Offshore fishery water chlorophyll-a concentration detection method and system based on deep learning
By constructing a physically constrained generative adversarial network and combining prior knowledge of marine optics with a spatial gradient consistency discrimination mechanism, a high-fidelity virtual chlorophyll a concentration sample is generated, solving the problem of scarce measured chlorophyll a data in nearshore areas and achieving high-precision concentration detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 福建省渔业资源监测中心
- Filing Date
- 2026-01-05
- Publication Date
- 2026-07-31
AI Technical Summary
Given the extreme scarcity of measured chlorophyll a data in nearshore areas, deep learning models suffer from insufficient training samples and poor generalization ability, making it difficult to achieve high-precision concentration detection.
A physical constraint generative adversarial network is constructed between multi-source remote sensing data and sparse measured data. Prior knowledge of marine optics and spatial gradient consistency discrimination mechanism are introduced. A high-fidelity virtual chlorophyll a concentration sample set is generated through a composite loss function, and a concentration inversion model with strong generalization ability is trained.
Under extremely low-density measured data conditions, a high-precision, high-resolution, and physically interpretable global inversion of chlorophyll a concentration in nearshore fishery waters was achieved. The inversion results are superior to traditional empirical algorithms and deep learning models without physical constraints.
Smart Images
Figure CN121884097B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method and system for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning. Background Technology
[0002] With the increasing demand for marine ecological environment monitoring and nearshore fishery resource management, chlorophyll a concentration, as a key indicator for measuring primary productivity and eutrophication levels in water bodies, has become a research hotspot in the interdisciplinary field of remote sensing and environmental science due to its high precision and high frequency of detection. Traditional methods mainly rely on on-site sampling combined with laboratory spectrophotometry or fluorescence methods. Although these methods have high accuracy, they are limited by manpower, cost, and spatiotemporal coverage, making it difficult to support large-scale, dynamic nearshore ecological assessments.
[0003] In recent years, optical remote sensing technology based on satellite or UAV platforms has been widely adopted due to its wide-area observation advantages, estimating chlorophyll a concentration by inverting water reflectance spectral data. However, the nearshore environment is highly complex, affected by multiple factors such as suspended sediment, colored soluble organic matter, and water depth changes, resulting in severe nonlinear aliasing of spectral signals. This leads to a decline in the generalization performance of traditional empirical models or semi-analytical algorithms in this scenario.
[0004] Deep learning-based chlorophyll a concentration inversion methods demonstrate the potential to overcome the limitations of traditional models due to their powerful nonlinear fitting and feature extraction capabilities. These methods typically construct end-to-end neural networks that directly map multi-band water reflectance to chlorophyll a concentration values, theoretically enabling adaptive learning of the spectral-concentration correlation patterns in complex water bodies.
[0005] However, effective training of deep learning models relies heavily on a large number of high-quality labeled samples, while nearshore areas face the reality of extremely scarce measured chlorophyll a data: offshore operations are costly and subject to weather and sea conditions, resulting in low frequency and sparse spatial distribution of ground-based ground-value data collection; the optical characteristics of water bodies in different sea areas vary significantly, making it difficult for models trained in one area to be transferred to other nearshore scenarios, further exacerbating the data bottleneck's constraint on the model's generalization ability.
[0006] In existing technologies, some studies have attempted to expand the training set through data augmentation or synthesis strategies. However, spectral data generated by simple noise injection or linear interpolation lacks physical consistency and cannot reflect the radiative transfer characteristics of real water bodies. Instead, it introduces false patterns that mislead model learning. Another approach uses remote sensing-measured paired data from open seas or lakes for pre-training and then fine-tuning for nearshore tasks. However, due to the complex composition and diverse optical properties of nearshore waters, there is a distribution shift between the source and target domains, resulting in limited transfer effects.
[0007] Furthermore, directly using chlorophyll a inversion results provided by global-scale satellite products (such as MODIS and Sentinel-3) as alternative labels is also insufficient to meet the requirements of high-precision modeling due to problems such as land pollution, atmospheric correction errors, and insufficient spatial resolution of nearshore pixels.
[0008] Therefore, under the condition of scarce measured data, how to generate physically reliable and spectrally accurate near-shore water training samples, and effectively integrate multi-source remote sensing knowledge to improve the model's cross-regional adaptability, has become an urgent technical problem to be solved. Summary of the Invention
[0009] This invention provides a method and system for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning, aiming to solve the technical problems of insufficient training samples and poor generalization ability of deep learning models caused by the extreme scarcity of measured chlorophyll a data in nearshore areas.
[0010] This method constructs a generative adversarial network architecture driven by physical constraints between multi-source remote sensing data and sparse measured data, introduces marine optical prior knowledge as a regularization condition for the generator and discriminator, and combines a spatiotemporal consistency loss function to dynamically calibrate the synthetic data. Thus, without the need for a large number of measured labels, a high-fidelity, physically interpretable virtual chlorophyll a concentration sample set is generated for training a concentration inversion model with strong generalization ability.
[0011] This invention provides a deep learning-based method for detecting chlorophyll a concentration in nearshore fishery waters, comprising:
[0012] Acquire multispectral or hyperspectral remote sensing image data of the target nearshore waters, wherein the remote sensing image data includes reflectance information in the visible to near-infrared bands;
[0013] Simultaneously acquire sparsely distributed field-measured chlorophyll a concentration data in the water area. The measured data is obtained by buoy stations, mobile sampling or fixed-point sampling, and its spatial coverage density is less than one effective sample point per 100 square kilometers.
[0014] A physical constraint generative adversarial network is constructed, which includes a generator and a discriminator. The generator takes remote sensing image data as input and outputs a predicted chlorophyll a concentration distribution map. The discriminator is used to distinguish the difference between the concentration map output by the generator and the actual measured concentration map.
[0015] The generator incorporates a nonlinear response relationship between chlorophyll a and remote sensing reflectance derived from an ocean radiative transfer model as a physical constraint module. This module expresses the mapping relationship between the remote sensing reflectance ratio and chlorophyll a concentration in different concentration ranges through a predefined piecewise power function. Specifically:
[0016] When the chlorophyll a concentration is less than 0.3 mg / m³, a power function of the blue-green band reflectance ratio is used for fitting.
[0017] When the concentration is between 0.3 and 5 mg / m³, a quadratic polynomial is used for fitting.
[0018] When the concentration is greater than 5 mg / m³, the logarithmic function of the difference between red light and near-infrared reflectance is used for fitting.
[0019] A spatial gradient consistency discrimination mechanism is introduced into the discriminator. This mechanism calculates the spatial gradient field of the generated density map and the corresponding remote sensing image, and constrains the alignment of the two on the edge structure to ensure that the generated result is consistent with the optical image in terms of spatial texture.
[0020] Define a composite loss function, which includes adversarial loss, physical constraint loss, spatiotemporal smoothing loss, and sparse measured point supervision loss;
[0021] The physical constraint loss is the mean square error between the remote sensing reflectance calculated by substituting the generated concentration value into the aforementioned piecewise power function and the original remote sensing reflectance.
[0022] The spatiotemporal smoothing loss is a constraint on the variation amplitude between adjacent time phases in the generated concentration map in the time dimension, and a second derivative penalty term between neighboring pixels in the spatial dimension.
[0023] The sparse measured point monitoring loss is the absolute error between the generated concentration map at the measured point location and the measured value.
[0024] The physical constraint generative adversarial network was trained end-to-end using the composite loss function until the generator could stably output a chlorophyll a concentration distribution map that conformed to physical laws and was consistent with sparse measured data.
[0025] The trained generator is used as the backbone model for chlorophyll a concentration inversion. Remote sensing image data of the water area to be detected is input, and the chlorophyll a concentration distribution results of the entire water area are output.
[0026] As one embodiment of the present invention, the multispectral or hyperspectral remote sensing image data comes from the Haiyang-1 satellite, the Gaofen series satellites, or the Sentinel-2 satellite, with a spatial resolution greater than 30 meters, a time revisit period of less than 5 days, and spectral bands including at least four center wavelengths of 490 nm, 560 nm, 620 nm, and 680 nm.
[0027] As one embodiment of the present invention, the on-site measured chlorophyll a concentration data is determined by acetone extraction-fluorescence method, the sampling depth is within 0.5 meters of the water surface, the time difference between the sampling time and the remote sensing transit time is less than 2 hours, and the geographical location of the sampling point is corrected by differential global positioning system with a positioning error of less than 10 meters.
[0028] As one embodiment of the present invention, the generator adopts an encoder-decoder structure. The encoder consists of 5 levels of convolutional blocks. Each level includes two 3×3 convolutional layers, a batch normalization layer, and a modified linear unit activation function. The number of channels is 64, 128, 256, 512, and 1024 respectively.
[0029] The decoder uses a transposed convolution upsampling path to fuse the feature maps of the corresponding levels of the encoder step by step. In the output layer, a hyperbolic tangent activation function is used to map the output value range to between -1 and 1, and then a linear transformation is used to convert it into a concentration range of 0 to 10 milligrams per cubic meter.
[0030] In one embodiment of the present invention, the discriminator adopts a fully convolutional discriminant network. Its input is the channel splicing tensor of remote sensing image and density map. The network structure includes four levels of downsampling convolutional blocks. Each level includes two 3×3 convolutional layers and a leakage correction linear unit activation function. The output of the last level is a single-channel probability map, which represents the probability that each spatial location belongs to the real sample.
[0031] As one embodiment of the present invention, the time dimension constraint in the spatiotemporal smoothing loss is only enabled when there is continuous multi-temporal remote sensing data. It is calculated as the sum of the squares of the pixel differences between the current temporal generation concentration map and the previous temporal generation concentration map at the same geographical location, and the difference is limited to a threshold of less than 0.5 milligrams per cubic meter per day.
[0032] This invention also provides a deep learning-based chlorophyll a concentration detection system for nearshore fishery waters, comprising:
[0033] The remote sensing data acquisition unit is used to acquire multispectral or hyperspectral remote sensing image data of the target's nearshore waters.
[0034] The measured data acquisition unit is used to acquire the sparsely distributed on-site measured chlorophyll a concentration data in the water area;
[0035] The physical constraint generative adversarial network building unit is used to construct a network architecture that includes a generator and a discriminator. The generator is embedded with a piecewise power function physical constraint module of marine optical priors, and the discriminator integrates a spatial gradient consistency discrimination mechanism.
[0036] The composite loss function definition unit is used to define a joint optimization objective that includes adversarial loss, physical constraint loss, spatiotemporal smoothing loss and sparse measured point supervision loss;
[0037] The model training unit is used to train the physical constraint generative adversarial network based on the remote sensing image data and the measured data, using the composite loss function.
[0038] The concentration inversion execution unit is used to take the trained generator as the inversion model, input the remote sensing image of the water area to be detected, and output a global chlorophyll a concentration distribution map.
[0039] In one embodiment of the present invention, in the physical constraint generative adversarial network construction unit, the piecewise power function physical constraint module is embedded in the loss calculation path of the generator in a differentiable form, and its parameters remain fixed during training and do not participate in gradient updates.
[0040] As one embodiment of the present invention, the model training unit adopts a two-stage training strategy: in the first stage, the encoder part of the generator is frozen, and only the decoder and physical constraint module are trained to quickly align sparse measured points; in the second stage, all parameters are unfrozen, end-to-end fine-tuning is performed, and the weight coefficient of the physical constraint loss is reduced, gradually decaying from the initial value of 1.0 to 0.3.
[0041] As one embodiment of the present invention, the concentration distribution map output by the concentration inversion execution unit is subjected to outlier removal and spatial interpolation smoothing by the post-processing module. The outlier removal is based on the three-standard-deviation criterion, and the spatial interpolation smoothing adopts the Kriging interpolation method with the range parameter set to 500 meters.
[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0043] 1. This invention effectively alleviates the problems of model overfitting and physical distortion caused by the extreme sparseness of measured samples by embedding prior knowledge of marine optical physics into a generative adversarial network in the form of differentiable piecewise functions.
[0044] 2. The introduced spatial gradient consistency discrimination mechanism ensures that the generated concentration map is strictly aligned with the geometric features of the remote sensing image in terms of spatial structure;
[0045] 3. The spatiotemporal smoothing constraint in the composite loss function improves the stability and spatial continuity of the model over time series.
[0046] 4. The two-stage training strategy balances the rapid convergence of sparse supervision signals with the fine optimization of global generalization ability.
[0047] 5. This invention achieves high-precision, high-resolution, and physically interpretable global inversion of chlorophyll a concentration in nearshore fishery waters, relying only on extremely low-density measured data. The inversion results are superior to traditional empirical algorithms and pure data-driven deep learning models without physical constraints. Attached Figure Description
[0048] Figure 1 This is a schematic diagram of the overall technical architecture of the method and system for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning proposed in this invention.
[0049] Figure 2 This is a schematic diagram of the core principle framework of the physical constraint generative adversarial network in this invention;
[0050] Figure 3 This is a flowchart illustrating the logical process of fusing multi-source remote sensing data and sparse measured data in this invention.
[0051] Figure 4 This is a schematic diagram of the principle framework of the generator embedding the marine optical prior piecewise power function physical constraint module in this invention;
[0052] Figure 5 This is a schematic diagram of the optimization framework consisting of the discriminator integrating the spatial gradient consistency discrimination mechanism and the composite loss function in this invention;
[0053] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal and the cloud in this invention. Detailed Implementation
[0054] Please refer to Figures 1 to 6 This invention provides a method and system for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning. Its core is to solve the technical problem of insufficient training samples and poor generalization ability of deep learning models caused by the extreme scarcity of measured chlorophyll a data in nearshore areas.
[0055] This method constructs a physically constrained generative adversarial network architecture, embeds marine optical prior knowledge as a differentiable physical constraint module in the generator, introduces a spatial gradient consistency discrimination mechanism in the discriminator, and defines a composite loss function that includes adversarial loss, physical constraint loss, spatiotemporal smoothing loss, and sparse measured point supervision loss. Thus, it generates a high-fidelity, physically interpretable virtual chlorophyll a concentration sample set with only extremely low density measured labels, and trains a concentration inversion model with strong generalization ability.
[0056] The method includes the following steps:
[0057] S1, acquire multispectral or hyperspectral remote sensing image data of the target nearshore waters;
[0058] S2, synchronously acquire the field measured chlorophyll a concentration data of sparsely distributed chlorophyll a in the water area;
[0059] S3, constructing a physically constrained generative adversarial network;
[0060] S4, embeds the nonlinear response relationship of chlorophyll a-remote sensing reflectance derived from the ocean radiative transfer model as a physical constraint module in the generator;
[0061] S5 introduces a spatial gradient consistency discrimination mechanism into the discriminator;
[0062] S6, Define the composite loss function;
[0063] S7. The physical constraint generative adversarial network is trained end-to-end using the composite loss function.
[0064] S8 uses the trained generator as the backbone model for chlorophyll a concentration inversion, inputs remote sensing image data of the water area to be detected, and outputs the chlorophyll a concentration distribution results of the entire water area.
[0065] In step S1, multispectral or hyperspectral remote sensing image data of the target nearshore waters are acquired, and the remote sensing image data includes reflectance information in the visible to near-infrared bands.
[0066] Specifically, the remote sensing image data comes from the Haiyang-1 satellite, the Gaofen series satellites, or the Sentinel-2 satellite. Its spatial resolution is greater than 30 meters, the time revisit period is less than 5 days, and the spectral bands include at least four center wavelengths: 490 nanometers, 560 nanometers, 620 nanometers, and 680 nanometers.
[0067] Remote sensing image data is provided in the form of surface reflectance. After atmospheric correction, the effects of aerosols, water vapor and Rayleigh scattering are eliminated to ensure that the reflectance value truly reflects the optical properties of the water body.
[0068] The image data is stored in GeoTIFF format and includes geographic coordinate information, projection parameters and band metadata. Each pixel corresponds to a multi-band reflectance vector of a geographic location.
[0069] Data acquisition is completed through the satellite ground receiving station or the interface of the National Remote Sensing Data Center, using the standard HTTP protocol for downloading, and file integrity is verified through MD5 checksum.
[0070] Image preprocessing includes geometric fine correction, cloud mask removal, stripe noise restoration, and band registration to ensure that multi-temporal data are strictly aligned in space.
[0071] In step S2, the sparsely distributed field-measured chlorophyll a concentration data in the water area are acquired simultaneously. The measured data is obtained by buoy stations, mobile sampling or fixed-point sampling, and its spatial coverage density is less than one effective sample point per 100 square kilometers.
[0072] The measured data were determined by acetone extraction-fluorescence method. The sampling depth was within 0.5 meters of the water surface. The time difference between the sampling time and the remote sensing transit time was less than 2 hours. The geographical location of the sampling point was corrected by differential global positioning system, and the positioning error was less than 10 meters.
[0073] Water samples were collected at a designated depth using a Niskin water sampler, immediately transported to the laboratory under refrigeration and protected from light, and then filtered, extracted, and subjected to fluorescence analysis within 24 hours.
[0074] The measured concentration values are in milligrams per cubic meter and are recorded in a structured database. Each record includes a sampling timestamp, latitude and longitude coordinates, water depth, temperature, salinity, and concentration value.
[0075] Data quality control includes duplicate sample comparison, blank sample subtraction, and instrument calibration curve verification to ensure that the measurement error is less than 5%.
[0076] The spatial distribution of the measured points was confirmed to be sparse by Voronoi diagram analysis, and the Euclidean distance between any two adjacent sampling points was greater than 10 kilometers.
[0077] In step S3, a physical constraint generative adversarial network is constructed, which includes a generator and a discriminator.
[0078] The generator takes remote sensing image data as input and outputs a predicted chlorophyll a concentration distribution map.
[0079] The discriminator is used to distinguish the differences between the concentration map output by the generator and the actual measured concentration map.
[0080] The generator adopts an encoder-decoder structure. The encoder consists of 5 levels of convolutional blocks. Each level contains two 3×3 convolutional layers, a batch normalization layer, and a modified linear unit activation function. The number of channels is 64, 128, 256, 512, and 1024, respectively.
[0081] The decoder employs a transposed convolutional upsampling path, fusing feature maps from corresponding encoder layers level by level. At the output layer, a hyperbolic tangent activation function maps the output value range to between -1 and 1, and then a linear transformation converts it to a concentration range of 0 to 10 milligrams per cubic meter. The discriminator uses a fully convolutional discriminant network. Its input is a channel-stitched tensor of the remote sensing image and the concentration map. The network structure contains four levels of downsampling convolutional blocks, each containing two 3×3 convolutional layers and a leakage-corrected linear unit activation function. The final output is a single-channel probability map, representing the probability that each spatial location belongs to a true sample.
[0082] The network weights were initialized using a Xavier normal distribution, with a learning rate of 0.0002. The optimizer was Adam, and the momentum parameters were set to 0.5 and 0.999.
[0083] In step S4, the nonlinear response relationship of chlorophyll a-remote sensing reflectance derived from the ocean radiative transfer model is embedded in the generator as a physical constraint module.
[0084] This module expresses the mapping relationship between the remote sensing reflectance ratio and chlorophyll a concentration in different concentration ranges using a preset piecewise power function. The specific form is as follows:
[0085] When the chlorophyll a concentration is less than 0.3 mg / m³, a power function of the blue-green band reflectance ratio is used for fitting; when the concentration is between 0.3 and 5 mg / m³, a quadratic polynomial is used for fitting.
[0086] When the concentration is greater than 5 mg / m³, the logarithmic function of the difference between the reflectance of red light and near-infrared bands is used for fitting.
[0087] The piecewise function is implemented in a differentiable form to ensure that the gradient can be backpropagated. Let the generator output concentration map be... The corresponding reflectance of the remote sensing image in the blue light band (490 nm) is The green light band (560 nanometers) is The red light band (620 nanometers) is The near-infrared band (680 nanometers) is The physical constraint module is defined as follows:
[0088] when hour,
[0089] ;
[0090] when hour,
[0091] ;
[0092] when hour,
[0093] ;
[0094] The above parameters are calibrated based on the classic ocean optical models OC2, OC3, and CI algorithm, and remain fixed during training, not participating in gradient updates. The physical constraint module is embedded after the generator output and before the loss calculation, used to back-calculate the generated concentration into the theoretical remote sensing reflectance and compare it with the original input reflectance.
[0095] In step S5, a spatial gradient consistency discrimination mechanism is introduced into the discriminator. This mechanism calculates the spatial gradient fields of the generated density map and the corresponding remote sensing image, and constrains the alignment of the two on the edge structure. Specifically, it is implemented as follows:
[0096] For generating concentration maps Calculate the first-order partial derivatives in the horizontal and vertical directions to obtain the gradient magnitude: ;
[0097] Green band of remote sensing images Similarly, calculate the gradient magnitude. .
[0098] The gradient consistency loss is defined as the negative of the normalized cross-correlation coefficient between the two factors, i.e.:
[0099] ;
[0100] in , These are the mean and standard deviation, respectively. The total number of pixels. This loss term is integrated into the discriminator's discrimination logic, enabling the discriminator to not only determine the authenticity of the concentration value but also assess whether its spatial texture is consistent with the edge structure of the optical image. This mechanism prevents the generator from producing blurred or structurally misaligned concentration distributions, ensuring that features such as algal bloom patches and fronts accurately correspond to water color anomalies in the remote sensing image.
[0101] In step S6, a composite loss function is defined. The composite loss function includes adversarial loss, physical constraint loss, spatiotemporal smoothing loss, and sparse measured point supervision loss.
[0102] The adversarial loss is expressed in the form of Wasserstein distance and is defined as the difference between the discriminator's mean output for real samples and the mean output for generated samples.
[0103] The physical constraint loss is the mean square error between the remote sensing reflectance derived by substituting the generated concentration value into the aforementioned piecewise power function and the original remote sensing reflectance, i.e.:
[0104] ;
[0105] The spatiotemporal smoothing loss includes spatial and temporal terms. The spatial term is a penalty for the second derivative between neighboring pixels of the generated concentration map, i.e.:
[0106] ;
[0107] The time term is enabled only when there is continuous multi-phase data. It is defined as the sum of squared differences between pixels at the same location in the concentration map of the current phase and the previous phase, and the difference is limited to a threshold of less than 0.5 mg / m³ per day.
[0108] The sparse measured point monitoring loss is the absolute error between the generated concentration map at the measured point location and the measured value, i.e.:
[0109] ;
[0110] in This represents the number of measured points.
[0111] The overall expression for the composite loss function is:
[0112] ;
[0113] Where the initial weight .
[0114] In step S7, the physical constraint generative adversarial network is trained end-to-end using the composite loss function.
[0115] The training adopts a two-stage strategy: the first stage freezes the encoder part of the generator and trains only the decoder and physical constraint module to quickly align sparse test points. This stage lasts for 5000 iterations.
[0116] The second stage involves unfreezing all parameters, performing end-to-end fine-tuning, and simultaneously reducing the weighting coefficient of the physical constraint loss from an initial value of 1.0 to 0.3, using an exponential decay function. , This represents the number of iteration steps.
[0117] The training batch size was set to 4, the input image was cropped to 256×256 pixel blocks, and data augmentation was performed using random rotation, flipping, and brightness perturbation.
[0118] The training process monitors the mean absolute error of sparse points on the validation set and the physical plausibility of the generated concentration map. Training is terminated when there is no improvement after 2000 consecutive iterations.
[0119] The final model saves the optimal checkpoint, and the average absolute error of the generated concentration map at the measured point is less than 0.3 mg / m³, and the correlation coefficient between the back-calculated reflectance and the original reflectance is greater than 0.9.
[0120] In step S8, the trained generator is used as the backbone model for chlorophyll a concentration inversion. Remote sensing image data of the water area to be detected is input, and the chlorophyll a concentration distribution across the entire water area is output. The output concentration map undergoes outlier removal and spatial interpolation smoothing via a post-processing module.
[0121] Outlier removal is based on the three-standard-deviation criterion, which means that pixels whose concentration values exceed the mean plus or minus three standard deviations are removed.
[0122] Spatial interpolation smoothing was performed using Kriging interpolation, with a range parameter set to 500 meters, a nugget effect of 0.01, and an off-slab value of 0.05. The final product was output in NetCDF format, including concentration values, uncertainty estimates, and timestamps. The spatial resolution was the same as the original remote sensing image resolution, and the temporal precision was the remote sensing transit time. This result can be directly used for fisheries resource assessment, red tide early warning, and water quality management decision support.
[0123] The system includes a remote sensing data acquisition unit, a measured data acquisition unit, a physical constraint generative adversarial network construction unit, a composite loss function definition unit, a model training unit, and a concentration inversion execution unit.
[0124] The remote sensing data acquisition unit automatically downloads and preprocesses multispectral or hyperspectral remote sensing images through the satellite data interface;
[0125] The measured data acquisition unit integrates real-time transmission from buoy stations, mobile sampling databases, and historical fixed-point sampling archives, all formatted uniformly and matched in time and space.
[0126] The physical constraint generative adversarial network building unit realizes the network construction of generator and discriminator, and embeds a piecewise power function physical constraint module and a spatial gradient consistency discrimination mechanism;
[0127] The composite loss function defines the unit configuration of various loss weights and calculation logic;
[0128] The model training unit executes a two-stage training process and manages computing resources and model versions.
[0129] The concentration inversion execution unit loads the trained generator, processes new remote sensing images in batches, and outputs standardized concentration products.
[0130] The system is deployed on a high-performance computing cluster, supports GPU-accelerated inference, and the processing time for a single scene image is less than 5 minutes.
[0131] In summary, this embodiment achieves high-precision, high-resolution, and physically consistent chlorophyll a concentration inversion under extremely sparse supervision by deeply integrating marine optical physics priors and deep generative models, thus solving the fundamental bottleneck of insufficient measured data in nearshore fishery water body monitoring.
[0132] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, 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 process, method, article, or apparatus.
[0133] 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 deep learning-based method for detecting chlorophyll a concentration in nearshore fishery waters, characterized in that, include: Acquire multispectral or hyperspectral remote sensing image data of the target nearshore waters, wherein the remote sensing image data includes reflectance information in the visible to near-infrared bands; Simultaneously acquire sparsely distributed field-measured chlorophyll a concentration data in the water area. The field-measured chlorophyll a concentration data is obtained by buoy stations, mobile sampling or fixed-point sampling, and its spatial coverage density is less than one effective sample point per 100 square kilometers. A physical constraint generative adversarial network is constructed, which includes a generator and a discriminator. The generator takes remote sensing image data as input and outputs a predicted chlorophyll a concentration distribution map. The discriminator is used to distinguish the difference between the concentration map output by the generator and the actual measured concentration map. The nonlinear response relationship between chlorophyll a and remote sensing reflectance derived from the ocean radiative transfer model is embedded in the generator as a physical constraint module. This module expresses the mapping relationship between the remote sensing reflectance ratio and chlorophyll a concentration in different concentration ranges through a preset piecewise power function. A spatial gradient consistency discrimination mechanism is introduced into the discriminator. This mechanism calculates the spatial gradient field of the generated density map and the corresponding remote sensing image, and constrains the alignment between the two on the edge structure. Define a composite loss function, which includes adversarial loss, physical constraint loss, spatiotemporal smoothing loss, and sparse measured point supervision loss; The physical constraint generative adversarial network was trained end-to-end using the composite loss function until the generator could stably output a chlorophyll a concentration distribution map that conformed to physical laws and was consistent with sparse measured data. The trained generator is used as the backbone model for chlorophyll a concentration inversion. Remote sensing image data of the water area to be detected is input, and the chlorophyll a concentration distribution results of the entire water area are output.
2. The method for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning according to claim 1, characterized in that, The remote sensing image data comes from the Haiyang-1 satellite, the Gaofen series satellites, or the Sentinel-2 satellite.
3. The method for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning according to claim 2, characterized in that, The on-site measured chlorophyll a concentration data were determined by acetone extraction-fluorescence method, and the geographical location of the sampling point was corrected by differential GPS.
4. The method for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning according to claim 3, characterized in that, The physical constraint module is embedded in the generator's loss calculation path in a differentiable form. Its parameters remain fixed during training and do not participate in gradient updates.
5. The method for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning according to claim 4, characterized in that, The introduction of a spatial gradient consistency discrimination mechanism in the discriminator includes: Calculate the first-order partial derivatives in the horizontal and vertical directions of the generated concentration map to obtain the concentration gradient magnitude; Calculate the gradient magnitude for the green band of the remote sensing image; The gradient consistency loss is defined as the negative value of the normalized cross-correlation coefficient between the concentration gradient magnitude and the remote sensing image gradient magnitude, and this loss is integrated into the discrimination logic of the discriminator.
6. The method for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning according to claim 5, characterized in that, The spatiotemporal smoothing loss in the composite loss function includes a spatial term and a temporal term; The spatial term is a second-order derivative penalty between neighboring pixels in the generated concentration map; The time term is the sum of squares of the pixel differences between the current time phase generated concentration map and the previous time phase generated concentration map at the same geographical location.
7. The method for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning according to claim 6, characterized in that, The end-to-end training of the physically constrained generative adversarial network using the composite loss function includes: In the first stage, the encoder part of the generator is frozen, and only the decoder and physical constraint module are trained to quickly align sparse test points. The second stage involves unfreezing all parameters, performing end-to-end fine-tuning, and gradually reducing the weighting coefficients of the physical constraint loss.
8. The method for detecting chlorophyll a concentration in nearshore fishery waters based on deep learning according to claim 7, characterized in that, The process of using the trained generator as the backbone model for chlorophyll a concentration inversion, inputting remote sensing image data of the water area to be detected, and outputting the chlorophyll a concentration distribution results for the entire water area, further includes: Outlier removal is performed on the output concentration distribution map, and the outlier removal is based on the three-standard-deviation criterion; Spatial interpolation smoothing is performed on the concentration distribution map after outlier removal, and the spatial interpolation smoothing adopts the Kriging interpolation method.
9. A deep learning-based chlorophyll a concentration detection system for nearshore fishery waters, characterized in that, include: The remote sensing data acquisition unit is used to acquire multispectral or hyperspectral remote sensing image data of the target nearshore waters, wherein the remote sensing image data includes reflectance information in the visible to near-infrared bands. The measured data acquisition unit is used to acquire the sparsely distributed on-site measured chlorophyll a concentration data in the water area. The on-site measured chlorophyll a concentration data is obtained by buoy stations, mobile sampling or fixed-point sampling, and its spatial coverage density is less than one effective sample point per 100 square kilometers. The physical constraint generative adversarial network building unit is used to construct a network architecture that includes a generator and a discriminator. The generator is embedded with a piecewise power function physical constraint module of marine optical priors, and the discriminator integrates a spatial gradient consistency discrimination mechanism. The composite loss function definition unit is used to define a joint optimization objective that includes adversarial loss, physical constraint loss, spatiotemporal smoothing loss and sparse measured point supervision loss; The model training unit is used to train the physical constraint generative adversarial network based on the remote sensing image data and the field measured chlorophyll a concentration data using the composite loss function. The concentration inversion execution unit is used to take the trained generator as the inversion model, input the remote sensing image of the water area to be detected, and output a global chlorophyll a concentration distribution map.
10. The deep learning-based chlorophyll a concentration detection system for nearshore fishery waters according to claim 9, characterized in that, In the physical constraint generative adversarial network construction unit, the piecewise power function physical constraint module is embedded in the loss calculation path of the generator in a differentiable form. Its parameters remain fixed during training and do not participate in gradient updates.