Marine environment dynamic monitoring system based on artificial intelligence and multi-source remote sensing cooperation

Abstract

The invention discloses a marine environment dynamic monitoring system based on cooperation of artificial intelligence and multi-source remote sensing, relates to the technical field of marine environment monitoring, and realizes multi-dimensional monitoring of a sea surface state by receiving multi-source data of a plurality of monitoring ends. Chlorophyll concentration mutation is detected in real time by setting sliding window difference, a subsequent processing flow is dynamically triggered, and the response speed of sudden events (such as red tide) is increased; aI is utilized to predict clock offset, interpolation reconstruction, weighted fusion and ocean current compensation technologies are combined, and space-time asynchronous errors of multi-source data are remarkably reduced; and generating a data cube fusing the backscattering coefficient, the temperature field and the chlorophyll gradient, and calculating a diffusion velocity field, thereby providing high-precision decision support for marine disaster early warning. The limitation of a traditional single data source is broken through, efficient cooperation of multi-modal data is driven through artificial intelligence, and the real-time performance, the accuracy and the dynamic analysis capability of marine environment monitoring are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of marine environmental monitoring, and specifically to a dynamic marine environmental monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing. Background Art

[0002] With the intensification of global climate change and the increasing frequency of marine economic activities, the importance of dynamic marine environmental monitoring has been significantly enhanced in the fields of disaster warning, ecological protection, and resource development. Traditional monitoring methods mainly rely on a single remote sensing data source (such as SAR satellites or infrared satellites) or local buoy observations, and there are the following technical bottlenecks: existing systems are difficult to effectively integrate heterogeneous data with different spatio-temporal resolutions and observation mechanisms (such as active microwave SAR, passive infrared remote sensing, and in-situ buoy data), resulting in limited early identification accuracy of marine phenomena (such as red tides and oil spills).

[0003] In recent years, artificial intelligence technology has shown potential in the field of remote sensing. For example, the LSTM network has been used for time series data prediction, and the convolutional neural network can improve the interpretation accuracy of SAR images. However, existing research has mainly focused on the optimization of a single data source, and a complete technical chain covering data triggering, spatio-temporal collaboration, and dynamic fusion has not been established. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides a dynamic marine environmental monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] A dynamic marine environmental monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing includes

[0007] a data acquisition module for receiving in parallel the backscattering coefficient matrix from SAR satellites, the infrared radiance temperature sequence from geostationary meteorological satellites, and the chlorophyll fluorescence pulse signal stream from a buoy array;

[0008] an event detection module connected to the data acquisition module for performing sliding window difference calculation on the chlorophyll fluorescence pulse signal stream of the buoy array, and generating an event trigger signal when the change rate of chlorophyll concentration between adjacent windows exceeds a preset threshold;

[0009] a clock offset prediction module connected to the event detection module, using a time series prediction model trained with historical clock deviation data, and outputting the clock offset amounts of each data source within a future predetermined time period in response to the event trigger signal;

[0010] a spatio-temporal calibration module connected to the clock offset prediction module and the data acquisition module, and performing the following operations:

[0011] Interpolate and reconstruct SAR satellite data within the time window before and after the event trigger time;

[0012] Compensate for time deviation of geostationary meteorological satellite data using linear weighted fusion method;

[0013] Calculate the spatial position compensation amount according to satellite orbit parameters and sea current velocity data;

[0014] The data fusion module is connected to the space-time calibration module to generate a three-dimensional geographic data cube containing backscattering coefficient, infrared temperature field and chlorophyll concentration gradient field, and calculate the diffusion velocity field of ocean phenomena.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] 1. Integrate and receive multi-source data as a standard, so as to monitor the ocean state from multiple dimensions, obtain multi-dimensional ocean environment state information, and effectively improve the monitoring effect of ocean state monitoring compared with the monitoring of a single type of data source;

[0017] 2. Realize dynamic monitoring of the ocean state through real-time data, so compared with the traditional monitoring methods of regular and fixed-point sampling, the timeliness of the monitoring results is effectively improved; set a threshold as the trigger condition of time to ensure the accuracy of the results and avoid false alarms and missed reports;

[0018] 3. Use historical clock deviation data as input to train and obtain a time series prediction model, which can predict the clock offset in response to an event trigger signal, and fully consider the influencing factors of multi-source data for the correction of the event trigger time. Compared with the traditional fixed calibration method, the correction result is more accurate;

[0019] 4. Use the clock offset to calibrate multi-source data, and after calibration, fuse multi-source data to generate a three-dimensional geographic data cube, and ensure the correlation between data through spatial and temporal alignment. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The disclosure of the present invention will be described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0021] Figure 1 is the module composition diagram of the present invention;

[0022] Figure 2 is the working schematic diagram of the abnormal feedback module of the present invention;

[0023] Figure 3 is the spatial deviation processing flow chart of the present invention. Detailed implementation manners

[0024] It is easy to understand that according to the technical solution of the present invention, without changing the essence of the present invention, those of ordinary skill in the art can propose various interchangeable structural manners and implementation manners. Therefore, the following detailed implementation manners and the accompanying drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as all of the present invention or as a limitation or restriction on the technical solution of the present invention.

[0025] Embodiment

[0026] Application overview:

[0027] The dynamic monitoring of the marine environment is of crucial significance to many fields such as marine resource development, ecological protection, and disaster warning. With the continuous development of technology, multi-source remote sensing technology has gradually become an important means to obtain marine environment information, and artificial intelligence technology has also demonstrated powerful capabilities in data processing and analysis. However, in the field of marine environment dynamic monitoring technology based on the collaboration of artificial intelligence and multi-source remote sensing, there are still many problems to be solved urgently.

[0028] Traditional monitoring means mainly rely on a single remote sensing data source or local buoy observations, and there are the following technical bottlenecks:

[0029] Insufficient multi-source data collaboration: Existing systems are difficult to effectively integrate heterogeneous data with different spatio-temporal resolutions and observation mechanisms (such as active microwave SAR, passive infrared remote sensing, and in-situ buoy data), resulting in limited early recognition accuracy of marine phenomena (such as red tides and oil spills). For example, although SAR data can capture micro-scale changes on the sea surface, it lacks biochemical parameters; while chlorophyll fluorescence signals can indicate algal blooms, but are easily restricted by sampling frequency and spatial coverage.

[0030] Spatio-temporal asynchrony problem: There are inherent clock biases (such as differences in satellite overpass times) and spatial registration errors (caused by ocean current movements or sensor perspective differences) in multi-platform observation data. Existing calibration methods mostly use static offset compensation and cannot meet the real-time monitoring requirements in a dynamic marine environment. Especially in sudden events (such as the rapid spread of red tides), uncorrected spatio-temporal biases will significantly reduce the reliability of data fusion.

[0031] Event response lag: Conventional monitoring systems rely on fixed-period data analysis and are difficult to capture the mutation characteristics of marine parameters. For example, a sudden increase in chlorophyll concentration often accompanies the initial stage of algal blooms, but traditional threshold detection methods may result in insufficient warning timeliness due to calculation delays.

[0032] In view of the above defects in the prior art, the basic idea of this application is to achieve efficient collaboration and dynamic analysis of multi-source marine observation data by constructing an intelligent data processing flow.

[0033] The specific implementation process is as follows: The parallel acquisition architecture is adopted to synchronously receive heterogeneous data from SAR satellites, geostationary meteorological satellites, and buoy arrays, forming the comprehensive observation ability for sea surface micro-scale features, temperature fields, and biochemical parameters. Based on the above multi-source heterogeneous data, an event-driven processing mechanism is adopted, and the change rate of chlorophyll concentration is monitored in real time through the sliding window difference algorithm. When an abnormal mutation is detected, the subsequent processing process is immediately triggered, thereby effectively improving the response speed of the system to abnormal ocean states.

[0034] For the processing of multi-source data, artificial intelligence technology is introduced to construct a spatio-temporal calibration system. The clock offset of each data source is dynamically predicted by using the time series prediction model trained with historical data. The SAR data, infrared data, and spatial positions are comprehensively calibrated through the clock offset. Finally, a three-dimensional geographical data cube containing multiple parameters is constructed, and the diffusion velocity field of ocean phenomena is calculated to provide high-precision dynamic analysis ability for ocean environment monitoring.

[0035] The integration of artificial intelligence technology and multi-source remote sensing observations provides an innovative technical path for dynamic ocean environment monitoring.

[0036] After introducing the basic concept of the present invention, the embodiments of the present invention will be specifically introduced below with reference to the accompanying drawings.

[0037] As Figure 1 shown, the dynamic ocean environment monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing includes

[0038] a data acquisition module, which is mainly used to realize the acquisition of multi-source remote sensing data. The module can receive the backscattering coefficient matrix from SAR satellites, the infrared radiance temperature sequence from geostationary meteorological satellites, and the chlorophyll fluorescence pulse signal stream from buoy arrays in parallel.

[0039] For the three types of multi-source data, different dimensions of the ocean can be monitored respectively;

[0040] Among them, the SAR satellite is used to obtain the sea surface backscattering coefficient matrix, and this data reflects the sea surface roughness (which can be used to detect oil spills, internal waves, etc.), corresponding to the monitoring of the sea surface micro-scale dimension;

[0041] The geostationary meteorological satellite can continuously receive the infrared radiance temperature sequence, and this data monitors the anomalies of sea surface temperature (SST) (such as red tides, warm eddies), corresponding to the monitoring of the temperature field dimension;

[0042] The buoy array can transmit the chlorophyll fluorescence pulse signal stream in real time, and this data can directly reflect the concentration of phytoplankton (a key indicator for red tide early warning), corresponding to the monitoring of the biochemical dimension.

[0043] An event detection module, connected to the data acquisition module, is used to perform sliding window difference calculation on the chlorophyll fluorescence pulse signal stream of the buoy array, and generate an event trigger signal when the change rate of chlorophyll concentration between adjacent windows exceeds a preset threshold.

[0044] Calculate the change rate of chlorophyll concentration in the buoy signal with a fixed time window. When the change rate exceeds the threshold, generate a trigger signal with a timestamp to mark potential ecological events (such as red tide outbreaks, algal bloom migrations), and the time point corresponding to the timestamp is the event trigger time.

[0045] The sliding window difference calculation of the event detection module specifically includes: setting a sliding time window with a length of n seconds, calculating the first-order difference value ΔC of the chlorophyll fluorescence intensity within the window; and activating the event trigger signal when ΔC exceeds the preset threshold for m consecutive windows; where m and n are both positive integers.

[0046] The time window and the preset threshold are both set manually and are adjusted in real time for different events, in different seasons, and in different sea areas; for example, during the high-incidence period of red tides, the preset threshold is reduced to 80% of the preset threshold during normal times.

[0047] Example:

[0048] Rapid red tide outbreak: The value of the time window is set to: n = 30 seconds, and the short window is used to capture mutant data.

[0049] Seasonal algal bloom: The value of the time window is set to: n = 300 seconds, and the long window smooths the noise.

[0050] The calculation process of the first-order difference value ΔC is:

[0051] ΔC = C t -C t-n where C t is the mean value of the chlorophyll fluorescence intensity of the current window, and C t-n is the mean value of the previous window. The first-order difference value ΔC can directly reflect the change rate of phytoplankton biomass. A negative value indicates the death of algae, and a positive value indicates the proliferation of algae.

[0052] The threshold trigger event adopts a dual-condition judgment method:

[0053] Condition 1: ΔC > threshold;

[0054] Condition 2: Condition 1 is satisfied for m consecutive windows.

[0055] By setting Condition 2, short-term noise (such as instantaneous disturbances caused by ship wakes) is filtered to ensure the accuracy of the monitoring results and improve the reliability of event detection.

[0056] Example:

[0057] The preset threshold is set to 0.3 μg / L / s, corresponding to the bi-optical characteristic threshold for red tide outbreak;

[0058] In the initial stage of red tide outbreak, if more than 50% of the nodes in the buoy array trigger signals within 3 windows, it is determined to be valid.

[0059] The clock offset prediction module, connected to the event detection module, adopts a time series prediction model trained with historical clock deviation data, and outputs the clock offsets of each data source within a future predetermined time period in response to the event trigger signal;

[0060] The time series prediction model is a dual-channel LSTM neural network, where:

[0061] The first channel inputs the historical clock deviation sequences of each data source, including the temperature drift and orbit perturbation;

[0062] The temperature drift is the clock deviation caused by each degree Celsius of satellite electronic devices, from the temperature-frequency characteristic curve of the satellite thermal control system (such as a clock deviation of 0.3 μs / s for every 1 °C increase).

[0063] The orbit perturbation includes the Doppler effect compensation caused by the change in altitude angle, obtained by integrating the angular acceleration recorded by the gyroscope, which affects the short-term stability of the spaceborne atomic clock (such as ±0.1 μs jitter caused by vibration noise).

[0064] The second channel receives the event trigger signal as the activation condition for time attention gating.

[0065] The functional purpose of this module is to eliminate the time inconsistency caused by satellite transmission and buoy clock drift, and can solve the problem of multi-platform clock asynchrony.

[0066] The specific process is as follows:

[0067] When the second channel receives the event trigger signal, the hidden state weight corresponding to the time step in the first channel is dynamically enhanced;

[0068] The enhancement method is: h′ t = h t · σ (W a · S trigger + b a );

[0069] Among them, h t represents the original hidden state of the LSTM at time step t, h′ t represents the original hidden state of the dynamically enhanced LSTM at time step t, representing the time series characteristics of storing historical clock deviations (such as temperature drift trend, orbit vibration mode);

[0070] S trigger is an event trigger signal, whose value is 0, 1 or a probability value, and is a pulse signal sent when the buoy detects a chlorophyll mutation;

[0071] W a is a trainable weight matrix used to control the influence intensity of the event signal on different hidden state components and is learned through backpropagation.

[0072] b a is a trainable bias term used to adjust the basic attention level and prevent the complete ignoring of the hidden state when not triggered.

[0073] σ is the Sigmoid activation function that compresses the attention weights into the interval (0, 1), and the formula is:

[0074] The following is illustrated with examples:

[0075] When the event is not triggered, S trigger = 0;

[0076] The formula degenerates to: h′ t = h t ·σ(b a );

[0077] At this time, the basic LSTM time series prediction function is maintained, and the bias term b a ensures that the model can still learn the long-term law of clock drift, such as the periodic deviation caused by the satellite orbit period.

[0078] If the event persists but does not reach the threshold, it is a soft trigger mode; at this time, a probabilistic activation signal (such as 50% intensity) is transmitted through pulse width modulation (PWM). It is applicable to suspected abnormal events, such as when the chlorophyll change rate is close to the threshold, and a smooth transition is achieved through probabilistic triggering to avoid jumps in the prediction results.

[0079] In the hard trigger mode, when the buoy detects a chlorophyll mutation (ΔC > threshold), a high-level signal is sent to the second channel, and the attention mechanism is immediately activated, that is, S trigger = 1;

[0080] The formula expands to: h′ t = h t ·σ(W a + b a );

[0081] At this time, the weight matrix W a learns the key features. For example, when a red tide outbreak causes a mutation in the buoy signal, the model automatically strengthens the weights of the hidden states 30 minutes before and after the current moment (increased by 3 to 5 times).

[0082] After the prediction model outputs the clock offset, the event trigger timestamp is automatically corrected (for example, if the deviation is +2 μs, the trigger time is adjusted forward by 2 μs).

[0083] Example:

[0084] The input data is the clock temperature drift sequence of HY-1C satellite (sampling rate 1 Hz);

[0085] At 09:47, the buoy detects that the chlorophyll change rate ΔC = 0.6 μg / L / s > the threshold = 0.5, and sends S trigger = 1.

[0086] Attention activation:

[0087] The weight of the 47th hidden unit changes from 0.38 to 0.91 (an increase of 140%), corresponding to the temperature mutation sensitive neuron.

[0088] Prediction effect:

[0089] The maximum clock deviation in the next 3 hours is predicted to be 1.8 μs (the actual measured value is 1.6 μs, and the error < 12%)

[0090] The data time alignment accuracy is improved from ±5 μs to ±0.7 μs, and the inversion accuracy of the red tide diffusion speed is increased by 22%.

[0091] The spatio-temporal calibration module, connected to the clock offset prediction module and the data acquisition module, performs the following operations:

[0092] Interpolate and reconstruct the SAR satellite data within the time window before and after the event trigger moment;

[0093] Compensate the time deviation of the geostationary meteorological satellite data by using the linear weighted fusion method;

[0094] Calculate the spatial position compensation amount according to the satellite orbit parameters and the sea current speed data.

[0095] The main function of this module is mainly aimed at the different acquisition frequencies and clocks of different data sources. By processing the SAR satellite and geostationary meteorological satellite data, the time is calibrated according to the clock offset to ensure that each data reflects the ocean environment in the same time period; at the same time, considering the influence of satellite orbit changes and sea currents on the buoy, the spatial position compensation amount is calculated, the data spatial position is corrected, and a unified spatial reference is established.

[0096] After calibration, the spatio-temporal accuracy of the data is improved, enabling the data of each layer of the three-dimensional geographic data cube to be accurately aligned and truly reflecting the ocean environment. It provides a reliable basis for data fusion, facilitates the algorithm to explore data connections, and accurately calculates the diffusion speed of ocean phenomena.

[0097] Next, the following explanations will be given for the three operations respectively:

[0098] The specific process of interpolating and reconstructing SAR satellite data within the time window before and after the event trigger time is as follows:

[0099] Receiving the clock offset:

[0100] The spatio-temporal calibration module receives the predicted clock offset from the clock offset prediction module. This clock offset reflects the deviation between the SAR satellite clock and the system standard clock.

[0101] Calculating the reference time:

[0102] The system records the event trigger time and uses it as a reference. The event trigger time is usually the timestamp determined according to the abnormal situation monitored by the chlorophyll fluorescence pulse signal stream of the buoy array.

[0103] Determining the start and end times of the time window:

[0104] Adjust the reference time according to the received clock offset.

[0105] Example:

[0106] Suppose the event trigger time is t trigger , and the clock offset is Δt;

[0107] Then the corrected time trigger time t corrected = t trigger + Δt.

[0108] If the buoy is triggered at 12:00:00 and the satellite clock is 1.5 μs fast, the corrected time is 12:00:00.0015.

[0109] The length of the time window is set to five minutes before and after the event trigger time. Centered on t corrected , it extends 5 minutes forward and backward each, forming an interval [t corrected - 300s, t corrected + 300s].

[0110] The cubic spline interpolation algorithm is used for interpolation and reconstruction. Within the determined time window [t corrected - 300s, t corrected + 300s], the SAR satellite data is reconstructed by cubic spline interpolation according to the preset interpolation step (such as 50 ms).

[0111] In this way, the clock offset participates in the calculation of the start and end times of the time window, ensuring the accuracy of the time window, enabling the interpolated and reconstructed data to be carried out within the accurate time range, and more truly reflecting the changes in the ocean environment during this time period.

[0112] The implementation process of compensating for time deviation by using linear weighted fusion for geostationary meteorological satellite data is as follows:

[0113] Obtain the clock offset:

[0114] The clock offset predicted by LSTM (such as the satellite clock is 2 seconds fast), correct the event trigger timestamp:

[0115] Assume the event trigger time is t trigger , and the clock offset is Δt;

[0116] Then the corrected time trigger time t corrected = t trigger + Δt.

[0117] Locate adjacent frames:

[0118] Find the two nearest frames of data Frame corrected and Frame i and Frame i+1 ;

[0119] Calculate the time difference t between the corrected time trigger time t corrected and the previous frame:

[0120]

[0121] Calculate the ratio of the front and rear frames:

[0122] The weight of the previous frame is The weight of the next frame is

[0123] Example: Himawari-8 infrared data, time interval T = 10 minutes; t corrected = 12:03, then t = 3 minutes.

[0124] The weight of the previous frame is: The weight of the next frame is:

[0125] Fusion effect:

[0126] The 12:00 frame (SST = 21.3 °C), 12:10 frame (SST = 21.7 °C), weighted by 0.7:0.3 to obtain the SST = 21.42 °C at the trigger time, with an error < 0.1 °C from the buoy measured value.

[0127] Calculating the spatial position compensation amount according to the satellite orbit parameters and sea current velocity data includes:

[0128] Calculating the sub-satellite point drift amount according to the orbital altitude and velocity vector of the SAR satellite;

[0129] Calculate the dynamic offset of the buoy based on the sea current velocity vector at the location of the buoy.

[0130] Among them, the specific process of calculating the subsatellite point drift based on the orbital altitude and velocity vector of the SAR satellite is as follows:

[0131] Calculate the subsatellite point drift based on the orbital altitude and velocity vector of the SAR satellite:

[0132] Denote the orbital altitude of the SAR satellite as h, and denote the velocity vector of the SAR satellite as Denote the radius of the earth as R e ;

[0133] According to the above parameters, first calculate the angular velocity ω of the subsatellite point:

[0134] The originally calculated time difference is Δk0. After considering the clock offset Δt, the actual time difference Δk used for calculation is Δk = Δk0 + Δt.

[0135] Then calculate the ground drift distance through the time difference Δk obtained after considering the clock offset Δt:

[0136] Δx = ω·(R e + h)·Δk·cos(θ), where θ is the satellite side view angle and Δx is the subsatellite point drift.

[0137] Finally, reverse translate the SAR image pixel coordinates by Δx to eliminate the geometric distortion caused by satellite movement.

[0138] Example:

[0139] The satellite is Sentinel-1, the orbital altitude h is 514 km, and the velocity vector is 7.6 km / s. After considering the clock offset Δt, the SAR image acquisition time interval Δk is 5 seconds, and the radius of the earth R e is 6371 km. The default satellite side view angle of Sentinel-1 is 34°.

[0140] Substituting the above calculations, we get:

[0141] ω ≈ 1.04×10 -3 rad / s;

[0142] Δx ≈ 38 m (along the orbit direction).

[0143] Just reverse translate the SAR image pixel coordinates according to the calculated Δx.

[0144] The specific process of calculating the dynamic offset of the buoy based on the sea current velocity vector at the location of the buoy is as follows:

[0145] Denote the sea current speed as The time difference between the buoy data and the SAR acquisition time is Δp0. After considering the clock offset Δt, the actual time difference Δp used for calculation is Δp = Δt + Δp0;

[0146] Calculate the offset vector:

[0147]

[0148] Convert the offset from the geographic coordinate system to the SAR image coordinate system (considering the UTM projection deformation);

[0149] Before data fusion, correct the positions of data such as the chlorophyll concentration reported by the buoy.

[0150] Example: The sea current speed of the HYCOM model / buoy GPS is 0.3 m / s (direction 120°);

[0151] After considering the clock offset Δt, the time difference Δp between the buoy data and the SAR acquisition time is 10 minutes;

[0152] Then it can be calculated that: Northeast coordinate system.

[0153] Actual usage example:

[0154] SAR data: GF-3 image (resolution 10 m)

[0155] Compensation effect:

[0156] Sub-satellite point drift correction 38.2 m → the positioning error of the internal wave front is reduced from 45 m to 6 m;

[0157] Buoy offset correction 122 m → the alignment degree between the chlorophyll peak and the SAR front is increased by 82%.

[0158] The data fusion module is connected to the spatio-temporal calibration module to generate a three-dimensional geographic data cube containing the backscattering coefficient, infrared temperature field, and chlorophyll concentration gradient field, and calculate the ocean phenomenon diffusion velocity field.

[0159] This module mainly integrates multi-source data such as SAR satellites, meteorological satellites, and buoy arrays into a three-dimensional geographic data cube, comprehensively and intuitively presenting multi-dimensional information of the ocean environment, and realizing the deep fusion of multi-modal data. And based on the calibrated spatio-temporal data, calculate the ocean phenomenon diffusion velocity field, grasp the development trend in real time, predict the diffusion direction and range, and assist in disaster warning and ecological protection.

[0160] The construction process of the three-dimensional geographic data cube is as follows:

[0161] Establish a 100m×100m geographic grid under the WGS84 coordinate system:

[0162] First, obtain the approximate geographic range of the monitoring area, which can be determined by the boundary information of satellite remote sensing images or the pre-set range of the monitoring area;

[0163] Based on the WGS84 coordinate system, divide the monitoring area into grids of 100m×100m. Each grid cell has a unique geographic coordinate identifier, which is accurately represented by longitude and latitude to indicate its position on the earth.

[0164] These grids serve as the basic spatial units for subsequent data storage and analysis, providing a unified spatial reference framework for data from different data sources.

[0165] Store data in layers in the vertical dimension:

[0166] Surface layer: Normalized SAR backscatter coefficient;

[0167] The SAR satellite is the main data source for obtaining the backscatter coefficient. The SAR emits microwave signals to the ground and receives the signals reflected back from the ground, and calculates the backscatter coefficient based on the intensity and characteristics of the reflected signals. The backscatter coefficient reflects the scattering ability of ground targets (mainly the sea surface in the marine environment) to microwave signals, and is closely related to factors such as the roughness of the sea surface and the height of sea waves.

[0168] The original SAR backscatter coefficient data may be affected by various factors, such as the characteristics of satellite sensors, observation angles, atmospheric conditions, etc., resulting in certain deviations and fluctuations in the data. In order to eliminate these effects, it is necessary to normalize the data.

[0169] The normalization process adopts the following steps:

[0170] First, calculate the statistical characteristics of the SAR backscatter coefficient data, such as the mean and standard deviation. Then, for each data point, use the formula for normalization, where x is the original backscatter coefficient value, μ is the mean of the data, and σ is the standard deviation of the data. In this way, all data is converted to a relatively unified scale, facilitating subsequent analysis and comparison.

[0171] The specific implementation is as follows:

[0172]

[0173] Middle layer: Atmospherically corrected infrared temperature field;

[0174] Infrared temperature field data are usually obtained by infrared sensors carried by meteorological satellites. The infrared sensors measure temperature by detecting the infrared radiation emitted from the Earth's surface (including the ocean surface). Objects with different temperatures emit infrared radiation with different intensities, and the sensor inversely calculates the surface temperature based on the intensity of the received infrared radiation signal.

[0175] During the transmission of infrared radiation from the Earth's surface to the satellite sensor, it will be affected by the absorption and scattering of various gases in the atmosphere (such as carbon dioxide, water vapor, etc.), resulting in attenuation and distortion of the signal received by the sensor, thereby affecting the accuracy of temperature measurement. Therefore, atmospheric correction is required.

[0176] The specific implementation is as follows in the table:

[0177] Attribute Technical details Data effect Data source Himawari-8 / FY-4A satellite 11μm band Sea Surface Temperature (SST) monitoring Atmospheric correction <![CDATA[Split window algorithm: SST = T 11 + A(T 11 - T 12 ) + B]]> Eliminating the influence of water vapor absorption Accuracy ±0.3℃ (under clear sky conditions) Warm eddy and upwelling identification

[0178] In the above table, A is the weight coefficient of atmospheric water vapor absorption, which is used to quantify the sensitivity of the radiation difference between the 11μm and 12μm bands to water vapor interference, and the typical value range is 0.8 - 1.5. B is the system deviation correction term, which is used to compensate for atmospheric residual errors and instrument system deviations, and the typical value range is -2.0 - 2.0°C.

[0179] Bottom layer: Chlorophyll concentration gradient field converted based on fluorescence intensity.

[0180] Fluorescence intensity data are mainly obtained by fluorescence sensors carried on buoy arrays. Chlorophyll is an important pigment in marine phytoplankton. When excited by light of a specific wavelength, chlorophyll will emit fluorescence, and the fluorescence intensity is closely related to the chlorophyll concentration.

[0181] To convert fluorescence intensity data into chlorophyll concentration, an empirical formula or a machine learning-based model is usually used. The following formula is used for processing, and the formula is Chl-a = a·ln(F) 3 +b·ln(F) 2 +c·ln(F)+d, where Chl-a represents the chlorophyll concentration, F is the fluorescence intensity value calibrated by the baseline, and the coefficients a, b, c, d are obtained by fitting field sampling data.

[0182] In practical applications, first, the original fluorescence intensity data obtained by the fluorescence sensor are calibrated by the baseline to remove the influence of background noise and instrument errors. Then, the calibrated fluorescence intensity value is substituted into the above formula to calculate the corresponding chlorophyll concentration. By analyzing and processing the chlorophyll concentration data at different positions and times, a chlorophyll concentration gradient field is obtained, which reflects the distribution and changes of phytoplankton in the ocean.

[0183] The specific implementation is as follows in the table:

[0184]

[0185] The conversion process of the chlorophyll concentration gradient field is as follows:

[0186] First, it is necessary to determine the fluorescence intensity value F calibrated by the baseline. The specific process is as follows:

[0187] The fluorometer carried by the buoy (such as the WET Labs ECO-FLRTD) measures the 685nm fluorescence pulse signal in real time.

[0188] Deduct the instrument electronic noise (measure the reference value under the condition of no light at night), and use the data of the 700nm channel to eliminate the scattering interference of suspended substances in the water body (formula: F cal = F 685 - 0.8×F 700 ); Correct the change of fluorescence quantum efficiency according to the data of the built-in temperature sensor.

[0189] Calculate the chlorophyll concentration according to the formula Chl-a = a·ln(F) 3 + b·ln(F) 2 + c·ln(F)+d;

[0190] a, b, c, d are obtained by fitting the on-site sampling data, and Chl-a is the chlorophyll concentration.

[0191] ln(F) is the natural logarithm conversion, which compresses the dynamic range and improves the sensitivity in the low concentration area.

[0192] Input the fluorescence intensity value F calibrated by the baseline;

[0193] Calculate the chlorophyll concentration according to the formula Chl-a = a·ln(F) 3 + b·ln(F) 2 + c·ln(F)+d;

[0194] Among them, a, b, c, d are obtained by fitting the on-site sampling data, and Chl-a is the chlorophyll concentration.

[0195] The process of coefficient fitting is as follows:

[0196] Synchronously collect the water sample (Niskin water sampler) and the fluorometer reading;

[0197] Determine the true chlorophyll concentration by the acetone extraction method;

[0198] Fit the polynomial coefficients by the least square method.

[0199] The generation process of the chlorophyll concentration gradient field is as follows:

[0200] It is implemented by using the Sobel operator. The mathematical expression of the convolution kernel is:

[0201] matrix Similarly, they are respectively denoted as G x and G y .

[0202] The gradient vector output is:

[0203]

[0204] The modulus length of is determined to be the red tide front when > 0.3 μg / L / m.

[0205] θ = arctan(G y / G x ), which is used for the algae aggregation direction.

[0206] Example:

[0207] Input the Chl-a values of a 3×3 grid:

[0208] Gradient calculation:

[0209] G x = (8.1 + 2×15.7 + 11.2) / 8 - (5.2 + 2×4.9 + 3.5) / 8 = 2.44;

[0210] G y = (3.5 + 2×9.8 + 11.2) / 8 - (5.2 + 2×6.8 + 8.1) / 8 = -0.31;

[0211] Modulus length Far exceeds the threshold, and it is determined to be a front.

[0212] The acquisition process of the ocean phenomenon diffusion velocity field is as follows:

[0213] Obtain the chlorophyll concentration gradient field after spatio-temporal calibration of two consecutive frames according to the fused three-dimensional geographical data cube:

[0214] Calculate the spatial displacement vector field based on the Horn-Schunck algorithm

[0215] Perform velocity vector correction by combining the clock offset output by the clock offset prediction module:

[0216] The clock offset Δt output by the clock prediction module;

[0217] The velocity correction formula is:

[0218] Where ΔT is the satellite revisit period, and grid_size represents the spatial grid resolution, which is fixed at 100 meters in this formula; this parameter directly correlates the velocity correction with the spatial scale, ensuring that the time deviation compensation has a clear corresponding relationship in physical space.

[0219] When Δt is positive, that is, the satellite clock is fast; the denominator takes ΔT - Δt, shortening the effective observation time;

[0220] When Δt is negative, that is, the satellite clock is slow; the denominator takes ΔT + Δt, shortening the effective observation time;

[0221] Example:

[0222] The red tide diffusion velocity v original = 0.15 m / s, the clock offset Δt = +0.85 s, that is, the satellite clock is fast, and ΔT = 600 s.

[0223] Correction calculation:

[0224]

[0225] Comparison of correction results:

[0226] Scene Original velocity Δt value Corrected velocity Rate of change Ecological significance Red tide diffusion 0.15m / s +0.85s 0.152m / s +1.3％ More accurately predicting the arrival time of the algal front Oil spill diffusion 0.08m / s -0.5s 0.079m / s -1.2％ Optimizing the simulation accuracy of the oil film drift trajectory

[0227] Actual application example:

[0228] Input data: Two-frame fusion cube at 08:00 and 08:10 (Chl-a peak from 12 → 18 μg / L);

[0229] Optical flow output: Frontal movement vector 0.2 m / s (direction NE 35°), with an observation error of <5% compared to the drone.

[0230] Early warning application: Predicting the impact on the aquaculture area after 6 hours and initiating protective measures in advance.

[0231] As Figure 2 shown, it also includes an anomaly feedback module for:

[0232] Monitoring the deviation value between the output data of the spatio-temporal calibration module and the measured data of the buoy; the deviation value includes time deviation and spatial deviation;

[0233] When the time deviation exceeds the preset threshold, trigger the online fine-tuning of the clock offset prediction model, and the fine-tuning method is:

[0234] Retain a copy of the original model parameters;

[0235] Perform gradient descent optimization using the real-time data of the 30 minutes before and after the current moment;

[0236] Update the model parameters when the loss function of the validation set drops by more than 10%;

[0237] When the spatial deviation exceeds the preset threshold, online fine-tuning of the spatio-temporal calibration module is triggered, and the spatio-temporal calibration module recalculates the sub-satellite point drift and the buoy dynamic offset.

[0238] The module plays the role of error closed-loop control. As the "quality controller" of the system, by monitoring the deviation between the data after spatio-temporal calibration and the measured value of the buoy, it dynamically optimizes the upstream model parameters to form a "monitoring - feedback - correction" closed loop. It independently processes time deviation (optimization of the clock prediction model) and spatial deviation (recalculation of calibration parameters) to ensure that the system continuously maintains sub-second time synchronization and sub-pixel spatial alignment.

[0239] The following separately explains the time deviation and the spatial deviation:

[0240] Time deviation:

[0241] Retain a copy of the original model parameters;

[0242] The purpose is to retain a stable "benchmark" during the fine-tuning process. Retaining a copy of the original parameters can ensure that in the fine-tuning process, if unsatisfactory results occur, it can roll back to the original state to avoid a significant decline in model performance due to over-adjustment. At the same time, the copy of the original parameters can also be used as a comparison reference to help evaluate the effect of fine-tuning.

[0243] Perform gradient descent optimization using the real-time data of 30 minutes before and after the current moment;

[0244] The real-time data of 30 minutes before and after the current moment is closely related to the moment when the fine-tuning is triggered and can reflect the latest situation of the clock offset under the current ocean environment state. A shorter time window can ensure the timeliness of the data, enabling the model to quickly adapt to the dynamic changes of the ocean environment. At the same time, the data volume of 30 minutes is also sufficient to contain enough information to capture the characteristics and laws of the clock offset, avoiding inaccurate model adjustment due to too little data or unnecessary computational burden due to too much data. In the actual execution process, the selected 30-minute real-time data is input into the clock offset prediction model, and the loss between the model prediction result and the actual value is calculated. Then, the gradient with respect to the model parameters is calculated according to the loss function.

[0245] Update the model parameters when the loss function of the validation set drops by more than 10%;

[0246] The validation set is a part of data divided from real-time data and is used to evaluate the performance of the model during the fine-tuning process. Setting the condition that the loss function of the validation set drops by more than 10% for updating the model parameters is to ensure that the fine-tuning of the model actually brings about an improvement in performance. If the decrease in the loss function of the validation set is not obvious, it indicates that the adjustment of the model may not effectively improve its ability to predict the actual clock offset. At this time, not updating the model parameters can avoid introducing unnecessary errors. Only when the loss function of the validation set drops by more than 10% is it considered that the performance of the model has been significantly improved after fine-tuning and it is worth updating the model parameters, so that the entire system can more accurately predict the clock offset and improve the accuracy of time calibration.

[0247] When the loss function of the validation set meets the condition of dropping by more than 10%,

[0248] the fine-tuned model parameters are formally updated into the clock offset prediction model, and the new parameters will be used for subsequent clock offset prediction, so that the model can better adapt to the dynamic changes of the ocean environment and maintain high-precision calibration in the time dimension.

[0249] Example:

[0250] Buoy #07, with a time deviation of +1.2 μs, exceeding the threshold;

[0251] The fine-tuning results are shown in the following table:

[0252] Index Before fine-tuning After fine-tuning Improvement Predicted MAE 0.8μs 0.3μs 62％ Red tide warning delay 45s 12s 73％

[0253] Spatial deviation, the process is as Figure 3 shown:

[0254] Spatial deviation calculation:

[0255]

[0256] In the formula, x calib is the calibrated target longitude coordinate, x buoy is the measured longitude coordinate of the buoy, y calib is the calibrated target latitude coordinate, y buoy is the measured latitude coordinate of the buoy, and Δf is the spatial position deviation.

[0257] Sub-satellite point drift update:

[0258] Δx = v sat ·Δt·cosθ, in the formula, v sat is the satellite orbital velocity, Δt is the time deviation amount, θ is the satellite side view angle, and Δx is the horizontal drift amount of the sub-satellite point.

[0259] Buoy dynamic offset correction:

[0260] In the formula, is the offset at the previous moment, v current is the three-dimensional ocean current velocity, and Δt′ is the time interval. is the corrected offset.

[0261] Example:

[0262] Initial data:

[0263] Satellite calibration coordinates: (122.352°E, 30.521°N);

[0264] Measured buoy coordinates: (122.351°E, 30.522°N);

[0265] Ocean current velocity: 0.3 m / s (eastward), -0.1 m / s (northward);

[0266] θ is set to 34°, v sat is set to 7600 m / s, Δt is set to +0.003 s, Δt′ is set to 60 s, is set to 5.2 m, v current is (0.3, -0.1) m / s.

[0267] Calculate the spatial deviation:

[0268] Exceed the inshore threshold (2 m), trigger correction.

[0269] Update the subsatellite point drift:

[0270] Δx = 7600 × 0.003 × cos34° ≈ 19.3 m.

[0271] Correct the buoy offset:

[0272]

[0273] Final output:

[0274]

[0275] Effect comparison:

[0276] Index Error before correction Error after correction Position deviation 111.2m 0.8m Red tide front positioning Deviation of 3 grids Deviation less than 0.1 grid

[0277] Through the closed-loop correction driven by multi-source data fusion and physical models, this process reduces the spatial positioning error from the hundred-meter level to the sub-meter level.

[0278] The technical scope of the present invention is not limited to the content described above. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of the present invention, and these deformations and modifications shall fall within the protection scope of the present invention.

Claims

1. An ocean environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing, characterized in that: including a data acquisition module, configured to receive in parallel a backscattering coefficient matrix from a SAR satellite, an infrared radiation brightness temperature sequence from a geostationary meteorological satellite, and a chlorophyll fluorescence pulse signal stream from a buoy array; an event detection module, electrically connected to the data acquisition module, configured to perform sliding window differential calculation on the chlorophyll fluorescence pulse signal stream of the buoy array, and generate an event trigger signal when the change rate of chlorophyll concentration between adjacent windows exceeds a preset threshold; a clock offset prediction module, electrically connected to the event detection module, adopting a time series prediction model trained with historical clock deviation data, and outputting the clock offsets of each data source within a future predetermined time period in response to the event trigger signal; a spatio-temporal calibration module, connected to the clock offset prediction module and the data acquisition module, and performing the following operations after obtaining the clock offsets: interpolating and reconstructing the SAR satellite data within a time window before and after the event trigger moment; compensating the time deviation of the geostationary meteorological satellite data by using a linear weighted fusion method; calculating a spatial position compensation amount according to satellite orbital parameters and sea current velocity data; a data fusion module, connected to the spatio-temporal calibration module, generating a three-dimensional geographical data cube including a backscattering coefficient, an infrared temperature field, and a chlorophyll concentration gradient field, and calculating an ocean phenomenon diffusion velocity field.

2. The marine environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing according to claim 1, characterized in that: The sliding window differential calculation of the event detection module specifically includes: setting a sliding time window with a length of n seconds, calculating the first difference value ΔC of the chlorophyll fluorescence intensity within the window; and activating the event trigger signal when ΔC exceeds the preset threshold for m consecutive windows; where m and n are both positive integers.

3. The marine environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing according to claim 1, characterized in that: The time series prediction model is a dual-channel LSTM neural network, where: The first channel inputs the historical clock deviation sequences of each data source, including the temperature drift amount and the orbital perturbation amount; the temperature drift amount is the clock deviation caused by each °C of the satellite electronic device, and the orbital perturbation amount includes the Doppler effect compensation caused by the change in the elevation angle; The second channel receives the event trigger signal as the activation condition for time attention gating.

4. The marine environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing according to claim 1, characterized in that: The process of interpolating and reconstructing the SAR satellite data within a time window before and after the event trigger moment is as follows: Receiving the clock offset output by the clock offset prediction module, and correcting the event trigger moment by using the clock offset; Taking the corrected event trigger moment as a reference, determining the front and back time windows, and interpolating and reconstructing the SAR satellite data within the time window; The interpolation and reconstruction adopts a cubic spline interpolation algorithm.

5. The marine environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing according to claim 1, characterized in that: The weight calculation method of the linear weighted fusion is: Let the time interval between two adjacent frames of meteorological satellite data be, the time difference between the event trigger time after clock offset correction and the previous frame be t, and the weight of the previous frame be The weight of the next frame is 6. The marine environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing according to claim 1, characterized in that: The calculation of the spatial position compensation amount includes: Calculating the sub-satellite point drift amount according to the orbital altitude and velocity vector of the SAR satellite; Calculating the buoy dynamic offset amount based on the sea current velocity vector at the position of the buoy.

7. The marine environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing according to claim 1, characterized in that: The construction of the three-dimensional geographical data cube includes: Establishing a 100m×100m geographical grid in the WGS84 coordinate system; Storing in layers in the vertical dimension: The surface layer: the normalized SAR backscattering coefficient; The middle layer: the infrared temperature field after atmospheric correction; The bottom layer: the chlorophyll concentration gradient field converted based on the fluorescence intensity.

8. The marine environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing according to claim 7, characterized in that: The conversion process of the chlorophyll concentration gradient field is: Input the fluorescence intensity value F calibrated by the baseline; Calculate the chlorophyll concentration according to the formula Chl-a = a·ln(F) 3 +b·ln(F) 2 +c·ln(F)+d Among them, a, b, c, and d are obtained by fitting field sampling data, and Chl-a is the chlorophyll concentration.

9. The marine environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing according to claim 1, characterized in that: The calculation of the diffusion velocity field of the marine phenomenon adopts the optical flow method, specifically including: Extract the chlorophyll concentration gradient field from two consecutive frames of fused data cubes; Calculate the spatial displacement vector field based on the Horn-Schunck algorithm; Perform velocity vector correction by combining the clock offset output by the clock offset prediction module.

10. The marine environment dynamic monitoring system based on the collaboration of artificial intelligence and multi-source remote sensing according to any one of claims 1 to 9, characterized in that: It also includes an anomaly feedback module for: Monitoring the deviation value between the output data of the spatio-temporal calibration module and the measured data of the buoy; the deviation value includes time deviation and spatial deviation; When the time deviation exceeds the preset threshold, trigger the online fine-tuning of the clock offset prediction model, and the fine-tuning method is: Retain a copy of the original model parameters; Perform gradient descent optimization using the real-time data of 30 minutes before and after the current moment; Update the model parameters when the loss function of the validation set drops by more than 10%; When the spatial deviation exceeds the preset threshold, trigger the online fine-tuning of the spatio-temporal calibration module, and the spatio-temporal calibration module recalculates the sub-satellite point drift and the buoy dynamic offset.

Images