Method and system for predicting concentration of nutrient salt in aquaculture area coupled with remote sensing data

Abstract

The present application relates to the field of environmental monitoring and remote sensing information processing technology, and discloses a method and system for predicting the concentration of nutrient salts in a breeding area coupled with remote sensing data. The method comprises: constructing a physically constrained ecological hydrodynamic model to generate synthetic remote sensing-nutrient salt pairing data; fusing multi-source remote sensing and measured data to form a high-dimensional feature tensor; separating breeding sensitive features and environmental interference features through an attention-guided domain-invariant feature decoupling network; transferring prior knowledge of the source domain large-scale breeding area using a dynamic weight cross-domain knowledge distillation architecture; and finally outputting real-time high-precision nutrient salt concentration distribution through a lightweight regression head and time series smoothing. The system includes corresponding functional units. The present application improves the prediction accuracy and stability in small sample scenarios, suppresses negative transfer, and provides support for precision breeding and eutrophication early warning.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring and remote sensing information processing technology, specifically relating to a method and system for predicting nutrient concentration in aquaculture areas by coupling remote sensing data. Background Technology

[0002] With the rapid development of aquaculture in marine and inland waters, dynamic monitoring of nutrient concentrations in aquaculture areas has become a crucial link in ensuring water quality safety and ecological balance. Remote sensing technology, with its advantages of wide-area, high-frequency, and non-contact observation, has become an important means of retrieving water environment parameters.

[0003] Nutrient concentration, as a core indicator reflecting the eutrophication level of water bodies, is crucial for accurate prediction in aquaculture management, pollution early warning, and ecological regulation. Current remote sensing inversion methods mainly rely on establishing empirical or semi-analytical models between water spectral reflectance and measured nutrient concentrations, typically requiring extensive synchronous field sampling data for model training and validation.

[0004] Machine learning-based remote sensing inversion methods have shown great potential in recent years, especially in large-scale, high-density sampling areas. However, for small-scale or remote aquaculture areas, the high cost and low frequency of field sampling result in a severe shortage of available remote sensing-measured paired samples. To alleviate the data shortage problem, researchers have attempted to introduce transfer learning strategies to apply models trained over large areas to smaller areas. However, differences in water optical properties, suspended matter composition, algal community structure, and lighting conditions among different aquaculture areas cause a severe domain shift between the source and target domains. This leads to a sharp decline in the generalization ability of the transfer model in the target small area, resulting in large prediction biases and poor stability.

[0005] Existing technologies generally face three major bottlenecks when dealing with remote sensing prediction of nutrients in small-sample aquaculture areas:

[0006] Traditional supervised learning models are prone to overfitting under conditions of scarce samples and are difficult to capture the complex nonlinear relationship between water spectra and nutrients.

[0007] Existing transfer learning methods often employ feature distribution alignment or adversarial training, failing to fully consider the shared physical mechanisms and commonalities in spectral responses across multiple regions, resulting in limited domain adaptation effects.

[0008] The lack of an intelligent learning framework capable of extracting generalizable prior knowledge from multiple heterogeneous regions and quickly adapting to new regions fails to meet the actual needs of real-time, accurate, and adaptive water quality monitoring in aquaculture areas.

[0009] Therefore, there is an urgent need for a nutrient concentration inversion method that can effectively integrate multi-source remote sensing data, overcome the influence of domain offset, and achieve high-precision prediction with a very small number of samples. Summary of the Invention

[0010] To address the aforementioned technical problems, embodiments of the present invention provide a method and system for predicting nutrient concentrations in aquaculture areas by coupling remote sensing data. This method establishes a virtual migration bridge with ecological process consistency between the source and target domains by constructing a synthetic data generation mechanism based on physical constraints.

[0011] A high-dimensional feature tensor is formed by spatiotemporal alignment and fusion of multi-source remote sensing observation data and field measured data.

[0012] An attention-guided domain-invariant feature decoupling module is introduced to separate the sensitive feature subspace affected by aquaculture activities from the redundant feature subspace affected by environmental background interference.

[0013] Based on this, a cross-domain knowledge distillation architecture with dynamic weight adaptation is designed to transfer the prior knowledge of the large-scale aquaculture area in the source domain to the small-scale target aquaculture area in a structured manner.

[0014] Ultimately, a lightweight online inference engine was used to achieve real-time, high-precision prediction of nutrient concentrations.

[0015] This invention provides a method for predicting nutrient concentrations in aquaculture areas by coupling remote sensing data, comprising:

[0016] Acquire historical remote sensing image data, field measured nutrient concentration data, and corresponding hydrological and meteorological driving data of large-scale aquaculture areas in the source region;

[0017] Based on the historical remote sensing image data, field measured nutrient concentration data and hydrological and meteorological driving data, an eco-hydraulic numerical model constrained by physical processes is constructed to generate a synthetic remote sensing-nutrient pairing dataset covering the geographical range and seasonal changes of small-scale aquaculture areas in the target domain.

[0018] Real-time multi-source remote sensing image data of a small-scale aquaculture area in the target domain is acquired. The multi-source remote sensing image data includes reflectance in the visible light band, reflectance in the near-infrared band, reflectance in the short-wave infrared band, and thermal infrared brightness temperature data.

[0019] Atmospheric correction, geometric fine correction and pixel-level spatiotemporal registration are performed on the real-time multi-source remote sensing image data to form a multispectral feature cube under a unified spatiotemporal reference.

[0020] The multispectral feature cube is input into a pre-trained domain-invariant feature decoupling network, which includes a channel attention-guided feature separation layer and a gradient inversion layer, used to output a feature map sensitive to aquaculture activities and a feature map of environmental background interference.

[0021] The aquaculture activity sensitive feature map and the synthetic remote sensing-nutrient pairing dataset are jointly input into the cross-domain knowledge distillation module. The cross-domain knowledge distillation module includes a teacher network and a student network. The teacher network has fixed parameters on the source domain data, and the student network receives the target domain features and minimizes the structured distillation loss with the output of the teacher network.

[0022] Based on the output of the student network, the predicted nutrient concentration values ​​of each pixel in the target aquaculture area are generated by a lightweight fully connected regression head.

[0023] The predicted nutrient concentration values ​​are post-processed using a sliding window time series smoothing algorithm to output the final real-time nutrient concentration distribution map.

[0024] This invention also provides a nutrient concentration prediction system for aquaculture areas coupled with remote sensing data, comprising:

[0025] The source domain data acquisition unit is used to acquire historical remote sensing image data, field measured nutrient concentration data, and corresponding hydrological and meteorological driving data of large-scale aquaculture areas in the source domain.

[0026] The synthetic data generation unit is used to construct an eco-hydraulic numerical model constrained by physical processes based on the historical remote sensing image data, field measured nutrient concentration data and hydro-meteorological driving data, and generate a synthetic remote sensing-nutrient pairing dataset covering the geographical range and seasonal changes of the target domain's small-scale aquaculture area.

[0027] The target domain remote sensing data acquisition unit is used to acquire real-time multi-source remote sensing image data of a small-scale aquaculture area in the target domain. The multi-source remote sensing image data includes visible light band reflectance, near-infrared band reflectance, short-wave infrared band reflectance, and thermal infrared brightness temperature data.

[0028] The remote sensing data preprocessing unit is used to perform atmospheric correction, geometric fine correction and pixel-level spatiotemporal registration on the real-time multi-source remote sensing image data to form a multispectral feature cube under a unified spatiotemporal reference.

[0029] A domain-invariant feature decoupling unit is used to input the multispectral feature cube into a pre-trained domain-invariant feature decoupling network. The domain-invariant feature decoupling network includes a channel attention-guided feature separation layer and a gradient inversion layer, which are used to output aquaculture activity sensitive feature map and environmental background interference feature map.

[0030] A cross-domain knowledge distillation unit is used to input the aquaculture activity sensitive feature map and the synthetic remote sensing-nutrient pairing dataset into the cross-domain knowledge distillation module. The cross-domain knowledge distillation module includes a teacher network and a student network. The teacher network has fixed parameters on the source domain data, and the student network receives the target domain features and minimizes the structured distillation loss with the output of the teacher network.

[0031] The nutrient concentration prediction unit is used to generate predicted nutrient concentration values ​​for each pixel in the target aquaculture area based on the output of the student network using a lightweight fully connected regression head.

[0032] The prediction result post-processing unit is used to post-process the predicted nutrient concentration values ​​using a sliding window time series smoothing algorithm to output the final real-time nutrient concentration distribution map.

[0033] As one embodiment of the present invention, the physical process-constrained ecological hydrodynamic numerical model specifically includes: a one-dimensional vertical mixing layer model, a phytoplankton growth dynamics equation, a nutrient cycle transformation equation, and wind and wave disturbance boundary conditions.

[0034] The input variables of the model include average daily solar radiation flux, surface water temperature, wind speed, precipitation intensity, initial nutrient concentration, and aquaculture feed amount.

[0035] The model outputs three-dimensional field data of chlorophyll a concentration, suspended matter concentration, and dissolved inorganic nitrogen concentration, which are matched with the spatiotemporal resolution of remote sensing images.

[0036] By coupling the three-dimensional field data with the radiative transfer model, the corresponding synthetic remote sensing reflectance data is generated through inversion, thereby forming a synthetic remote sensing-nutrient pairing dataset.

[0037] As one embodiment of the present invention, the construction process of the multispectral feature cube specifically includes: stacking the atmospherically corrected remote sensing images in band order to form an initial feature tensor;

[0038] The initial feature tensor is masked and clipped using the vector boundary of the aquaculture area;

[0039] The bilinear interpolation method was used to uniformly resample images acquired by different satellite platforms to a spatial resolution of 10 meters;

[0040] By using a timestamp alignment strategy, image data from multiple transits within the same day are fused according to the principle of maximizing the effective pixels to form a multispectral feature cube with no cloud gaps on a single day.

[0041] As one embodiment of the present invention, the channel attention-guided feature separation layer of the domain-invariant feature decoupling network specifically includes:

[0042] Global average pooling layer, two fully connected layers, and Sigmoid activation function;

[0043] The global average pooling layer compresses the spatial dimension of each channel of the input feature map;

[0044] The two fully connected layers constitute a bottleneck structure, and the number of hidden units is 1 / 16 of the number of input channels;

[0045] The Sigmoid activation function outputs the importance weights of each channel.

[0046] The importance weights are multiplied channel by channel with the original input feature map to obtain the aquaculture activity sensitive feature map; the remaining part is connected through residuals to obtain the environmental background interference feature map; the gradient inversion layer keeps the data flow unchanged during forward propagation and multiplies the gradient by a negative constant factor during backward propagation to suppress the interference of environmental background features on the classification boundary.

[0047] As one embodiment of the present invention, the structured distillation loss in the cross-domain knowledge distillation module specifically includes:

[0048] Mean squared error loss at the feature map level, KL divergence loss at the prediction distribution level, and consistency loss of nutrient concentration gradient field.

[0049] The mean squared error loss is calculated to determine the pixel-level difference between the intermediate feature maps of the student network and the teacher network under the same input.

[0050] The KL divergence loss measures the relative entropy of the output probability distributions of the two;

[0051] The consistency loss is obtained by extracting the spatial gradient fields of nutrient concentration output by students and teachers respectively using the Sobel operator, and calculating their L1 norm distance.

[0052] The total distillation loss function is formed by the weighted sum of the three losses, with weighting coefficients of 0.5, 0.3, and 0.2, respectively.

[0053] As one embodiment of the present invention, the lightweight fully connected regression head specifically includes three fully connected layers with the number of neurons being 256, 64 and 1 respectively; batch normalization and ReLU activation functions are used between each layer; the last layer does not use an activation function and directly outputs continuous nutrient concentration values; the input of the regression head is the vector of the feature map of the last layer of the student network after global average pooling.

[0054] As one embodiment of the present invention, the sliding window time series smoothing algorithm specifically includes: setting the time window length to 7 days;

[0055] For each pixel location, the historical prediction sequence is used to construct a seven-dimensional vector by taking the prediction values ​​of the current time and the three days before and after it.

[0056] A Hanning window weighted average filter is applied to the vector with window function coefficients of 0.08, 0.22, 0.32, 0.4, 0.32, 0.22, and 0.08; the smoothed daily nutrient concentration value is output to suppress abnormal fluctuations caused by cloud remnants or sensor noise in a single remote sensing observation.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] 1. This invention effectively alleviates the problem of scarce remote sensing samples in small-scale aquaculture areas by constructing a synthetic data generation mechanism constrained by physical processes. The generated synthetic data not only has the apparent characteristics of remote sensing observation, but also embeds real ecological hydrodynamic processes, thereby improving the data foundation quality for transfer learning.

[0059] 2. By using a domain-invariant feature decoupling network, the feature subspace sensitive to aquaculture activities is accurately separated, effectively suppressing domain shift interference caused by geographical location, climate background, and water optical properties.

[0060] 3. Through a structured cross-domain knowledge distillation architecture, efficient and stable knowledge transfer from a large-scale source domain to a small-scale target domain is achieved, avoiding the negative transfer phenomenon caused by excessive differences in feature distribution in traditional transfer learning.

[0061] 4. By using a lightweight online inference engine and time series smoothing post-processing, the real-time and stability requirements for nutrient concentration monitoring in aquaculture areas are met. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the overall technical architecture of the method and system for predicting nutrient concentration in aquaculture areas by coupling remote sensing data, as proposed in this invention.

[0063] Figure 2 This is a schematic diagram of the core principle framework of the attention-guided domain-invariant feature decoupling module in this invention;

[0064] Figure 3 This is a logical flowchart of the physical process constraint-based remote sensing-nutrient pairing data generation stage in this invention.

[0065] Figure 4 This is a logical flowchart of the multi-source remote sensing image preprocessing and multispectral feature cube construction stage in this invention.

[0066] Figure 5 This is a logical flow diagram of the cross-domain knowledge distillation architecture with dynamic weight adaptation in this invention;

[0067] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal prediction system and the cloud data processing unit in this invention. Detailed Implementation

[0068] Please refer to Figures 1 to 6 This invention provides a method and system for predicting nutrient concentration in aquaculture areas by coupling remote sensing data, aiming to solve the technical problem that the scarcity of remote sensing samples in small-scale aquaculture areas leads to the failure of transfer learning generalization and the inability to meet the needs of real-time monitoring.

[0069] This method establishes a virtual migration bridge with ecological process consistency between the source and target domains by constructing a synthetic data generation mechanism constrained by physical processes. It utilizes multi-source remote sensing observation data and field-measured data for spatiotemporal alignment and fusion to form a high-dimensional feature tensor. An attention-guided domain-invariant feature decoupling module is introduced to separate the sensitive feature subspace affected by aquaculture activities from the redundant feature subspace affected by environmental background interference. Based on this, a dynamically weighted, adaptive cross-domain knowledge distillation architecture is designed to transfer prior knowledge from the large-scale aquaculture area in the source domain to the small-scale target aquaculture area in a structured manner. Finally, a lightweight online inference engine enables real-time, high-precision prediction of nutrient concentrations.

[0070] The method includes the following steps:

[0071] S1, acquire historical remote sensing image data of large-scale aquaculture areas in the source region, on-site measured nutrient concentration data, and corresponding hydrological and meteorological driving data;

[0072] S2, Based on the historical remote sensing image data, field measured nutrient concentration data and hydrological and meteorological driving data, construct an eco-hydraulic numerical model constrained by physical processes, and generate a synthetic remote sensing-nutrient pairing dataset covering the geographical range and seasonal changes of small-scale aquaculture areas in the target domain.

[0073] S3, acquire real-time multi-source remote sensing image data of the small-scale aquaculture area in the target domain. The multi-source remote sensing image data includes visible light band reflectance, near-infrared band reflectance, short-wave infrared band reflectance, and thermal infrared brightness temperature data.

[0074] S4, Atmospheric correction, geometric fine correction and pixel-level spatiotemporal registration are performed on the real-time multi-source remote sensing image data to form a multispectral feature cube under a unified spatiotemporal reference.

[0075] S5, the multispectral feature cube is input into a pre-trained domain-invariant feature decoupling network, which includes a channel attention-guided feature separation layer and a gradient inversion layer, used to output a breeding activity sensitive feature map and an environmental background interference feature map.

[0076] S6, input the aquaculture activity sensitive feature map and the synthetic remote sensing-nutrient pairing dataset into the cross-domain knowledge distillation module. The cross-domain knowledge distillation module includes a teacher network and a student network. The teacher network has fixed parameters on the source domain data, and the student network receives the target domain features and minimizes the structured distillation loss with the output of the teacher network.

[0077] S7. Based on the output of the student network, the predicted nutrient concentration values ​​of each pixel in the target aquaculture area are generated through a lightweight fully connected regression head.

[0078] S8, the predicted nutrient concentration values ​​are post-processed using a sliding window time series smoothing algorithm to output the final real-time nutrient concentration distribution map.

[0079] In step S1, historical remote sensing image data of the large-scale aquaculture area in the source region, on-site measured nutrient concentration data, and corresponding hydrological and meteorological driving data are acquired.

[0080] The source area large-scale aquaculture zone refers to a large aquaculture area with long-term, continuous, and high-density on-site sampling records. Its remote sensing image data comes from multiple Earth observation satellites, including but not limited to the Landsat series, Sentinel-2 satellite, and Gaofen series satellites.

[0081] The historical remote sensing image data should span at least 3 years, have a spatial resolution greater than 30 meters, and cover the complete aquaculture cycle, including the feeding period, growth period, harvest period, and rest period.

[0082] Nutrient concentration data measured on-site were collected regularly by professional water quality monitoring vessels or buoy stations, with a sampling frequency of once a week. The measured indicators included major nutrient components such as dissolved inorganic nitrogen, phosphate, and silicate, and the sampling depth was 0.5 meters below the surface. Hydrometeorological driving data included daily average solar radiation flux, surface water temperature, wind speed, precipitation intensity, evaporation, tidal height, and aquaculture feed input. Data sources included the National Meteorological Administration, the National Marine Forecasting Center, and aquaculture management systems. The temporal resolution was daily, and the spatial location was aligned with the center of the remote sensing pixels.

[0083] All data is standardized and stored according to a unified coordinate system (such as WGS84) and time standard (UTC) to form a structured source domain database.

[0084] In step S2, based on the historical remote sensing image data, field measured nutrient concentration data, and hydrological and meteorological driving data, an eco-hydrodynamic numerical model constrained by physical processes is constructed to generate a synthetic remote sensing-nutrient pairing dataset covering the geographical range and seasonal changes of small-scale aquaculture areas in the target domain.

[0085] The ecological hydrodynamic numerical model consists of a one-dimensional vertical mixing layer model, phytoplankton growth dynamics equations, nutrient cycle transformation equations, and wind and wave disturbance boundary conditions.

[0086] A one-dimensional vertical mixing layer model describes the vertical mixing process of water under the influence of wind stress and heat flux. The governing equations are:

[0087] ;

[0088] in For the depth of the mixing layer, For reference water density, For wind stress, For bottom frictional stress, This represents the net heat flux. The phytoplankton growth kinetics equation uses the modified Monod model, and the expression is:

[0089] ;

[0090] in Phytoplankton biomass, For the maximum specific growth rate, To determine the concentration of dissolved inorganic nitrogen, It is the half-saturation constant. For light intensity, The light suppression constant is... The mortality rate.

[0091] The nutrient cycle transformation equation describes the transformation of elements such as nitrogen, phosphorus, and silicon in dissolved, particulate, and biological states, including processes such as mineralization, assimilation, and sedimentation.

[0092] The wave disturbance boundary conditions are calculated based on the measured wind speed to determine the wave energy input, which drives the turbulent mixing of the surface water.

[0093] The model input variables include daily average solar radiation flux, surface water temperature, wind speed, precipitation intensity, initial nutrient concentration, and aquaculture feed amount. The output is three-dimensional field data of chlorophyll a concentration, suspended solids concentration, and dissolved inorganic nitrogen concentration, which are matched with the spatiotemporal resolution of remote sensing images.

[0094] Subsequently, the three-dimensional field data is input into the water radiative transfer model to simulate the water radiance in different bands (blue, green, red, near-infrared, and short-wave infrared) and convert it into apparent reflectance, thereby generating synthetic remote sensing reflectance data with the same spectral response function and noise characteristics as the real remote sensing image.

[0095] Finally, the synthetic remote sensing reflectance data and the corresponding dissolved inorganic nitrogen concentration field are paired pixel by pixel to form a synthetic remote sensing-nutrient paired dataset. This dataset covers the entire latitude and longitude range of the target domain's small-scale aquaculture area and includes typical hydrological and meteorological scenarios for spring, summer, autumn, and winter, ensuring the spatiotemporal generalization ability of transfer learning.

[0096] In step S3, real-time multi-source remote sensing image data of small-scale aquaculture areas in the target domain are acquired. Small-scale aquaculture areas in the target domain refer to small or emerging aquaculture areas lacking long-term field sampling; their remote sensing data is downloaded in real-time via satellite ground receiving stations or cloud service platforms.

[0097] Multi-source remote sensing imagery data includes reflectance in the visible light band (440 nm to 670 nm), near-infrared band (840 nm), short-wave infrared band (1610 nm and 2200 nm), and thermal infrared brightness temperature data (11 μm band). Data sources include the Sentinel-2 multispectral imager, the Gaofen-1 wide-field camera, and the Fengyun-3 medium-resolution spectrophotometer, ensuring at least one effective overflight per day. All imagery data carries precise timestamps and orbital parameters for subsequent spatiotemporal alignment.

[0098] In step S4, atmospheric correction, geometric fine correction and pixel-level spatiotemporal registration are performed on the real-time multi-source remote sensing image data to form a multispectral feature cube under a unified spatiotemporal reference.

[0099] Atmospheric correction employs a method combining dark target subtraction and radiative transfer simulation. First, deep-water dark pixels in the image are identified, and aerosol optical thickness is estimated. Then, the surface reflectance is retrieved using the MODTRAN model. Geometric fine correction uses a high-precision digital elevation model and coastline vectors as control points. A second-order polynomial transformation is used to project the image onto the UTM coordinate system, with residuals controlled within 0.5 pixels.

[0100] Pixel-level spatiotemporal registration first resamples images from different satellite platforms to a spatial resolution of 10 meters, and uses bilinear interpolation to avoid high-frequency information distortion;

[0101] Subsequently, images that passed through the area multiple times on the same day were fused according to the principle of the largest effective pixel. That is, for each pixel location, the observation value marked as "clear sky" by the cloud mask and with the highest signal-to-noise ratio was selected first. If there was no effective data on the day, the data was backdated to a maximum of 3 days and the nearest neighbor interpolation was used to fill in the gaps.

[0102] Finally, the processed reflectivity of each band and the thermal infrared brightness temperature are stacked in band order to form a structure with a size of [missing information]. Multispectral feature cube, in which and The cube represents the spatial dimension of the target aquaculture area, with five channels corresponding to the blue, green, red, near-infrared, and thermal infrared bands, respectively. This cube serves as the input to the subsequent feature decoupling network.

[0103] In step S5, the multispectral feature cube is input into a pre-trained domain-invariant feature decoupling network.

[0104] This network consists of a channel attention-guided feature separation layer and a gradient inversion layer. The channel attention-guided feature separation layer first performs global average pooling on the input feature map to obtain the statistical response of each channel;

[0105] Then, a nonlinear transformation is performed through a bottleneck structure consisting of two fully connected layers (the number of hidden units is 1 / 16 of the number of input channels);

[0106] Finally, the importance weights of each channel are output through the Sigmoid activation function. These weights are then multiplied channel by channel with the original input feature map to obtain aquaculture activity-sensitive feature map. This feature map prominently reflects changes in the optical properties of the water body caused by human activities such as feeding, excretion, and agitation.

[0107] The remaining part is connected via residuals. An environmental background interference feature map is obtained, which includes signals caused by natural factors such as geographical location, climate, and river input.

[0108] The gradient reversal layer keeps the data flow unchanged during forward propagation, but multiplies the gradient of the environmental background interference feature map by a negative constant factor (with a value of -0.1) during backward propagation. This allows the backbone network to actively suppress its dependence on environmental background features during optimization, thereby enhancing domain invariance.

[0109] The network is pre-trained on a mixed set of source domain synthetic data and a small number of target domain labeled samples. The loss function consists of a weighted sum of classification loss and domain discrimination loss.

[0110] In step S6, the aquaculture activity sensitive feature map and the synthetic remote sensing-nutrient pairing dataset are input together into the cross-domain knowledge distillation module.

[0111] This module comprises a teacher network and a student network. The teacher network is a deep convolutional neural network fully trained on synthetic data from the source domain. Its structure is an encoder-decoder, with a ResNet34 backbone for the encoder and a U-Net skip connection structure for the decoder. The output is a nutrient concentration distribution map. The teacher network parameters are fixed during the distillation process. The student network has a similar structure to the teacher network but with fewer layers, retaining only the first three residual blocks to reduce computational complexity.

[0112] The student network receives a sensitive feature map of aquaculture activities in the target domain as input and outputs the corresponding nutrient concentration prediction.

[0113] The structured distillation loss consists of three parts: mean squared error loss at the feature map level. KL divergence loss at the predicted distribution level and the consistency loss of the nutrient concentration gradient field. The mean squared error loss calculates the pixel-level difference between the student and teacher feature maps in the intermediate layers of the network; the KL divergence loss measures the relative entropy of the output probability distributions of the two, where the outputs are normalized by Softmax; the consistency loss extracts the spatial gradient fields of nutrient concentrations of the student and teacher outputs respectively using the Sobel operator and calculates their L1 norm distance.

[0114] The total distillation loss function is:

[0115] ;

[0116] The student network learns the structured knowledge of the teacher network by minimizing this loss function, thus achieving cross-domain transfer.

[0117] In step S7, based on the output of the student network, the predicted nutrient concentration values ​​of each pixel in the target aquaculture area are generated using a lightweight fully connected regression head.

[0118] The regression head is attached after the feature map of the last layer of the student network. First, global average pooling is performed on the feature map to obtain a feature vector of length 512.

[0119] Then, the neurons are passed through three fully connected layers with 256, 64, and 1 neurons respectively; batch normalization and ReLU activation function are used between each layer.

[0120] The final layer does not use an activation function and directly outputs continuous nutrient concentration values ​​(in milligrams per liter). This regression head has few parameters, fast inference speed, and is suitable for deployment on edge computing devices.

[0121] In step S8, the predicted nutrient concentration values ​​are post-processed using a sliding window time series smoothing algorithm.

[0122] The time window length is set to 7 days. For the historical prediction sequence of each pixel location, the prediction values ​​of the current moment and the 3 days before and after it are used to form a seven-dimensional vector.

[0123] Apply a Hanning window weighted average filter to the vector with window function coefficients of 0.08, 0.22, 0.32, 0.4, 0.32, 0.22, and 0.08; output the smoothed daily nutrient concentration value.

[0124] This post-processing suppresses abnormal fluctuations in single remote sensing observations caused by cloud remnants, fog, or sensor noise, improving temporal continuity and operational availability.

[0125] The final output is a real-time nutrient concentration distribution map in GeoTIFF format with a spatial resolution of 10 meters and an update frequency of once a day.

[0126] The system includes a source domain data acquisition unit, a synthetic data generation unit, a target domain remote sensing data acquisition unit, a remote sensing data preprocessing unit, a domain-invariant feature decoupling unit, a cross-domain knowledge distillation unit, a nutrient concentration prediction unit, and a prediction result post-processing unit.

[0127] The source domain data acquisition unit is responsible for retrieving historical remote sensing images, measured nutrient concentrations, and hydrological and meteorological data from satellite data centers, water quality monitoring stations, and meteorological service platforms, and for standardizing formats and controlling quality.

[0128] The synthetic data generation unit integrates an eco-hydraulic numerical model and a radiative transfer model, and synthesizes remote sensing-nutrient pairing data covering the spatiotemporal range of the target domain based on source domain data.

[0129] The target domain remote sensing data acquisition unit automatically downloads the latest multi-source remote sensing images of the target aquaculture area via API interface or FTP protocol.

[0130] The remote sensing data preprocessing unit performs atmospheric correction, geometric fine correction and spatiotemporal fusion, and outputs a multispectral feature cube.

[0131] The domain-invariant feature decoupling unit loads a pre-trained network to perform feature separation on the input cube and output sensitive and interference feature maps.

[0132] The cross-domain knowledge distillation unit operates on a teacher-student architecture, using synthetic data to guide student networks in learning domain-invariant representations. The nutrient concentration prediction unit generates preliminary predictions using a lightweight regression head.

[0133] The prediction result post-processing unit applies a sliding window smoothing algorithm to output the final product.

[0134] Each unit achieves efficient data flow through message queues and shared memory, supporting distributed deployment and cloud-based collaborative inference.

[0135] This embodiment achieves high-precision, real-time prediction of nutrient concentration in small-scale aquaculture areas with only a few labeled samples using the above-described method and system. It solves the problem of generalization failure caused by domain offset in traditional transfer learning and provides reliable technical support for the precision management of aquaculture.

[0136] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish an entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0137] 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 predicting nutrient concentration in aquaculture areas using coupled remote sensing data, characterized in that, include: Acquire historical remote sensing image data, field measured nutrient concentration data, and corresponding hydrological and meteorological driving data of large-scale aquaculture areas in the source region; Based on the historical remote sensing image data, field measured nutrient concentration data and hydrological and meteorological driving data, an eco-hydraulic numerical model constrained by physical processes is constructed to generate a synthetic remote sensing-nutrient pairing dataset covering the geographical range and seasonal changes of small-scale aquaculture areas in the target domain. Real-time multi-source remote sensing image data of a small-scale aquaculture area in the target domain is acquired. The multi-source remote sensing image data includes reflectance in the visible light band, reflectance in the near-infrared band, reflectance in the short-wave infrared band, and thermal infrared brightness temperature data. Atmospheric correction, geometric fine correction and pixel-level spatiotemporal registration are performed on the real-time multi-source remote sensing image data to form a multispectral feature cube under a unified spatiotemporal reference. The multispectral feature cube is input into a pre-trained domain-invariant feature decoupling network, which includes a channel attention-guided feature separation layer and a gradient inversion layer, used to output a feature map sensitive to aquaculture activities and a feature map of environmental background interference. The aquaculture activity sensitive feature map and the synthetic remote sensing-nutrient pairing dataset are jointly input into the cross-domain knowledge distillation module. The cross-domain knowledge distillation module includes a teacher network and a student network. The teacher network has fixed parameters on the source domain data, and the student network receives the target domain features and minimizes the structured distillation loss with the output of the teacher network. Based on the output of the student network, the predicted nutrient concentration values ​​of each pixel in the target aquaculture area are generated by a lightweight fully connected regression head. The predicted nutrient concentration values ​​are post-processed using a sliding window time series smoothing algorithm to output the final real-time nutrient concentration distribution map. Based on the historical remote sensing imagery data, field-measured nutrient concentration data, and hydrological and meteorological driving data, a physically constrained eco-hydraulic numerical model is constructed to generate a synthetic remote sensing-nutrient pairing dataset covering the geographical range and seasonal variations of small-scale aquaculture areas in the target domain, including: An eco-hydraulic numerical model was constructed, consisting of a one-dimensional vertical mixing layer model, phytoplankton growth dynamics equations, nutrient cycling and transformation equations, and wind and wave disturbance boundary conditions. Using the average daily solar radiation flux, surface water temperature, wind speed, precipitation intensity, initial nutrient concentration, and aquaculture feed amount as model input variables, the output is three-dimensional field data of chlorophyll a concentration, suspended solids concentration, and dissolved inorganic nitrogen concentration that match the spatiotemporal resolution of remote sensing images. The three-dimensional field data is coupled with the water body radiative transfer model to generate corresponding synthetic remote sensing reflectance data, forming a synthetic remote sensing-nutrient paired dataset. The process of performing atmospheric correction, geometric fine correction, and pixel-level spatiotemporal registration on the real-time multi-source remote sensing image data to form a multispectral feature cube under a unified spatiotemporal reference includes: The atmospherically corrected remote sensing images are stacked in band order to form the initial feature tensor; The initial feature tensor is masked and clipped using the vector boundary of the aquaculture area; The bilinear interpolation method was used to uniformly resample images acquired by different satellite platforms to a spatial resolution of 10 meters; By using a timestamp alignment strategy, image data that have passed through multiple times on the same day are fused according to the principle of the maximum effective pixels to form a multispectral feature cube with no cloud gaps on a single day. The process involves inputting the multispectral feature cube into a pre-trained domain-invariant feature decoupling network. This network includes a channel attention-guided feature separation layer and a gradient inversion layer, used to output a feature map sensitive to aquaculture activities and a feature map showing environmental background interference. The spatial dimension of each channel of the input feature map is compressed using a global average pooling layer; The compressed feature vector is nonlinearly transformed through a bottleneck structure consisting of two fully connected layers, where the number of hidden units in the bottleneck structure is 1 / 16 of the number of input channels. The importance weights of each channel are output by the Sigmoid activation function and multiplied with the original input feature map channel by channel to obtain the aquaculture activity sensitive feature map; Calculate the environmental background interference feature map using residual connections; During backpropagation, the gradient of the environmental background interference feature map is multiplied by a negative constant factor to suppress its impact on network optimization.

2. The method for predicting nutrient concentration in aquaculture areas by coupling remote sensing data according to claim 1, characterized in that, The step of inputting the aquaculture activity sensitive feature map and the synthetic remote sensing-nutrient pairing dataset into the cross-domain knowledge distillation module includes: The teacher network is trained on source domain data and its parameters are fixed. The teacher network is an encoder-decoder structure. The student network receives a sensitive feature map of aquaculture activities in the target domain as input, and its network depth is less than that of the teacher network. Calculate the structured distillation loss between the student network and the teacher network. The structured distillation loss includes the mean square error loss at the feature map level, the KL divergence loss at the prediction distribution level, and the consistency loss of the nutrient concentration gradient field. The student network parameters are updated by minimizing the structured distillation loss.

3. The method for predicting nutrient concentration in aquaculture areas by coupling remote sensing data according to claim 2, characterized in that, In the structured distillation loss, the mean squared error loss at the feature map level calculates the pixel-level difference between the intermediate feature maps of the student network and the teacher network under the same input. The KL divergence loss at the prediction distribution level measures the relative entropy of the output probability distributions of the two. The consistency loss of the nutrient concentration gradient field is obtained by extracting the spatial gradient fields of nutrient concentration output by students and teachers respectively using the Sobel operator, and calculating their L1 norm distance. The total distillation loss function is formed by the weighted sum of the three losses.

4. The method for predicting nutrient concentration in aquaculture areas by coupling remote sensing data according to claim 3, characterized in that, The process of generating predicted nutrient concentration values ​​for each pixel in the target aquaculture area using a lightweight fully connected regression head based on the output of the student network includes: Global average pooling is performed on the feature map of the last layer of the student network to obtain the feature vector; The feature vector is passed through three fully connected layers in sequence; Batch normalization and ReLU activation function are used between each layer, and no activation function is used in the last layer, which directly outputs continuous nutrient concentration values.

5. The method for predicting nutrient concentration in aquaculture areas by coupling remote sensing data according to claim 4, characterized in that, The post-processing of the predicted nutrient concentration values ​​using a sliding window time series smoothing algorithm includes: Set the time window length to 7 days; For each pixel location, the historical prediction sequence is used to construct a seven-dimensional vector by taking the prediction values ​​of the current time and the three days before and after it. Apply a Hanning window weighted average filter to the vector; Output the smoothed nutrient concentration values ​​for the day.

6. A nutrient concentration prediction system for aquaculture areas coupled with remote sensing data, characterized in that, include: The source domain data acquisition unit is used to acquire historical remote sensing image data, field measured nutrient concentration data, and corresponding hydrological and meteorological driving data of large-scale aquaculture areas in the source domain. The synthetic data generation unit is used to construct an eco-hydraulic numerical model constrained by physical processes based on the historical remote sensing image data, field measured nutrient concentration data and hydro-meteorological driving data, and generate a synthetic remote sensing-nutrient pairing dataset covering the geographical range and seasonal changes of the target domain's small-scale aquaculture area. The target domain remote sensing data acquisition unit is used to acquire real-time multi-source remote sensing image data of a small-scale aquaculture area in the target domain. The multi-source remote sensing image data includes visible light band reflectance, near-infrared band reflectance, short-wave infrared band reflectance, and thermal infrared brightness temperature data. The remote sensing data preprocessing unit is used to perform atmospheric correction, geometric fine correction and pixel-level spatiotemporal registration on the real-time multi-source remote sensing image data to form a multispectral feature cube under a unified spatiotemporal reference. A domain-invariant feature decoupling unit is used to input the multispectral feature cube into a pre-trained domain-invariant feature decoupling network. The domain-invariant feature decoupling network includes a channel attention-guided feature separation layer and a gradient inversion layer, which are used to output aquaculture activity sensitive feature map and environmental background interference feature map. A cross-domain knowledge distillation unit is used to input the aquaculture activity sensitive feature map and the synthetic remote sensing-nutrient pairing dataset into the cross-domain knowledge distillation module. The cross-domain knowledge distillation module includes a teacher network and a student network. The teacher network has fixed parameters on the source domain data, and the student network receives the target domain features and minimizes the structured distillation loss with the output of the teacher network. The nutrient concentration prediction unit is used to generate predicted nutrient concentration values ​​for each pixel in the target aquaculture area based on the output of the student network using a lightweight fully connected regression head. The prediction result post-processing unit is used to post-process the predicted nutrient concentration values ​​through a sliding window time series smoothing algorithm and output the final real-time nutrient concentration distribution map. The synthetic data generation unit is used for: An eco-hydraulic numerical model was constructed, consisting of a one-dimensional vertical mixing layer model, phytoplankton growth dynamics equations, nutrient cycling and transformation equations, and wind and wave disturbance boundary conditions. Using the average daily solar radiation flux, surface water temperature, wind speed, precipitation intensity, initial nutrient concentration, and aquaculture feed amount as model input variables, the output is three-dimensional field data of chlorophyll a concentration, suspended solids concentration, and dissolved inorganic nitrogen concentration that match the spatiotemporal resolution of remote sensing images. The three-dimensional field data is coupled with the water body radiative transfer model to generate corresponding synthetic remote sensing reflectance data, forming a synthetic remote sensing-nutrient paired dataset. The process of performing atmospheric correction, geometric fine correction, and pixel-level spatiotemporal registration on the real-time multi-source remote sensing image data to form a multispectral feature cube under a unified spatiotemporal reference includes: The atmospherically corrected remote sensing images are stacked in band order to form the initial feature tensor; The initial feature tensor is masked and clipped using the vector boundary of the aquaculture area; The bilinear interpolation method was used to uniformly resample images acquired by different satellite platforms to a spatial resolution of 10 meters; By using a timestamp alignment strategy, image data that have passed through multiple times on the same day are fused according to the principle of the maximum effective pixels to form a multispectral feature cube with no cloud gaps on a single day. The process involves inputting the multispectral feature cube into a pre-trained domain-invariant feature decoupling network. This network includes a channel attention-guided feature separation layer and a gradient inversion layer, used to output a feature map sensitive to aquaculture activities and a feature map showing environmental background interference. The spatial dimension of each channel of the input feature map is compressed using a global average pooling layer; The compressed feature vector is nonlinearly transformed through a bottleneck structure consisting of two fully connected layers, where the number of hidden units in the bottleneck structure is 1 / 16 of the number of input channels. The importance weights of each channel are output by the Sigmoid activation function and multiplied with the original input feature map channel by channel to obtain the aquaculture activity sensitive feature map; Calculate the environmental background interference feature map using residual connections; During backpropagation, the gradient of the environmental background interference feature map is multiplied by a negative constant factor to suppress its impact on network optimization.

Images