagery, comprising: receiving (S1330) sonar imagery obtainable by a sonar device (SD) in a body of water; searching (S1340), based on the sonar imagery, for at least one candidate in-image locationin the sonar imagery, the at least one such candidate in-image location, when found, potentially representative of a respective discharge event of fluid dischargeable into the body of water; processing (S1350) sonar image values of the sonar imagery at the at least one found candidate in-image location into a characterization result that characterises the discharge event, and providing (S1360) the characterization result.

16. Method of training, based on training data, the machine learning model as in claim 2.

17. Method of providing training data for training the machine learning model of claim 2.

18. A computer program element, which, when being executed by at least one processing unit, is adapted to cause a processing unit or system to perform the method as per claim 15, 16 or17.

19. At least one computer readable medium having stored thereon the program element of claim18, and / or having stored thereon the trained machine learning model as per claim 2.

20. Use of a higher frequency sonar device having a horizontal imaging geometry for collecting sensor imagery for processing into characterization of a discharge event (DE).

Images