Unmanned aerial vehicle fish health monitoring method and system based on marine ranching

Abstract

The application relates to the technical field of ocean ranching monitoring, and discloses a method and system for unmanned aerial vehicle fish health monitoring based on an ocean ranching, which controls a flight platform to lower a sensing unit to a water layer indicated by a dynamic target depth parameter based on electromechanical servo linkage, collects baseline data such as video, multispectral data, dissolved oxygen and flow rate and spatial distribution echo data, extracts video and multispectral features to determine physiological abnormalities, combines with environmental parameters in and outside a net cage to determine a physical retardation condition of a net cover to exclude false alarms caused by oxygen deficiency stress, uses spatial distribution echo data to deduce a biological quantity and deduce a theoretical feeding amount, converts the theoretical feeding amount into a final feeding amount based on an environmental oxygen-carrying bottom line constraint to perform feeding, uses a water surface image to perform visual residual feed inverse operation to deduce a residual feed remaining weight, obtains an actual effective feeding load to perform closed-loop correction on baseline parameters of the system. The application overcomes the defects of fixed detection blind area and single data false alarm, realizes environmental safety constraint and self-adaptive correction of a feeding strategy.

Description

Technical Field

[0001] This invention relates to the field of marine ranching aquaculture monitoring technology, specifically to a method and system for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching. Background Technology

[0002] With the expansion of marine ranching, drones equipped with sensors are increasingly being used to monitor and assess the health status of fish. Existing monitoring systems typically employ a fixed-depth suspended sampling mode or rely on a single type of sensor for data collection. Because farmed fish dynamically move between different water layers according to changes in hydrological conditions such as water temperature, light, and dissolved oxygen, fixed-depth detection methods struggle to pinpoint the actual activity areas of the target group. This leads to discrepancies between the optical images collected by the sensors and the actual water quality data, creating blind spots in observation. In net cage aquaculture environments, organisms attached to the nets can easily obstruct water exchange and cause localized hypoxia within the cages. Fish stressed by this environment often exhibit behavioral changes such as decreased swimming speed or abnormal surfacing. Existing monitoring systems, relying solely on apparent movement characteristics or single images, are prone to misinterpreting this stress behavior caused by low oxygen levels as physiological disorders in fish. The systems lack spatial environmental comparison capabilities, resulting in frequent false alarms.

[0003] In daily feeding management, current automated feeding mechanisms primarily rely on theoretical growth models to output fixed control commands. This open-loop feeding mode fails to consider the actual state of the aquatic environment, ignoring the increased metabolic oxygen consumption of fish after large-scale feeding. When dissolved oxygen levels in the cage water are low, feeding the full amount according to theoretical predictions can easily breach the safe oxygen-carrying limit of the aquatic environment, leading to metabolic hypoxia risks for the fish population. Furthermore, existing feeding processes lack effective quantitative detection methods for residual feed on the water surface after material distribution, making it impossible for the system to obtain data on the actual effective feeding amount of the target group. This lack of a closed-loop feedback mechanism prevents the underlying growth model and feeding parameters from being calibrated with real data, and the cumulative calculation deviations generated over long-term operation ultimately lead to feed waste and bottom pollution. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for monitoring fish health using unmanned aerial vehicles (UAVs) in marine ranching. This solves the problems of existing marine ranching monitoring systems, which typically rely on fixed water layers or single sensors to collect data under complex sea conditions. These systems struggle to dynamically track the water layers where fish are active and are prone to misinterpreting environmental stresses such as low oxygen levels as physiological disorders in fish based on single data characteristics. Furthermore, traditional automatic feeding strategies lack deep coupling with the actual oxygen-carrying capacity of the environment, easily leading to metabolic hypoxia after feeding. Additionally, the lack of a closed-loop correction mechanism for feeding amounts based on actual feeding effects results in feed waste and accumulated errors in the system's baseline parameters.

[0005] To address the above problems, the present invention provides the following technical solution: The first aspect of this invention provides a method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching, comprising the following steps: The control flight platform uses electromechanical servo linkage to lower the onboard sensing unit to the water layer position indicated by the dynamic target depth parameter, and simultaneously collect multi-source heterogeneous reference data and spatial distribution echo data of the target group. The multi-source heterogeneous reference data includes continuous video frames of the target group, multispectral reflectance characteristic data, and dissolved oxygen concentration and water flow velocity values ​​in the same water layer environment. Extracting continuous video frames and multispectral reflectance feature data for dual determination of physiological abnormalities, and coupling dissolved oxygen concentration values ​​with dynamic target depth parameters for analysis, combined with reference environmental parameters outside the cage to establish the final determination of physical obstruction of the netting; The biomass of the target population is inverted using spatially distributed echo data to derive the basic theoretical feeding parameters. Based on the environmental oxygen carrying capacity baseline constraint mechanism, the basic theoretical feeding parameters are transformed into the final execution feeding parameters for feeding. The system uses a surface camera to capture wide-angle images of the water surface and performs visual inverse calculations on the remaining bait to deduce the actual remaining weight of the remaining bait. The actual effective feeding load data is obtained by subtracting the actual remaining weight of the remaining bait from the final feeding amount parameters, and cross-cycle closed-loop correction is performed on the underlying baseline parameters of the system.

[0006] This invention addresses the blind zone problem in fixed-depth sampling by employing an acoustic guidance and electromechanical servo linkage mechanism. Multi-dimensional benchmark data cross-diagnosis is used to eliminate false alarm interference from single sensors. An environmental physical oxygen-carrying baseline constraint is introduced into the feeding decision-making process, and a closed-loop adaptive feedback for the feeding strategy is constructed by combining a visual residual bait inverse calculation mechanism.

[0007] Furthermore, the control flight platform, based on electromechanical servo linkage, lowers the onboard sensing unit to the water layer location indicated by dynamic target depth parameters, including: Using scanning sonar to acquire time-of-flight parameters from spatially distributed echo data, a three-dimensional point cloud spatial distribution feature within the target water area is generated based on the time-of-flight parameters. The entire water body is divided into several equidistant physical depth levels along the vertical descent direction. The number of point clouds contained in each physical depth level is calculated, and the number of point clouds is compared with the spatial volume corresponding to that level to establish the corresponding mapping relationship between vertical spatial depth and biomass density. The physical depth level with the largest biomass density value is extracted, and the center depth value of that level is established as the current dynamic target depth parameter. The difference between the dynamic target depth parameter and the current depth of the sensing unit is calculated, and the difference is converted into a motor rotation control quantity to drive the winch mechanism to move the sensing unit vertically to the water layer position indicated by the dynamic target depth parameter.

[0008] By utilizing the spatial distribution of acoustic three-dimensional point clouds to reflect the longitudinal aggregation characteristics of the target population, and by dynamically tracking high-density water layers through an electromechanical servo system, the physical validity of optical and multispectral sample acquisition is ensured.

[0009] Furthermore, when establishing the corresponding mapping relationship, the frustum-shaped spatial volume corresponding to each physical depth level is calculated by combining the factory acoustic emission angle parameters of the scanning sonar, and the point cloud quantity is compared with the frustum-shaped spatial volume. If multiple physical depth levels have the same biomass density value or the difference is within the set error allowable range, the physical depth level with the smallest depth value is selected as the dynamic target depth parameter. After the drive winch mechanism controls the sensing unit to reach the designated water layer position, a static delay window of a predetermined duration is set to wait for the sensor to physically balance. After the delay ends, a data acquisition trigger signal is sent synchronously to obtain multi-source heterogeneous reference data.

[0010] By combining the acoustic emission angle with geometric compensation for the detection volume at different depths, the accuracy of density calculation is improved; the static delay mechanism can meet the time requirement for the environmental probe to reach chemical reaction stability.

[0011] Furthermore, continuous video frames and multispectral reflectance feature data are extracted for dual determination of physiological anomalies, and dissolved oxygen concentration values ​​are coupled with dynamic target depth parameters for analysis, including: Based on continuous video frames, the decrease rate of the current average swimming speed parameter relative to historical benchmark data is extracted. When abnormal spectral reflectance patches are identified in the multispectral reflectance feature data and the decrease rate exceeds the swimming speed attenuation threshold, a candidate physiological abnormality warning signal is generated. When the dissolved oxygen concentration is lower than the safe dissolved oxygen limit and the dynamic target depth parameter is less than the shallow activity threshold, a preliminary judgment command for netting blockage is generated and the candidate physiological abnormality warning signal is kept in a pending confirmation state.

[0012] Establish a joint identification logic based on movement characteristics and abnormal surface spectra, and set up a disease warning when a shallow hypoxic environment is detected, thus eliminating physiological misjudgments caused by hypoxic stress behaviors.

[0013] Furthermore, the final determination of the physical obstruction of the netting is established by combining external environmental parameters, including: The flight platform is positioned to the free water area on the upstream side of the net cage, and the sensing unit is lowered to a reference water layer with the same depth parameters as the dynamic target to obtain reference dissolved oxygen concentration and reference water flow velocity values. When the reference water flow velocity value is determined to be greater than the set effective flow velocity baseline, the flow velocity attenuation difference ratio between the reference water flow velocity value and the water flow velocity value is calculated, and the absolute concentration difference between the reference dissolved oxygen concentration value and the dissolved oxygen concentration value is also calculated. When the flow velocity attenuation difference ratio is greater than the set flow velocity stagnation threshold and the absolute concentration difference is greater than the set concentration difference threshold, it is determined that there is a low oxygen state inside the net cage caused by the physical obstruction of water exchange by the netting, and a netting physical obstruction warning is generated. If the flow velocity attenuation difference ratio is not greater than the flow velocity stagnation threshold or the absolute concentration difference is not greater than the concentration difference threshold, it is determined that the environmental obstruction is not established, and the candidate physiological abnormality warning signal is confirmed as a formal physiological abnormality warning signal and output.

[0014] A differential verification mechanism of internal and external hydrological parameters was adopted to eliminate external interference caused by the stagnation of background ocean currents and to confirm the physical obstruction of the netting.

[0015] Furthermore, the biomass of the target population is inverted using spatially distributed echo data to derive the basic theoretical feeding parameters. Based on the environmental oxygen-carrying baseline constraint mechanism, the basic theoretical feeding parameters are transformed into the final execution feeding parameters for feeding, including: Based on the target intensity parameters in the spatially distributed echo data, body length characteristic data is inverted and divided into multiple body length intervals. Using an allometric growth model combined with preset morphological empirical constants, the biomass of the subpopulation in each interval is derived. The total biomass parameter is then extracted and multiplied with a dynamic feeding rate parameter set based on water temperature to generate the basic theoretical feeding rate parameter. It is then determined whether the current dissolved oxygen concentration in the water area is greater than the set minimum dissolved oxygen threshold parameter. If not, a control command is directly output to set the final feeding rate parameter to zero, and the current feeding calculation is terminated. If the value is greater than the minimum dissolved oxygen threshold parameter, the usable safe dissolved oxygen concentration is obtained by subtracting the dissolved oxygen concentration value from the dissolved oxygen concentration value. Combined with the effective water volume inside the cage, the remaining oxygen-carrying capacity parameter of the current water area is calculated. The expected metabolic oxygen load is calculated by multiplying the basic theoretical feeding amount parameter with the associated oxygen consumption coefficient per unit mass of feed. When the expected metabolic oxygen load is greater than the remaining oxygen-carrying capacity parameter, the feed restriction cutoff mechanism is triggered. The remaining oxygen-carrying capacity parameter is divided by the associated oxygen consumption coefficient to obtain the maximum safe feeding weight. The maximum safe feeding weight is output as the final feeding amount parameter.

[0016] The oxygen demand caused by feeding is numerically matched and constrained with the actual physical oxygen-carrying capacity of the water body to avoid feeding causing the dissolved oxygen in the water body to fall below the safe threshold.

[0017] Furthermore, by using the surface camera component to capture wide-angle images of the water surface and performing visual residual bait inverse operations, cross-cycle closed-loop corrections are made to the system's underlying baseline parameters, specifically including: After the system completes feeding and the set physical settling delay, the flight platform is controlled to hover at a set altitude. A wide-angle image of the water surface is captured using the surface camera component, and the ground-to-ground flight altitude parameters are recorded simultaneously. After image preprocessing, the wide-angle image is input into the uneaten bait image segmentation model, which outputs a corresponding binarized semantic segmentation mask. The total number of pixels belonging to the foreground is counted. Combining the ground-to-ground flight altitude parameters with the factory calibration parameters of the surface camera component, the actual physical area of ​​the water surface represented by a single pixel in the current image is calculated based on the pinhole imaging geometry principle. The total physical distribution area is obtained by multiplying the total number of foreground pixels by the actual physical area of ​​the water surface. Combined with the average physical cross-sectional area and average weight constant of a single feed pellet, the actual remaining weight of the uneaten bait is derived in reverse. A moving average time series of the actual effective feeding load data is constructed to calculate the system's long-term feeding deviation rate. This long-term feeding deviation rate is then used to perform reverse smoothing compensation on the dynamic feeding rate parameter, feeding activity correction coefficient, or acoustic biomass estimation correction coefficient, completing the adaptive correction of the underlying baseline parameters.

[0018] Establish a calculation process from image pixel area to actual physical weight to achieve actual feeding assessment and baseline correction based on visual inversion.

[0019] Furthermore, the bait residue image segmentation model adopts a fully convolutional neural network architecture. Its input data dimension is a three-channel color image with a fixed resolution, and the output result is a binary semantic segmentation mask representing the physical distribution of residual feed particles. The internal layers of the fully convolutional neural network architecture include an encoder layer consisting of alternating continuous convolutional and pooling layers for extracting multi-scale image features, a decoder layer containing upsampling and convolutional layers, and a skip connection structure connecting the corresponding encoder and decoder layers. The training samples of the bait residue image segmentation model come from a set of manually labeled multi-source water surface bait residue images, and the label data is the corresponding manually delineated binary mask. During network training, the binary cross-entropy loss function is used to iterate the network parameters.

[0020] Define the neural network structure and data processing format to achieve pixel extraction of residual bait in the water surface environment.

[0021] Furthermore, when acquiring actual effective feeding load data, the process includes: calculating the ratio of the actual remaining weight of the bait to the final feeding amount parameter; determining whether the ratio is less than the set minimum residual bait threshold, and simultaneously acquiring the attitude tilt data of the flight platform, which characterizes the current wind, wave, and sea conditions; if the ratio is less than the minimum residual bait threshold, and the attitude tilt data generated to maintain hovering is greater than the safe and stable angle, then the visual sampling data acquired in this instance is deemed invalid, and the residual bait weight inversion result of this instance is not used to calculate the actual effective feeding load data.

[0022] Cross-validation of flight attitude data with visual inversion results prevents data deviations caused by feed being blown away by high winds and waves.

[0023] A second aspect of the present invention provides an unmanned aerial vehicle (UAV) fish health monitoring system based on marine ranching, comprising: Memory is used to store computer program instructions; The processor is configured to implement the unmanned aerial vehicle (UAV) fish health monitoring method based on marine ranching as described in any of the foregoing embodiments when executing computer program instructions.

[0024] Provide the underlying equipment foundation for implementing the above monitoring methods, and realize the corresponding closed-loop monitoring and control functions.

[0025] This invention provides a method and system for monitoring fish health using unmanned aerial vehicles (UAVs) in marine ranching. It has the following beneficial effects: 1. This invention generates three-dimensional point cloud features by acquiring spatially distributed echo data using scanning sonar, calculates the water layer with the highest biomass density as the dynamic target depth parameter, and controls an electromechanical servo mechanism to lower the sensing unit to that water layer. This design overcomes the shortcomings of traditional fixed-depth water quality and image monitoring, which are prone to detection blind spots, ensuring that the data collected by the optical components and environmental probes directly originate from the actual activity area of ​​the fish school, thus improving the physical representativeness and accuracy of the bottom-level sensing data.

[0026] 2. This invention extracts continuous video frame swimming speed changes and multispectral reflectance features for dual physiological abnormality determination. When shallow hypoxic conditions are detected, a disease warning is activated, and a drone is controlled to travel to the free waters outside the net cage to obtain environmental reference data for internal and external differential comparison. This multimodal data cross-validation logic eliminates interference from external ocean currents, accurately distinguishes between stress avoidance behavior caused by hypoxia and actual physiological lesions, and accurately establishes the physical obstruction of the net, reducing the system's false alarm rate.

[0027] 3. This invention utilizes acoustic data to invert biomass to derive the basic theoretical feeding amount, combines it with the water body's remaining oxygen-carrying capacity to implement feeding amount cutoff constraints, and uses wide-angle images of the water surface to segment residual feed pixels and inversely derive its weight to correct baseline parameters after feeding. This mechanism avoids the risk of population metabolic hypoxia caused by overfeeding in the control feedforward stage, and transforms open-loop feeding into a closed-loop system in the control feedback stage, relying on real and effective feeding data to eliminate long-term cumulative errors in the feeding model and reduce feed waste. Attached Figure Description

[0028] Figure 1 This is a diagram of the system physical architecture of the present invention; Figure 2 This is a control flowchart of the present invention; Figure 3 This is a flowchart of the dynamic optimization detection and multimodal data acquisition method of the present invention; Figure 4 This is a flowchart of the multi-source data coupling diagnosis and differential feature verification method of the present invention; Figure 5 This is a flowchart of the biomass acoustic inversion and environmental constraint feedforward feeding method of the present invention; Figure 6 This is a flowchart of the visual residual bait inverse calculation and feeding baseline closed-loop feedback method of the present invention.

[0029] Among them, 10. Flight platform; 11. Winch mechanism; 12. Edge computing unit; 13. Surface camera assembly; 20. Sensing unit; 21. Scanning sonar; 22. Multispectral module; 23. Optical camera; 24. Environmental probe; 241. Dissolved oxygen sensor; 242. Water temperature sensor; 243. Flow meter; 25. Communication cable; 30. Cloud platform; 40. Feeding terminal. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] See attached document Figure 1 , Figure 1 This is a system physical architecture diagram according to an embodiment of the present invention. The present invention provides a drone-based fish health monitoring system for marine ranching, which may include: a flight platform 10, a sensing unit 20, a cloud platform 30, and a feeding terminal 40. Two-way data communication links are established between the device nodes via a wireless broadband communication network.

[0032] The flight platform 10 is an industrial-grade multi-rotor UAV, with a winch mechanism 11 mounted on its belly and an edge computing unit 12 built in. A surface camera assembly 13 is independently located on the underside of the flight platform 10. The edge computing unit 12 is used to perform data parsing and computation at the edge and output low-level control commands.

[0033] The sensing unit 20 is physically connected to the winch mechanism 11 via a communication cable 25. The sensing unit 20 integrates a scanning sonar 21, a multispectral module 22, an optical camera 23, and an environmental probe 24. The scanning sonar 21 is preferably a multibeam imaging sonar, a mechanically rotating scanning sonar, or other underwater acoustic detection equipment capable of outputting echo distance, azimuth angle, pitch angle, and echo intensity, enabling the edge computing unit 12 to calculate and obtain three-dimensional spatial distribution data of the target water area based on these parameters. The environmental probe 24 specifically includes a dissolved oxygen sensor 241, a water temperature sensor 242, and a current meter 243.

[0034] The cloud platform 30 is deployed on a remote server and is responsible for receiving data uploaded by the edge computing unit 12 and issuing global scheduling commands. The feeding terminal 40 is deployed at the edge of the marine ranch cage and is used to respond to feeding commands and perform mechanical scattering operations.

[0035] See attached document Figure 2 , Figure 2This is a control flowchart according to an embodiment of the present invention. The present invention provides a method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching, comprising the following steps: S10, the flight platform 10 takes off and navigates to hover above the center water area of ​​the target cage. The edge computing unit 12 issues a command to control the winch mechanism 11 to release the cable and lower the sensing unit 20 to the shallow water surface. The scanning sonar 21 is activated and performs vertical spatial acoustic detection on the underwater target area to obtain spatial distribution echo data of the target group. The spatial distribution echo data includes at least the distance information, angle information and echo intensity information of the target echo. The edge computing unit 12 generates the spatial distribution results of the target group in different water layers based on the distance information and angle information. The edge computing unit 12 analyzes the echo data, extracts the depth value of the water layer with the highest density of the target group as the dynamic target depth parameter, and drives the winch mechanism 11 to reel in and release the communication cable 25 to adjust the sensing unit 20 to the water layer indicated by the dynamic target depth parameter, and establishes a synchronous multi-dimensional data acquisition environment at the same layer. S20, the edge computing unit 12 extracts the characteristics of the group swimming speed change in the image captured by the optical camera 23, and simultaneously extracts the abnormal characteristics of the multispectral reflectance of the organism surface obtained by the multispectral module 22. When the two characteristics meet the preset alarm conditions at the same time, a candidate physiological abnormality warning signal is generated. After completing the investigation of hypoxia stress and net physical obstruction factors, it is determined whether to output a formal physiological abnormality warning signal to the cloud platform 30. At the same time, the edge computing unit 12 extracts the current environmental dissolved oxygen parameters collected by the dissolved oxygen sensor 241, and combines them with the dynamic target depth parameters fed back by the scanning sonar 21. When it is determined that there is hypoxia data and the dynamic target depth parameters are in the shallow water layer, the flight platform 10 is controlled to move laterally to the free water outside the net cage and be lowered to detect again. The difference in flow velocity attenuation and the difference in dissolved oxygen inside and outside the net cage are extracted. After confirming that the difference meets the judgment conditions, a net physical obstruction warning is output. S30, the edge computing unit 12 uses the acoustic feature data acquired by the scanning sonar 21 to establish a biological population structure model, derives the basic theoretical feeding amount parameters, and extracts the dissolved oxygen storage data of the water body acquired by the dissolved oxygen sensor 241 to calculate the metabolic oxygen consumption demand caused by the target population's full feeding; the edge computing unit 12 compares the metabolic oxygen consumption demand with the dissolved oxygen storage data of the water body, performs truncation constraint processing on the basic theoretical feeding amount parameters to generate the final execution feeding amount parameters, and sends them to the feeding terminal 40 for execution through the cloud platform 30; S40, after the feeding terminal 40 completes the mechanical feeding action and after a preset settling delay, the flight platform 10 activates the water surface camera component 13 to take pictures of the water surface area. The edge computing unit 12 extracts the residual feed feature area data in the image and converts it into the actual remaining weight of the residual feed. The edge computing unit 12 calculates the actual feeding amount of the target group by deducting the actual remaining weight of the residual feed from the final feeding amount parameter, and sends the actual feeding amount back to the cloud platform 30 to save it as historical benchmark data, which is used to correct the generation logic of the basic theoretical feeding amount parameter for the next operating cycle.

[0036] See attached document Figure 3 , Figure 3 This is a flowchart of a dynamic optimization detection and multimodal data acquisition method according to an embodiment of the present invention. To reduce the impact of water turbidity on the blindness of optical sensors and overcome the lack of representativeness in single-depth sampling, the present invention introduces a control mechanism involving acoustic guidance and electromechanical servo linkage. When the optical image quality is lower than a preset sharpness threshold, the edge computing unit 12 can pause the physiological anomaly determination based on the optical camera 23 and the multispectral module 22, and use the population aggregation density change, activity depth change, or acoustic motion characteristics obtained from the scanning sonar 21 as auxiliary anomaly judgment criteria. Specifically, step S10 can be further divided into the following sub-steps: S101, the flight platform 10 hovers at the set aerial coordinates, and the edge computing unit 12 generates an initial line-laying control command. Upon receiving this control command, the winch mechanism 11 drives its built-in servo motor to release the communication cable 25, lowering the sensing unit 20 to a pre-set shallow water area. Considering the impact of surface wave disturbance, the depth of this shallow water area is typically set within 1 to 2 meters below the water surface. After the sensing unit 20 enters the water, the scanning sonar 21 is activated and emits sound waves towards the bottom of the net cage or within a preset scanning angle range, while simultaneously receiving echo signals reflected from underwater targets. For the transmission, reception, and preliminary filtering and noise reduction of the sound wave signals, those skilled in the art can use conventional underwater acoustic signal processing methods; the underlying signal acquisition process is well-known in the field and will not be described further here.

[0037] S102, the edge computing unit 12 extracts the filtered and denoised acoustic echo signal and analyzes the time-of-flight and reflection intensity parameters carried in the signal. Based on the acoustic propagation speed and time-of-flight parameters, the edge computing unit 12 calculates the physical distance of each reflection point relative to the scanning sonar 21, and combines the azimuth, elevation, or mechanical scanning angle parameters output by the scanning sonar 21 to convert each reflection point from polar coordinates to spatial coordinates, thereby generating a three-dimensional point cloud spatial distribution feature within the target water area. The edge computing unit 12 obtains the maximum physical depth parameter of the target cage and divides the entire water body into several equidistant physical depth levels along the vertical descent direction. To balance computational efficiency and spatial resolution, the division step size between adjacent physical depth levels can be set to 0.5 meters to 1 meter. In practical implementation, it should be understood that the transmitted beam of the scanning sonar 21 exhibits a conical distribution, and the cross-sectional area covered by the beam at different depths increases with distance. Therefore, the edge computing unit 12, in conjunction with the sonar's factory-set acoustic emission angle parameters, calculates the frustum-shaped spatial volume corresponding to each physical depth level. The edge computing unit 12 counts the number of point clouds contained in each physical depth level and compares the number of point clouds with the corresponding frustum-shaped spatial volume of that level to establish the corresponding mapping relationship between vertical spatial depth and biomass density values ​​within the target water area.

[0038] S103, the edge computing unit 12 compares and filters the biomass density values ​​of each physical depth level in the above mapping relationship. Preferably, the system extracts the physical depth level with the largest biomass density value and establishes the center depth value of this level as the current dynamic target depth parameter. This dynamic target depth parameter represents the main physical water layer where the target group gathers at the current time point. If the calculation results show that multiple physical depth levels have the same biomass density value or the difference is within the set error allowable range, the edge computing unit 12 selects the level with the smaller depth value, i.e., closer to the water surface, as the dynamic target depth parameter to reduce the energy consumption and time delay of subsequent electromechanical actions. The edge computing unit 12 calculates the difference distance between the dynamic target depth parameter and the shallow surface depth where the sensing unit 20 is currently located. The system combines the physical size data of the wire reel inside the winch mechanism 11 and converts this difference distance into a corresponding motor rotation control quantity. The edge computing unit 12 transmits the rotation control quantity to the motor drive module of the winch mechanism 11. The winch mechanism 11 operates according to the set number of rotations, driving the sensing unit 20 to rise or fall vertically underwater. Combined with the feedback data from the depth encoder, it controls the sensing unit 20 to reach the water layer position indicated by the dynamic target depth parameter.

[0039] S104, considering that the environmental probe 24 requires a certain physical equilibrium time to detect water quality chemical parameters, the edge computing unit 12 sets a static delay window of several seconds after the control sensing unit 20 reaches the designated water layer position. After the delay ends, the edge computing unit 12 synchronously sends a data acquisition trigger signal through the communication cable 25. The optical camera 23, multispectral module 22, and environmental probe 24 integrated inside the sensing unit 20 respond to the trigger signal and enter the working state. The optical camera 23 acquires continuous underwater movement video frames of the target group. The multispectral module 22 emits a detection beam with a specific center wavelength to the target group and receives the reflected light, extracting multispectral reflectance characteristic data of the group surface. The dissolved oxygen sensor 241, water temperature sensor 242, and flow meter 243 included in the environmental probe 24 synchronously read the dissolved oxygen concentration, absolute water temperature, and water flow velocity values ​​in the same water layer environment. The heterogeneous data acquired by each sensing module are all tagged with a unified timestamp by the edge computing unit 12 and transmitted back to the local storage space of the edge computing unit 12 via the communication cable 25. The system obtains underlying parameters under the same spatiotemporal reference by using electromechanical guidance actions in the spatial dimension and triggering mechanisms in the temporal dimension.

[0040] See attached document Figure 4 , Figure 4 This is a flowchart of a multi-source data coupling diagnosis and differential feature verification method according to an embodiment of the present invention. For the acquired multi-dimensional heterogeneous data, the present invention eliminates false alarm interference from a single sensor by constructing a cross-correlation logic of behavioral, physiological, and environmental parameters. Step S20 can be further divided into the following sub-steps: S201, the edge computing unit 12 retrieves continuous swimming video frames acquired by the optical camera 23. The edge computing unit 12 processes the video frames using a target tracking algorithm to extract the current average swimming speed parameter of the target group. For target recognition and motion trajectory extraction processing of the video frames, those skilled in the art can use conventional machine vision tracking algorithms; the underlying image processing process is well-known in the field and will not be elaborated here. The edge computing unit 12 retrieves locally stored historical baseline swimming speed parameters and calculates the decrease rate of the current average swimming speed parameter relative to the historical baseline data. Simultaneously, the edge computing unit 12 analyzes the multispectral reflectance feature data acquired by the multispectral module 22 and extracts the spectral reflectance spectrum of the target group's surface. The edge computing unit 12 compares this spectral reflectance spectrum with a preset healthy group baseline spectrum, which is extracted from historical archives established by the system during non-abnormal warning periods. The edge computing unit 12 focuses on calculating the reflectance difference of the target surface under specific sensitive bands, such as extracting specific narrowband light reflectance data characterizing the redshift phenomenon of inflammation and congestion. The system identifies whether there are abnormal spectral reflectance patches on the surface of the target group that deviate from the normal wavelength range. Local inflammation causing capillary congestion or parasite-induced mucus shedding significantly alters the reflection characteristics of specific wavelengths of light, and these abnormal spectral reflectance patches typically indicate potential physiological lesions. The edge computing unit 12 establishes dual-constraint judgment conditions. When the swimming speed decrease rate exceeds a set swimming speed attenuation threshold (which can be set to 20% in specific implementation scenarios) and abnormal spectral reflectance patches are simultaneously detected, the system determines that the target group has entered a latent period of physiological lesions and generates a candidate physiological abnormality warning signal. This candidate physiological abnormality warning signal is temporarily stored in the local pending confirmation queue of the edge computing unit 12 for subsequent investigation of environmental stress factors before determining whether to upload it to the cloud platform 30.

[0041] S202, to further investigate stress-related anomalies caused by deteriorating physical environment, the edge computing unit 12 extracts the dissolved oxygen concentration value obtained by the dissolved oxygen sensor 241 in the current water layer. The edge computing unit 12 compares this dissolved oxygen concentration value with the set safe dissolved oxygen limit (e.g., 6 mg / L for most marine aquaculture fish). Simultaneously, the edge computing unit 12 retrieves the dynamic target depth parameter established in the aforementioned steps. Since this depth parameter indicates the main physical water layer where the current group is gathering, the edge computing unit 12 determines whether the dynamic target depth parameter is less than the set shallow activity threshold (which can be set to within 5 meters based on the physical size of the net cage). When the dissolved oxygen concentration value is lower than the safe dissolved oxygen limit and the dynamic target depth parameter is less than the shallow activity threshold, it indicates that the target group is exhibiting abnormal surfacing behavior due to oxygen deficiency in the deep water. At this time, the edge computing unit 12 generates a preliminary judgment command for net blockage. When generating the initial judgment command for net blockage, the system determines that the current decrease in the swimming speed of the fish group may be due to hypoxia stress. The edge computing unit 12 keeps the aforementioned candidate physiological abnormality warning signals in a pending confirmation state and does not upload them to the cloud platform 30 for the time being to avoid false alarms caused by non-pathological factors.

[0042] S203, relying solely on the low-oxygen state inside the net cage to determine net clogging is prone to misjudgment due to overall hypoxia or stagnant ocean currents in the external ocean environment. To perform differential physics verification in the spatial dimension, the edge computing unit 12 generates a trajectory bias control signal. The flight platform 10 receives this signal and moves horizontally to hover above a set coordinate point in the free water on the upstream side of the net cage, avoiding disturbance to the wake area behind it by the net cage structure. The upstream side can be determined based on the water flow direction collected by the current meter 243, the preset tidal current direction data of the sea area, the flow direction data of the on-site buoy, or the movement direction of floating objects in the image. The edge computing unit 12 controls the winch mechanism 11 to operate, lowering the sensing unit 20 to a physical water layer position equal to the depth parameters of the dynamic target inside the net cage. At this reference water layer, the environmental probe 24 restarts, and the edge computing unit 12 obtains the reference dissolved oxygen concentration value and the reference water flow velocity value through the dissolved oxygen sensor 241 and the current meter 243, respectively. To prevent computational anomalies during periods of extremely low ocean currents, such as slack tides, the edge computing unit 12 first determines whether the reference water flow velocity value is greater than the set effective velocity baseline. If the velocity baseline condition is met, the edge computing unit 12 calculates the attenuation difference ratio between the reference water flow velocity value outside the net cage and the water flow velocity value inside the net cage, and simultaneously calculates the absolute difference in dissolved oxygen concentration between the reference dissolved oxygen concentration value outside the net cage and the dissolved oxygen concentration value inside the net cage. When the reference water flow velocity value is not greater than the set effective velocity baseline, the edge computing unit 12 determines that the current flow field conditions do not meet the requirements for differential verification of net physical blockage, suspends the output of the net physical blockage warning, and marks this operating cycle as a low-velocity invalid verification cycle. At this time, the edge computing unit 12 can either generate an environmental low-oxygen warning signal based solely on the dissolved oxygen concentration difference, or wait for the next sampling cycle to re-execute the differential verification. When the velocity attenuation difference ratio is greater than the set velocity blockage threshold, and the absolute difference in dissolved oxygen concentration is greater than the set concentration difference threshold, the edge computing unit 12 confirms that a significant flow field isolation has been formed between the inside and outside of the net cage. Based on this, the system eliminates interference from the background ocean water quality and establishes that the low oxygen state inside the cage is caused by the netting deposits obstructing water exchange. The edge computing unit 12 generates a physical obstruction warning and reports it to the cloud platform 30. If the above differential conditions are not met, the system determines that environmental obstruction is not established. When a candidate physiological abnormality warning signal has been generated in step S201, the edge computing unit 12 confirms the candidate physiological abnormality warning signal as a formal physiological abnormality warning signal and uploads it to the cloud platform 30. When no candidate physiological abnormality warning signal is generated in step S201, the edge computing unit 12 can generate an environmental background low oxygen warning signal or maintain the current monitoring state. If the above differential conditions are met, and a candidate physiological abnormality warning signal has been generated in step S201, the priority of the candidate physiological abnormality warning signal is canceled or reduced, and the physical obstruction warning signal is uploaded first.

[0043] See attached document Figure 5 , Figure 5 This is a flowchart of a biomass acoustic inversion and environmental constraint feedforward feeding method according to an embodiment of the present invention. After completing the diagnosis of environmental and physiological states, the system establishes adaptive feedforward control logic by coupling the biomass derivation model with the environmental oxygen carrying capacity baseline. Step S30 can be further divided into the following sub-steps: S301, the edge computing unit 12 extracts the acoustic echo feature data acquired by the scanning sonar 21 during the detection phase. The edge computing unit 12 analyzes the target intensity parameters in the echo signal and combines them with a preset acoustic mapping relationship, which characterizes the correspondence between the target acoustic scattering intensity and the individual physical body length. When adjacent target echoes meet the preset spatial separation conditions, the edge computing unit 12 uses the single target echo as the body length inversion object; when adjacent target echoes overlap or the target population density is higher than the preset density threshold, the edge computing unit 12 inverts the quantity distribution of each body length interval based on the average volume scattering intensity per unit water layer, the historically calibrated target intensity distribution model, and the target quantity estimation model. By substituting the empirical constant of the acoustic reflection cross-sectional area pre-calibrated for a specific farmed fish species, the edge computing unit 12 calculates the body length feature data or body length interval distribution data of the target individuals inside the cage. After acquiring a large amount of individual body length characteristic data, the edge computing unit 12 uses a histogram statistical algorithm to divide the overall target group into multiple body length intervals according to physical size differences, and counts the number of individuals in each body length interval. For the specific implementation of the target clustering algorithm and the division of body length distribution intervals, those skilled in the art can use conventional unsupervised clustering algorithms and statistical processing methods. The underlying data classification process is a well-known technology in this field and will not be elaborated here.

[0044] S302, the edge computing unit 12 invokes a pre-set biological allometric growth model. This model reflects the power function evolution relationship between the individual weight and physical body length of aquatic organisms. The edge computing unit 12 extracts morphological empirical constants such as the power exponent and proportionality coefficient pre-established for the current farmed fish species from the model. Using this allometric growth model, combined with the central body length values ​​of each body length interval, the average individual weight corresponding to each interval is calculated and derived. The edge computing unit 12 multiplies the average individual weight by the number of individuals in that interval to obtain the subpopulation biomass of each interval. The system accumulates and summarizes the subpopulation biomass of all intervals, outputting the total biomass parameter of the target population. Based on this, the edge computing unit 12 reads the current environmental water temperature data obtained by the water temperature sensor 242, and extracts the corresponding dynamic feeding rate parameter according to the preset water temperature and feeding activity correlation. The total biomass parameter is multiplied by the dynamic feeding rate parameter to generate the basic theoretical feeding amount parameter expected to adapt to the feeding needs of the current population.

[0045] S303. Considering that fish will experience a special dynamic process after consuming large amounts of feed, leading to a significant increase in the group's metabolic oxygen demand, if the environmental water body cannot support this additional oxygen consumption, it will cause a risk of hypoxia in the group. Therefore, the edge computing unit 12 establishes an environmental oxygen carrying capacity baseline constraint mechanism. The edge computing unit 12 extracts the current dissolved oxygen concentration value of the water body obtained by the dissolved oxygen sensor 241 and retrieves the preset minimum dissolved oxygen baseline parameter. This baseline parameter represents the low oxygen critical value required to maintain the basic physiological survival of this specific fish species, and in conventional engineering implementation, it can be set between 3 mg / L and 4 mg / L based on the tolerance of the target fish species. The edge computing unit 12 first determines whether the current dissolved oxygen concentration value of the water body is greater than the minimum dissolved oxygen baseline parameter. If the determination result is no, the system determines that the current water body has lost its safe carrying capacity, and the edge computing unit 12 directly outputs a control command to ultimately execute a feeding amount parameter of zero and terminates the current feeding calculation process. If the determination result is yes, the edge computing unit 12 subtracts the minimum dissolved oxygen baseline parameter from the current dissolved oxygen concentration value of the water body to obtain the current usable safe dissolved oxygen concentration of the water body. The system acquires the physical dimensions of the target cage and calculates the effective water volume. The edge computing unit 12 multiplies the available safe dissolved oxygen concentration by the effective water volume and performs corresponding unit conversions based on the dissolved oxygen concentration unit and the effective water volume unit, expressing the remaining oxygen-carrying capacity parameter in mass units. This calculates the remaining oxygen-carrying capacity parameter of the current cage water environment. The remaining oxygen-carrying capacity parameter is a conservative static carrying capacity estimate for the current feeding cycle. In other embodiments, the edge computing unit 12 can also dynamically correct the remaining oxygen-carrying capacity parameter by combining the water exchange coefficient, current water flow velocity, water temperature correction coefficient, and the target population's basic respiratory oxygen consumption coefficient.

[0046] S304, Edge computing unit 12 retrieves the associated oxygen consumption coefficient after a unit mass of feed intake. This coefficient is typically obtained through prior laboratory closed respiration chamber testing and calibration. Edge computing unit 12 multiplies the aforementioned basic theoretical feeding amount parameter with the associated oxygen consumption coefficient to calculate the expected metabolic oxygen consumption load caused by the target population consuming the full amount of this batch of feed. Edge computing unit 12 establishes a comparison and judgment logic, comparing the expected metabolic oxygen consumption load with the remaining oxygen carrying capacity parameter. When the expected metabolic oxygen consumption load is less than or equal to the remaining oxygen carrying capacity parameter, it indicates that the current aquatic environment is expected to be able to support the oxygen demand caused by the full amount of feed intake. At this time, edge computing unit 12 confirms the basic theoretical feeding amount parameter as the final execution feeding amount parameter. When the expected metabolic oxygen consumption load is greater than the remaining oxygen carrying capacity parameter, it indicates that feeding according to the theoretical amount will break through the safe oxygen carrying capacity of the environment. Edge computing unit 12 triggers a feed restriction cutoff mechanism, and by dividing the remaining oxygen carrying capacity parameter by the associated oxygen consumption coefficient, it inversely derives the maximum safe feeding weight allowed by the environment. Edge computing unit 12 forcibly reduces the basic theoretical feeding amount, using the maximum safe feeding weight as the final execution feeding amount parameter. After the decision is made, edge computing unit 12 transmits the final execution feeding amount parameter to cloud platform 30, which converts it into an electromechanical spreading command and sends it to feeding terminal 40 to execute the corresponding action. Specifically, cloud platform 30 converts the final execution feeding amount parameter into low-level control parameters that can be executed by feeding terminal 40, based on the correspondence between the feeding amount and the feeding motor speed, feeding time, valve opening, or spreading disc speed pre-calibrated by feeding terminal 40.

[0047] See attached document Figure 6 , Figure 6 This is a flowchart of a visual residual bait inverse calculation and feeding baseline closed-loop feedback method according to an embodiment of the present invention. To form a closed-loop verification of feeding decisions and reduce the model accumulation error caused by open-loop control, the present invention provides a mechanism for actual feeding assessment and baseline correction based on visual inversion. Step S40 can be further divided into the following sub-steps: S401, the cloud platform 30 sends a control command to the feeding terminal 40, driving the feeding terminal 40 to perform the corresponding material scattering action. After the feeding action is completed, the system triggers the physical settling delay timing program. Considering the floating characteristics of feed pellets on the water surface and the average feeding cycle of the aquaculture group, the settling delay is usually set in the range of 5 to 10 minutes. In this embodiment, the visual inversion of uneaten feed is preferably applicable to floating pellet feed, or slow-sinking pellet feed that is still in the visible area of ​​the water surface within the preset settling delay; for sinking feed, the system can replace or supplement the water surface camera component 13 with a downward-looking underwater camera component, an acoustic uneaten feed detection module, or other underwater uneaten feed identification unit to obtain underwater uneaten feed distribution data. After the delay timing ends, the flight platform 10 receives the command and moves to a set height directly above the target feeding area and hovers there. This set height can be set in the range of 3 to 5 meters above the water surface according to the focal length of the mounted camera and the required field of view. The edge computing unit 12 controls the water surface camera component 13 to capture wide-angle images of the water surface below, and simultaneously records the ground flight altitude parameters of the current flight platform 10.

[0048] S402, the edge computing unit 12 inputs the acquired wide-angle image of the water surface into a pre-set bait image segmentation model. This bait image segmentation model employs a fully convolutional neural network architecture, which internally includes an encoder layer for extracting multi-scale image features, a decoder layer for restoring spatial resolution, and skip connections connecting the corresponding encoder and decoder layers. The encoder layer consists of multiple consecutive convolutional and pooling layers connected in series, and the decoder layer includes corresponding upsampling and convolutional layers. Before inputting the image into the model, the edge computing unit 12 performs image preprocessing logic, scaling the wide-angle image to the fixed resolution required by the model and normalizing the pixel values ​​to the range of 0 to 1. The input data dimension of the bait image segmentation model is a fixed-resolution three-channel color image, and the output result is a binarized semantic segmentation mask of the same resolution. The foreground pixels of the mask correspond to the physical distribution of residual feed particles on the actual water surface, while the background pixels correspond to water surface reflections and other floating interference. Before deploying this model to the edge computing unit 12, technicians need to complete the model construction and training process on the cloud platform 30. The training samples were derived from a multi-source set of surface bait images collected in the field under different lighting and wave conditions and manually labeled. The label data consisted of corresponding manually drawn binary masks. The model training process employed a binary cross-entropy loss function combined with an adaptive moment estimation optimization algorithm to iterate the network parameters until the network loss converged. The edge computing unit 12 performed forward propagation calculations by running the bait image segmentation model, outputting the corresponding binary semantic segmentation mask and counting the total number of pixels belonging to the foreground in the image.

[0049] S403, the edge computing unit 12 calls the factory-calibrated parameters of the water surface camera component 13, combines them with the acquired ground-to-ground flight altitude parameters, and performs distortion correction and perspective correction on the wide-angle image of the water surface based on the camera focal length, sensor size, lens distortion parameters, and attitude angle data of the flight platform 10 relative to the water surface. It then calculates the actual physical area of ​​the water surface represented by a single pixel in the current image based on the pinhole imaging geometry principle. In this geometric derivation, the physical area is directly proportional to the square of the ground-to-ground flight altitude parameter. The edge computing unit 12 multiplies the total number of foreground pixels by the actual physical area of ​​a single pixel to obtain the total physical distribution area of ​​the residual feed particles on the water surface. Based on this, the edge computing unit 12 retrieves the average physical cross-sectional area and average weight constant of a single particle for the current batch of feed. The edge computing unit 12 estimates the total number of residual feed particles floating on the water surface by dividing the total physical distribution area by the average physical cross-sectional area. Subsequently, the edge computing unit 12 multiplies the total number of residual feed particles by the average weight constant to deduce the actual remaining weight of the residual feed.

[0050] S404, the edge computing unit 12 extracts the final execution feeding amount parameters and the actual remaining weight of the uneaten bait that were generated and issued earlier, and first determines the validity of the visual sampling data of the uneaten bait. If the visual sampling data of the uneaten bait is determined to be invalid, the edge computing unit 12 does not use the uneaten bait weight inversion result in the calculation of the actual feeding amount, and marks this operation cycle as an abnormal sampling cycle; if the visual sampling data of the uneaten bait is determined to be valid, the edge computing unit 12 uses the final execution feeding amount parameters to deduct the actual remaining weight of the uneaten bait, and calculates the actual effective feeding load data of the target group. In order to prevent the feed from drifting out of the camera's field of view due to strong ocean currents or gusts and thus creating a calculation dead zone, the edge computing unit 12 calculates the ratio of the actual remaining weight of the uneaten bait to the final execution feeding amount parameters. If the ratio is less than the set minimum residual bait threshold (e.g., 5% in specific implementations), and the attitude tilt angle data of the flight platform 10 acquired by the system to maintain hovering is greater than the set safe and stable angle, indicating that the current sea state is high wind and waves, the edge computing unit 12 determines that the visual sampling data is invalid and does not use the residual bait weight inversion result to calculate or update the actual effective feeding load data. In other embodiments, the edge computing unit 12 can also comprehensively judge the validity of the visual sampling of residual bait based on the water flow velocity collected by the current meter 243, the drift velocity of the residual bait area in the continuous image frames, the lateral compensation velocity of the flight platform 10, or the external wind speed data; when any of the above parameters exceeds the corresponding sampling validity threshold, it is determined that the current residual bait image does not meet the weight inversion conditions. If the data is determined to be valid, the edge computing unit 12 sends the actual effective feeding load data back to the cloud platform 30 as a benchmark feedback parameter. The cloud platform 30 summarizes the actual effective feeding load data of multiple consecutive operating cycles in history to construct a moving average time series of actual feeding amount. The cloud platform 30 compares the long-term trend of this time series with the expected metabolic oxygen consumption load calculated in the corresponding historical period to deduce the long-term feeding deviation rate of the system. The cloud platform 30 uses this long-term feeding deviation rate to preferentially smooth and compensate for the dynamic feeding rate parameter, feeding activity correction coefficient, or acoustic biomass estimation correction coefficient, thereby completing the adaptive correction of the system's underlying baseline parameters for the next operating cycle. When the system further acquires manually sampled weighing data, graded fish weight data, or other reliable growth calibration data, the cloud platform 30 uses the aforementioned calibration data to smooth and compensate for the morphological empirical constants built into the allometric growth model. Through this cross-cycle feedback logic, the system achieves continuous iterative updates of the aquaculture growth model and feeding strategy. For the processing of moving average time series data and the calculation of conventional deviation rates, those skilled in the art can use standard digital signal processing and statistical algorithms; the underlying data calculation rules are well-known technologies in the field and will not be elaborated here.

[0051] In this embodiment, the swimming velocity decay threshold, safe dissolved oxygen limit, shallow activity threshold, effective flow velocity baseline, flow velocity stagnation threshold, concentration difference threshold, minimum dissolved oxygen baseline parameter, associated oxygen consumption coefficient, dynamic feeding rate parameter, and residual feed sampling effectiveness threshold can be pre-calibrated based on the target farmed fish species, cage size, farming density, feed type, and marine hydrological conditions. Calibration can be completed through historical aquaculture data statistics, manual sampling measurement, laboratory closed breathing chamber testing, field control experiments, or a combination of the above methods. Under different implementation scenarios, the above thresholds and empirical parameters can be stored in the parameter table of the edge computing unit 12 or the cloud platform 30 and can be periodically updated as historical operating data accumulates.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring fish health using unmanned aerial vehicles (UAVs) in marine ranching, characterized in that, Includes the following steps: The control flight platform uses electromechanical servo linkage to lower the onboard sensing unit to the water layer position indicated by the dynamic target depth parameter, and simultaneously collects multi-source heterogeneous reference data and spatial distribution echo data of the target group. The multi-source heterogeneous reference data includes continuous video frames of the target group, multispectral reflectance characteristic data, and dissolved oxygen concentration and water flow velocity values ​​in the same water layer environment. Extract the continuous video frames and the multispectral reflectance feature data for dual physiological abnormality determination, and couple the dissolved oxygen concentration value with the dynamic target depth parameter for analysis. Combine the reference environmental parameters outside the net cage to establish the final determination of the physical obstruction of the net. The biomass of the target population is inverted using the spatially distributed echo data to derive the basic theoretical feeding parameters. Based on the environmental oxygen carrying capacity baseline constraint mechanism, the basic theoretical feeding parameters are converted into the final execution feeding parameters for feeding. The system uses a surface camera to capture wide-angle images of the water surface and performs visual inverse calculations on the remaining bait to deduce the actual remaining weight of the remaining bait. The actual effective feeding load data is obtained by subtracting the actual remaining weight of the remaining bait from the final feeding amount parameter, and the underlying baseline parameters of the system are corrected across cycles.

2. The method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching according to claim 1, characterized in that, The control flight platform, based on electromechanical servo linkage, lowers the onboard sensing unit to the water layer position indicated by dynamic target depth parameters, specifically including: The time-of-flight parameter in the spatially distributed echo data is obtained using scanning sonar, and the three-dimensional point cloud spatial distribution characteristics inside the target water area are generated based on the time-of-flight parameter. The entire water body is divided into several equidistant physical depth levels along the vertical descent direction. The number of point clouds contained in each physical depth level is calculated, and the number of point clouds is compared with the spatial volume corresponding to the level to establish the corresponding mapping relationship between vertical spatial depth and biomass density. Extract the physical depth level with the highest biomass density value, and establish the center depth value of this level as the current dynamic target depth parameter; The difference between the dynamic target depth parameter and the current depth of the sensing unit is calculated, and the difference is converted into a motor rotation control quantity to drive the winch mechanism to move the sensing unit vertically to the water layer position indicated by the dynamic target depth parameter.

3. The method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching according to claim 2, characterized in that, When establishing the corresponding mapping relationship, the frustum-shaped spatial volume corresponding to each physical depth level is calculated by combining the factory acoustic emission angle parameters of the scanning sonar, and the number of point clouds is compared with the frustum-shaped spatial volume. If multiple physical depth levels have the same biomass density value or the difference is within the set error allowable range, the physical depth level with the smallest depth value is selected as the dynamic target depth parameter. After the driving winch mechanism controls the sensing unit to reach the designated water layer position, a static delay window of a predetermined duration is set to wait for the sensor to physically balance. After the delay ends, a data acquisition trigger signal is sent synchronously to obtain the multi-source heterogeneous reference data.

4. The method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching according to claim 1, characterized in that, The process of extracting the continuous video frames and the multispectral reflectance feature data for dual physiological anomaly determination, and coupling the dissolved oxygen concentration value with the dynamic target depth parameter for analysis, includes: Based on the continuous video frames, the decrease ratio of the current average swimming speed parameter relative to the historical benchmark data is extracted. When abnormal spectral reflectance patches are identified in the multispectral reflectance feature data and the decrease ratio exceeds the swimming speed attenuation threshold, a candidate physiological abnormality warning signal is generated. When the dissolved oxygen concentration is lower than the safe dissolved oxygen limit and the dynamic target depth parameter is less than the shallow activity threshold, a preliminary judgment command for netting blockage is generated and the candidate physiological abnormality warning signal is kept in a pending confirmation state.

5. The method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching according to claim 4, characterized in that, The final determination of the physical obstruction of the netting, based on reference environmental parameters outside the net cage, includes: The flight platform is controlled to move to the free water area on the upstream side of the cage, and the sensing unit is lowered to the reference water layer position with the same depth parameter as the dynamic target to obtain the reference dissolved oxygen concentration value and the reference water flow velocity value. When it is determined that the reference water flow velocity value is greater than the set effective flow velocity base, the ratio of the flow velocity attenuation difference between the reference water flow velocity value and the water flow velocity value is calculated, and the absolute difference in concentration between the reference dissolved oxygen concentration value and the dissolved oxygen concentration value is calculated at the same time. When the flow velocity attenuation difference ratio is greater than the set flow velocity obstruction threshold and the absolute concentration difference is greater than the set concentration difference threshold, it is established that there is a low oxygen state inside the net cage caused by the physical obstruction of water exchange by the net attachment material, and a net physical obstruction warning is generated. If the flow rate attenuation difference ratio is not greater than the flow rate stagnation threshold or the absolute concentration difference is not greater than the concentration difference threshold, then the environmental stagnation is determined to be invalid, and the candidate physiological abnormality warning signal is confirmed as a formal physiological abnormality warning signal and output.

6. The method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching according to claim 1, characterized in that, The process of using the spatially distributed echo data to invert the biomass of the target population to derive the basic theoretical feeding parameters, and then converting these parameters into final execution feeding parameters based on the environmental oxygen-carrying baseline constraint mechanism, includes: Based on the target intensity parameter in the spatially distributed echo data, the body length characteristic data is inverted and divided into multiple body length intervals. The subpopulation biomass of each interval is derived using an allometric growth model combined with preset morphological empirical constants. The total biomass parameter is extracted and multiplied with the total biomass parameter and the dynamic feeding rate parameter set based on water temperature to generate the basic theoretical feeding amount parameter. Determine whether the dissolved oxygen concentration in the current water area is greater than the set minimum dissolved oxygen baseline parameter; if it is not greater, directly output the control command that the final feeding amount parameter is zero and terminate the current feeding calculation; if it is greater, subtract the minimum dissolved oxygen baseline parameter from the dissolved oxygen concentration to obtain the usable safe dissolved oxygen concentration, and calculate the remaining oxygen carrying capacity parameter of the current water area in combination with the effective water volume inside the cage. The expected metabolic oxygen demand is calculated by multiplying the basic theoretical feeding amount parameter by the associated oxygen demand coefficient per unit mass of feed. When the expected metabolic oxygen demand is greater than the remaining oxygen carrying capacity parameter, the feed restriction cutoff mechanism is triggered. The remaining oxygen carrying capacity parameter is divided by the associated oxygen demand coefficient to obtain the maximum safe feeding weight. The maximum safe feeding weight is then output as the final feeding amount parameter.

7. The method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching according to claim 1, characterized in that, The process of using a surface camera to capture wide-angle images of the water surface and performing visual residual bait inverse calculations to correct the system's underlying baseline parameters across cycles includes: After the system completes the feeding and the set physical settling delay, the flight platform is controlled to move to a set height and hover. The water surface camera component is used to capture a wide-angle image of the water surface and simultaneously record the ground flight altitude parameters. After performing image preprocessing, the wide-angle image of the water surface is input into the residual bait image segmentation model, and the corresponding binarized semantic segmentation mask image is output. The total number of pixels belonging to the foreground is counted. Combining the ground flight altitude parameters with the factory internal calibration parameters of the water surface camera component, the actual physical area of ​​the water surface represented by a single pixel in the current image is calculated based on the geometric principle of pinhole imaging. Multiply the total number of foreground pixels by the actual physical area of ​​the water surface to obtain the total physical distribution area. Combine the average physical cross-sectional area and average weight constant of a single feed pellet to deduce the actual remaining weight of the uneaten feed. A moving average time series of the actual effective feeding load data is constructed to calculate the long-term feeding deviation rate of the system. The long-term feeding deviation rate is then used to perform reverse smoothing compensation on the dynamic feeding rate parameter, feeding activity correction coefficient, or acoustic biomass estimation correction coefficient to complete the adaptive correction of the underlying baseline parameters.

8. The method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching according to claim 7, characterized in that, The residual feed image segmentation model adopts a fully convolutional neural network architecture. Its input data dimension is a three-channel color image with a fixed resolution, and the output result is the binarized semantic segmentation mask image that represents the physical distribution state of residual feed particles. The internal layers of the fully convolutional neural network architecture include an encoder layer consisting of alternating convolutional and pooling layers for extracting multi-scale image features, a decoder layer containing upsampling and convolutional layers, and a skip connection structure connecting the corresponding encoder and decoder layers. The training samples for the bait residue image segmentation model are derived from a set of multi-source surface bait residue images that have been manually labeled. The label data is the corresponding manually drawn binary mask image. During the network training process, the binary cross-entropy loss function is used to iterate the network parameters.

9. The method for monitoring fish health using unmanned aerial vehicles (UAVs) based on marine ranching according to claim 7, characterized in that, When acquiring the actual effective food intake data, the following steps are included: Calculate the ratio of the actual remaining weight of the uneaten bait to the final feeding amount parameter; Determine whether the ratio is less than the set minimum threshold for extremely low residual bait, and simultaneously acquire the attitude tilt data of the flight platform, which characterizes the current wind, wave and sea conditions. If the ratio is less than the minimum residual bait threshold, and the attitude tilt angle data generated to maintain the hovering state is greater than the safe and stable angle, then the visual sampling data acquired in this instance is deemed invalid, and the actual effective feeding load data is not calculated using the residual bait weight inversion result of this instance.

10. A drone-based fish health monitoring system for marine ranching, characterized in that, include: Memory is used to store computer program instructions; A processor, configured to implement, when executing the computer program instructions, the method for monitoring fish health using unmanned aerial vehicles based on marine ranching as described in any one of claims 1 to 9.

Images