Chlorophyll-a prediction method based on spatial heterogeneity perception graph time sequence adversarial network

Abstract

The chlorophyll a prediction method based on spatial heterogeneity perception graph timing confrontation network relates to the technical field of chlorophyll a prediction, and is used for solving the problems that the subjectivity is strong in response to spatial heterogeneity by artificial partition, the partition boundary is not fine enough, and the statistical characteristic difference in the region is large, etc.The reconstructed daily scale chlorophyll a concentration remote sensing data and numerical simulation sea surface temperature data SST are taken as inputs, through spatial heterogeneity partition based on the time evolution behavior of chlorophyll a, graph convolution network GCN spatial coding, time convolution network TCN time coding and regional discriminator constraint, short-term prediction of the spatial distribution of chlorophyll a concentration in the future several days is realized, so that the representation ability of the model to the inhomogeneous change process of chlorophyll a in the complex offshore sea area is improved.

Description

Technical Field

[0001] This invention relates to the field of chlorophyll a forecasting technology, specifically a chlorophyll a forecasting method based on a spatial heterogeneous sensing graph temporal adversarial network. Background Technology

[0002] Chlorophyll a, as the main photosynthetic pigment of phytoplankton, is often used as a key indicator for assessing phytoplankton biomass and net primary productivity. Therefore, accurate forecasting of short-term chlorophyll a changes is crucial. Current chlorophyll a concentration forecasting methods are mostly based on remote sensing observation data and environmental ancillary variables, using deep learning models for spatiotemporal modeling. Some schemes encode historical chlorophyll a sequences and ancillary variables using improved deep learning models such as ConvLSTM or multi-layer convolutional networks to capture short-term changes and periodic patterns, and enhance the expression of local spatial features through spatial convolution. Building on this, other technical solutions further utilize spatial heterogeneity to divide the study area into multiple sub-regions, constructing temporal and spatial inputs for each sub-region, and extracting temporal evolution and spatial distribution features through spatiotemporal attention networks or LSTM models. Finally, the results from each sub-region are fused to obtain an overall forecast map, thereby better reflecting the spatial heterogeneity characteristics of different sub-regions.

[0003] While existing short-term chlorophyll a forecasting methods can reflect the temporal variation characteristics of marine areas to some extent, the spatiotemporal variation of chlorophyll a concentration in nearshore waters exhibits high complexity. Local biological metabolic activities, riverine and coastal biogenic inputs lead to significant differences in the growth and decay rates of chlorophyll a in different regions. Furthermore, the indirect influence of tides, mixing, and hydrodynamic conditions contributes to the complex heterogeneity of chlorophyll a spatial distribution. To address this spatial heterogeneity, existing methods employ artificial sub-regional division for training. However, this artificial division method suffers from strong subjectivity, insufficiently refined sub-regional boundaries, and significant differences in statistical characteristics within regions. It fails to fully reflect the dynamic behavior of chlorophyll in different sub-regions within the real marine environment, thus limiting the model's ability to characterize local variations and inter-regional correlations. Summary of the Invention

[0004] The purpose of this invention is to provide a chlorophyll a prediction method based on a temporal adversarial network of spatial heterogeneous perception maps, which addresses the problems of strong subjectivity in artificial partitioning in response to spatial heterogeneity, insufficiently refined partition boundaries, and large differences in statistical characteristics within regions.

[0005] The technical solution adopted by this invention to solve its technical problem is: a chlorophyll a prediction method based on spatial heterogeneous perception map temporal adversarial network, which specifically includes the following steps.

[0006] S1 Data Acquisition and Processing: Acquire diurnal chlorophyll a concentration remote sensing data of the study sea area, and perform spatial cropping, land-sea masking, invalid grid removal, logarithmic transformation and missing data reconstruction on the chlorophyll a concentration remote sensing data to form a temporally continuous and spatially complete chlorophyll a concentration field.

[0007] S2 Spatial Heterogeneous Partitioning: Based on the chlorophyll a concentration field obtained in step S1, the effective sea area grid within the study area is spatially heterogeneous to obtain several sub-regions with similar chlorophyll a temporal variation characteristics.

[0008] Based on the spatial heterogeneous partitioning results obtained in step S2, S3 constructs a chlorophyll a short-term forecast GCN-TCN generator that integrates GCN and TCN.

[0009] Based on the spatial heterogeneous partitioning results obtained in step S2, S4 constructs a discriminator to constrain the spatial structure rationality and inter-regional relationship rationality of the generator output results. It also constructs a region discrimination module and a region relationship discrimination module, and builds a chlorophyll a short-term forecasting model that includes spatial heterogeneous partitioning, GCN-TCN generator, region discrimination module, region relationship module and staged adversarial training strategy, namely the Spatial Heterogeneous Perception Map Temporal Adversarial Network SH-GTAN.

[0010] S5 performs sample construction, normalization, and inference restoration on the SH-GTAN model constructed in step S4, restoring the normalized daily variation of chlorophyll a concentration output by the generator to the future chlorophyll a concentration field in the real concentration space.

[0011] Further, step S1 includes the following sub-steps: S1.1 Using chlorophyll a concentration remote sensing products provided by the European Space Agency's Climate Change Initiative Ocean Color Project as samples, and spatially cropping the sample data according to the study area, and extracting effective sea area grids by combining land-sea masking; S1.2 Removing grid points from the samples that have no chlorophyll a concentration remote sensing data throughout the entire study period; S1.3 Performing a base-10 logarithmic transformation on the chlorophyll a concentration remote sensing data; S1.4 Using the DINEOF data interpolation method based on empirical orthogonal functions to reconstruct missing data from the chlorophyll a concentration remote sensing data, and recovering the missing grid values ​​through empirical orthogonal function decomposition and iterative filling process, thereby obtaining a temporally continuous and spatially complete chlorophyll a concentration field; dividing the reconstructed chlorophyll a concentration field into a validation set and a training set; introducing sea surface temperature (SST) data obtained from numerical simulation as an environmental auxiliary variable to characterize the impact of the sea surface thermal environment on short-term changes in chlorophyll a.

[0012] Further, step S2 includes the following sub-steps: S2.1 Take the chlorophyll a concentration time series of each grid point in the effective sea area grid as a sample to be partitioned; S2.2 Perform a fast Fourier transform on the time series and divide the period into multiple period frequency bands, and calculate the energy distribution of each training set in different period frequency bands; S2.3 Normalize the energy value of each grid point in each period frequency band into a frequency band energy composition vector, so that the sum of the square proportions of the energy in each frequency band of each grid point is 1; S2.4 Based on the frequency band energy composition vector obtained in step S2.3, partition the data through K-means clustering analysis, and determine the number of partitions according to the model validation set prediction performance under different numbers of partitions.

[0013] Further, step S3 includes the following sub-steps: S3.1 Treat each grid point of the effective sea area grid in the study area as a graph node, and map the chlorophyll a concentration time series and SST time series corresponding to the graph node as node features; S3.2 Calculate edge weights using a Gaussian function based on the spatial proximity relationship between graph nodes to construct a distance graph; S3.3 Construct a two-layer globally shared GCN encoding layer to extract the local spatial structure features of chlorophyll a concentration and SST; S3.4 Based on the spatial heterogeneous partitioning results obtained in step S2, and on the basis of the globally shared GCN results, set up a region adaptation module for different sub-regions to achieve lightweight region-specific transformation of node features for each sub-region; S3.5 Construct a four-layer TCN encoding module for each sub-region, with the convolution dilation coefficient of each layer increasing sequentially to gradually expand the temporal receptive field, and take the output of the last moment in the historical input window as... For each graph node, extract its temporal evolution features within the historical input window based on its corresponding temporal state features. S3.6 When introducing the SST auxiliary variable, the generator independently spatiotemporally encodes the chlorophyll a concentration sequence and the SST sequence. S3.7 Concatenates the spatial embedding features corresponding to the last moment of the historical input window with the temporal evolution features obtained from TCN encoding, then maps them to a fusion hidden state through a multi-layer fully connected network, fusing the spatial and temporal states of each graph node. S3.8 Sets regional forecast heads for different spatial sub-regions; each regional forecast head outputs the daily change in chlorophyll a concentration for the next 7 days based on the fusion hidden state of the graph nodes in that sub-region. S3.9 After the generator outputs the daily change in chlorophyll a concentration for the next 7 days, it compares it with the actual daily change label of chlorophyll a concentration, and uses mean squared error to construct the generator's supervised error loss.

[0014] Furthermore, in step S3.3, the GCN encoding layer retains both the individual features and the neighborhood weighted average features for each graph node, calculated using the following formula: ;in, Indicates the first The output features of each graph node after GCN encoding; A learnable weight matrix representing its own characteristics; This represents the input features of graph node i; A learnable weight matrix representing neighborhood features; This indicates a weighted average aggregation of features from neighboring nodes. Represents graph nodes The set of neighboring nodes; The distance weight between graph node i and graph node j is represented. The input features of graph node j are represented by GCN. After processing, the model obtains the node spatial embedding features through normalization and nonlinear activation functions.

[0015] Further, step S4 includes the following sub-steps: S4.1 Constructing the historical condition input of the discriminator, the historical condition input receives the same historical input sequence as the generator, including the chlorophyll a concentration sequence and SST sequence of the past several days; S4.2 Constructing the future change input of the discriminator, the future change input receives the daily change sequence of chlorophyll a concentration for the next 7 days, the input includes the real chlorophyll a concentration daily change label sequence as the real sample, and the daily change sequence of chlorophyll a concentration for the next 7 days output by the generator as the generated sample.

[0016] Further, step S4 includes the following sub-steps: S4.3 Perform regional averaging or weighted aggregation on the node features belonging to the same spatial sub-region to obtain the regional embedding features of each sub-region; S4.4 Construct a regional discrimination module to perform true / false discrimination on each regional embedding feature, output the discrimination score of the corresponding sub-region, and summarize the discrimination scores of multiple sub-regions to obtain the overall regional discrimination result; S4.5 Based on the spatial adjacency relationship of the graph nodes of the effective sea area grid on the latitude and longitude plane, determine the pairs of adjacent sub-regions, and then combine the regional embedding features of each pair of sub-regions to form regional relationship features, construct a regional relationship discrimination module, and output the corresponding regional relationship discrimination score; S4.6 Weightedly fuse the overall regional discrimination result obtained in step S4.4 and the regional relationship discrimination score obtained in step S4.5 to obtain the final output score D of the discriminator. total : ;in, The weighting coefficients for the region discrimination term; Indicates the discrimination score within the region; These are the weighting coefficients for the regional relationship discrimination term; The score indicates the relationship between regions.

[0017] Further, step S4 includes the following sub-steps: S4.7 Introduces a gradient-penalized Wasserstein adversarial training strategy to train the discriminator. In each round of training, the true daily variation of chlorophyll a concentration label sequence and the generator's output sequence of the daily variation of chlorophyll a concentration for the next 7 days are first input into the discriminator to obtain the true sample score and the generated sample score; then, linear interpolation is performed between the true daily variation of chlorophyll a concentration label sequence and the generator's output sequence of the daily variation of chlorophyll a concentration for the next 7 days, and the gradient penalty term of the discriminator output on the interpolated sample is calculated to constrain the discriminator to satisfy Lipschitz continuity; the discriminator loss function is: ;in, This represents the discriminator loss function; Represents the mathematical expectation operator; This represents the discriminator function or the discriminator output score; This represents the sequence of future chlorophyll a changes output by the generator; This represents a sequence representing the actual future changes in chlorophyll a. Indicates the gradient penalty weights; S4.8 represents the gradient penalty term; based on the adversarial feedback from the discriminator, construct the adversarial constraint term for the generator. ;in, This represents the adversarial loss obtained from feedback from the discriminator.

[0018] Further, step S4 includes the sub-step: S4.9 Constructing the complete loss function of the generator; during generator training, supervised error loss and adversarial loss are introduced simultaneously; the complete loss function of the generator is: ;in, These are the weighting coefficients for supervised error loss; This indicates that there is a supervised error loss; Weighting coefficients to counteract loss; supervised error loss. The calculation formula is: ;in, Indicates batch size; Indicates the number of days in the future forecast; N represents the number of valid sea area nodes; Indicates the first output of the generator The number of samples, the k-th forecast day, and the... Daily variation of chlorophyll a concentration at each node in the graph. This label represents the daily change in the actual chlorophyll a concentration.

[0019] Further, step S5 includes the following sub-steps: S5.1 Constructing the model input variables and the forecast target. The SH-GTAN model input includes the chlorophyll a concentration sequence and SST sequence for the past several days. The forecast target is set as the daily change in chlorophyll a concentration in the true concentration space. , ;in This represents the chlorophyll a concentration on day t+1. This represents the chlorophyll a concentration on day t. This represents the change in chlorophyll a concentration from day t to day t+1; S5.2 performs maximum-min normalization on the logarithmic chlorophyll a concentration remote sensing data variable and the SST variable: ;in, Represents the normalized variable value; This represents the variable to be normalized; Represents variables in the training set The minimum value; Represents variables in the training set The maximum value; for the daily variation label of chlorophyll a concentration, a symmetrical normalization using an inverse hyperbolic sine function is adopted to reduce the impact of abnormally large variations on the training process; in the inference stage of the SH-GTAN model, S5.3 transforms the normalized daily variation output by the generator back to the true concentration space to obtain the chlorophyll a concentration variation for each future day; then, using the true chlorophyll a concentration on the reporting date as the initial field, the future chlorophyll a concentration is recovered by daily accumulation: ;in, This represents the chlorophyll a concentration predicted by the SH-GTAN model on day t+k. This represents the chlorophyll a concentration predicted by the SH-GTAN model on day t+k-1. This represents the change in chlorophyll a concentration predicted by the SH-GTAN model from day t+k-1 to day t+k.

[0020] The beneficial effects of this invention are: the spatial partitioning of this invention is objective, and the clustering and partitioning based on chlorophyll a behavior features avoids the problems of strong subjectivity and rough boundaries of manual partitioning; the adaptability to complex nearshore sea areas is enhanced, and the complex coastline and heterogeneous nodes are handled by GCN, so that the model can take into account both the shared spatial information of the whole sea area and the local differential changes; the spatial rationality of the forecast results is improved, and the generator output is constrained by the region discriminator and the region relationship discriminator, so as to take into account both the consistency within the sub-region and the linkage features between regions. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2Comparison of chlorophyll a forecast performance under different historical input days;

[0023] Figure 3 This is a graph showing how the RMSE of the validation set changes with the number of training epochs during model training.

[0024] Figure 4 This is a forecast performance chart of the 7-day average chlorophyll a from the reporting date.

[0025] Figure 5 A graph showing the variation of chlorophyll a forecast performance over the next 1-7 days with the forecast lead time;

[0026] Figure 6 A comparison of the forecast performance of SH-GTAN and ConvLSTM within the extreme performance percentile range;

[0027] Figure 7 A scatter plot comparing the predicted and actual values ​​of chlorophyll a from different models.

[0028] Figure 7 In this context, Chl-a represents chlorophyll a. Detailed Implementation

[0029] like Figure 1 As shown, this invention uses reconstructed diurnal chlorophyll a concentration remote sensing data and numerically simulated sea surface temperature (SST) data as input. Through spatial heterogeneous partitioning based on chlorophyll a temporal evolution behavior, graph convolutional network (GCN) spatial encoding, temporal convolutional network (TCN) temporal encoding, and region discriminator constraints, it achieves short-term forecasts of the spatial distribution of chlorophyll a concentration over several days, thereby improving the model's ability to characterize the heterogeneous changes in chlorophyll a in complex nearshore sea areas. This invention provides a chlorophyll a forecasting method based on a spatially heterogeneous sensing graph temporal adversarial network, specifically including the following steps.

[0030] S1 Data Acquisition and Processing.

[0031] We acquired diurnal chlorophyll a concentration remote sensing data for the study sea area, and then performed spatial cropping, land-sea masking, invalid grid removal, logarithmic transformation, and missing data reconstruction on the chlorophyll a concentration remote sensing data to form a temporally continuous and spatially complete chlorophyll a concentration field.

[0032] S1.1 uses chlorophyll a concentration remote sensing products provided by the European Space Agency's Climate Change Initiative Ocean Color Project as the data source, selects annual and daily chlorophyll a concentration remote sensing data for the study period as samples, and spatially crops the sample data according to the study sea area, and extracts the effective sea area grid by combining the land-sea mask.

[0033] S1.2 Grid points that have no chlorophyll a concentration remote sensing data at all during the entire study period are removed from the sample.

[0034] S1.3 performs a base-10 logarithmic transformation on the chlorophyll a concentration remote sensing data for subsequent data reconstruction, spatial partitioning, and model training.

[0035] S1.4 employs the DINEOF data interpolation method based on empirical orthogonal functions to reconstruct missing data from chlorophyll a concentration remote sensing data, addressing the common problem of high missing data rates in diurnal chlorophyll a concentration remote sensing data. Missing grid values ​​are recovered through empirical orthogonal function decomposition and iterative filling processes, thereby obtaining a temporally continuous and spatially complete chlorophyll a concentration field. In the reconstructed chlorophyll a concentration field, data from the study period are divided into a validation set and a training set. In addition to chlorophyll a concentration remote sensing data, this invention also introduces SST data obtained from numerical simulation as an environmental auxiliary variable to characterize the impact of sea surface thermal environment on short-term chlorophyll a changes and to assist the model in extracting environmental driving information in chlorophyll a changes.

[0036] S2 spatial heterogeneous partitioning.

[0037] Based on the chlorophyll a concentration field obtained in step S1, the effective sea area grid within the study area is spatially heterogeneous to obtain several sub-regions with similar chlorophyll a temporal variation characteristics, providing a spatial basis for subsequent partitioning modeling, regional feature extraction, and regional relationship discrimination.

[0038] S2.1 uses the chlorophyll a concentration time series of each grid point in the effective sea area grid as a sample to be partitioned. To avoid using validation year information in spatial partitioning and model parameter training, frequency domain features are extracted only from the chlorophyll a time series in the training set for spatial partitioning.

[0039] S2.2 performs a Fast Fourier Transform on the time series and divides the period into five periodic frequency bands: greater than 180 days, 90-180 days, 30-90 days, 10-30 days, and 3-10 days. It then calculates the energy distribution of each training set within different periodic frequency bands.

[0040] S2.3 normalizes the energy value of each grid point in each period frequency band into a frequency band energy composition vector, so that the sum of the square proportions of the energy in each frequency band of each grid point is 1.

[0041] S2.4 Based on the frequency band energy composition vector obtained in step S2.3, K-means clustering analysis is used to partition the data, and the number of partitions is determined according to the prediction performance of the model validation set under different numbers of partitions.

[0042] Based on the spatial heterogeneous partitioning results obtained in step S2, S3 constructs a short-term chlorophyll a forecast generator that integrates GCN and TCN.

[0043] S3.1 treats each grid point of the effective sea area grid in the study area as a graph node, and maps the chlorophyll a concentration time series and SST time series corresponding to the graph node as node features.

[0044] S3.2 Calculate edge weights using a Gaussian function based on the spatial proximity relationships between graph nodes to construct a distance graph. Furthermore, add self-connecting edges to each graph node to preserve its own information.

[0045] S3.3 Constructs a two-layer globally shared GCN coding layer to extract chlorophyll a concentration and local spatial structure features of SST.

[0046] The GCN encoding layer retains both the individual features and the neighborhood weighted average features of each graph node. The calculation formula is as follows: .in, Indicates the first The output features of each graph node after GCN encoding; A learnable weight matrix representing its own characteristics; This represents the input features of graph node i; A learnable weight matrix representing neighborhood features; This indicates a weighted average aggregation of features from neighboring nodes. Represents graph nodes The set of neighboring nodes; The distance weight between graph node i and graph node j is represented. This represents the input features of graph node j. After GCN processing, the model obtains the node spatial embedding features through normalization and nonlinear activation functions.

[0047] S3.4 Based on the spatial heterogeneous partitioning results obtained in step S2, and on the basis of sharing the GCN results across the entire domain, a region adaptation module is set up for different sub-regions to achieve lightweight region-specific transformation of the node features of each sub-region.

[0048] S3.5 constructs a 4-layer TCN encoding module for each sub-region, with the convolution dilation coefficient of each layer increasing sequentially to gradually expand the temporal receptive field. The output of the last moment in the historical input window is taken as the temporal state feature of the corresponding graph node, and the temporal evolution feature of each graph node within the historical input window is extracted.

[0049] S3.6 When the SST auxiliary variable is introduced, the generator performs independent spatiotemporal encoding on the chlorophyll a concentration sequence and the SST sequence respectively.

[0050] S3.7 concatenates the spatial embedding features corresponding to the last moment of the historical input window with the temporal evolution features obtained by TCN encoding, and then maps them into fused hidden states through a multi-layer fully connected network to fuse the spatial and temporal states of each graph node.

[0051] S3.8 sets up regional forecast heads for different spatial sub-regions. Each regional forecast head outputs the daily change in chlorophyll a concentration for the next 7 days based on the fusion hidden state of the map nodes in that sub-region.

[0052] After the S3.9 generator outputs the daily change in chlorophyll a concentration for the next 7 days, it compares it with the actual daily change in chlorophyll a concentration label, and uses mean squared error to construct the generator's supervised error loss.

[0053] Based on the spatial heterogeneous partitioning results obtained in step S2, S4 constructs a discriminator to constrain the rationality of the spatial structure and the rationality of the inter-regional relationships in the generator output results.

[0054] S4.1 Construct the historical condition input for the discriminator. The historical condition input receives the same historical input sequence as the generator, including chlorophyll a concentration sequences and SST sequences from the past few days.

[0055] S4.2 Constructing the future change input for the discriminator. The future change input receives the daily change sequence of chlorophyll a concentration for the next 7 days. The input includes the real chlorophyll a concentration daily change label sequence (as the real sample) and the generator's output daily change sequence of chlorophyll a concentration for the next 7 days (as the generated sample). The discriminator uses a spatiotemporal coding structure similar to the generator, with one layer for the GCN and two layers for the TCN.

[0056] S4.3 Perform regional averaging or weighted aggregation on the node features belonging to the same spatial sub-region to obtain the regional embedding features of each sub-region.

[0057] S4.4 constructs a region discrimination module, which performs true / false discrimination on the embedded features of each region, outputs the discrimination score of the corresponding sub-region, and summarizes the discrimination scores of multiple sub-regions to obtain the overall region discrimination result.

[0058] S4.5 determines adjacent sub-region pairs based on the spatial adjacency relationship of graph nodes in the latitude and longitude plane of the effective sea area grid, then combines the regional embedding features of each pair of sub-regions to form regional relationship features, constructs a regional relationship discrimination module, and outputs the corresponding regional relationship discrimination score.

[0059] S4.6 The overall region discrimination result obtained in step S4.4 and the region relationship discrimination score obtained in step S4.5 are weighted and fused to obtain the final output score D of the discriminator. total : .in, The weighting coefficients for the region discrimination term; Indicates the discrimination score within the region; These are the weighting coefficients for the regional relationship discrimination term; The score indicates the relationship between regions.

[0060] S4.7 introduces a gradient-penalized Wasserstein adversarial training strategy to train the discriminator. Specifically, in each training round, the discriminator is first input with the true daily chlorophyll a concentration label sequence and the generator's output daily chlorophyll a concentration change sequence for the next 7 days, respectively, to obtain the true sample score and the generated sample score. Then, linear interpolation is performed between the true daily chlorophyll a concentration label sequence and the generator's output daily chlorophyll a concentration change sequence for the next 7 days, and the gradient penalty term of the discriminator output on the interpolated sample is calculated to constrain the discriminator to satisfy Lipschitz continuity. The discriminator loss function is: .in, This represents the discriminator loss function; Represents the mathematical expectation operator; This represents the discriminator function or the discriminator output score; This represents the sequence of future chlorophyll a changes output by the generator; This represents a sequence representing the actual future changes in chlorophyll a. Indicates the gradient penalty weights; This represents the gradient penalty term.

[0061] S4.8 Based on the adversarial feedback from the discriminator, construct the adversarial constraint terms for the generator. .in, This represents the adversarial loss obtained from feedback from the discriminator.

[0062] S4.9 Construct the complete loss function for the generator. During generator training, both supervised error loss and adversarial loss are introduced. The complete loss function for the generator is: .in, These are the weighting coefficients for supervised error loss; This indicates that there is a supervised error loss; Weighting coefficients to counteract the loss. Supervised error loss. Using the mean squared error form, the calculation formula is as follows: .in, Indicates batch size; Indicates the number of days in the future forecast; N represents the number of valid sea area nodes; Indicates the first output of the generator The number of samples, the k-th forecast day, and the... Daily variation of chlorophyll a concentration at each node in the graph. This label represents the daily change in the actual chlorophyll a concentration.

[0063] To leverage the discriminator's role in supervising large-scale processes while preventing it from compromising model stability, it is necessary to assign [a certain function] during the early stages of model training. Larger value, medium term Gradually reduce to 0, and later rely solely on The method imposes constraints and requires the discriminator to be updated twice before each generator update. The early, middle, and late stages each account for one-third of the total training rounds. This approach allows the discriminator to quickly develop effective discrimination capabilities in the early stages, efficiently supervises large-scale processes, and avoids its adverse impact on generator numerical convergence in the later stages of training.

[0064] The short-term chlorophyll a forecasting model, which consists of spatial heterogeneous partitioning, a GCN-TCN generator, a region discrimination module, a region relationship discrimination module, and a staged adversarial training strategy, is named Spatial Heterogeneous Perception Graph Temporal Adversarial Network (SH-GTAN).

[0065] S5 performs sample construction, normalization, and inference restoration on the SH-GTAN model constructed in step S4, restoring the normalized daily variation of chlorophyll a concentration output by the generator to the future chlorophyll a concentration field in the real concentration space.

[0066] S5.1 Constructing Model Input Variables and Forecast Targets: The SH-GTAN model inputs include chlorophyll a concentration sequences and SST sequences from the past several days. The forecast target is not directly set as the future absolute chlorophyll a concentration, but rather as the daily variation of chlorophyll a concentration in the true concentration space. , .in This represents the chlorophyll a concentration on day t+1. This represents the chlorophyll a concentration on day t. This represents the change in chlorophyll a concentration from day t to day t+1.

[0067] S5.2 Perform maximum-min normalization on the logarithmic chlorophyll a concentration remote sensing data variable and the SST variable: .in, Represents the normalized variable value; This represents the variable to be normalized; Represents variables in the training set The minimum value; Represents variables in the training set The maximum value. For the daily variation label of chlorophyll a concentration, symmetric normalization is performed using an inverse hyperbolic sine function to reduce the impact of abnormally large variations on the training process.

[0068] S5.3 In the inference phase of the SH-GTAN model, the normalized daily variation output by the generator is inversely transformed back to the true concentration space to obtain the chlorophyll a concentration variation for each future day. Then, using the true chlorophyll a concentration on the reporting date as the initial field, the future chlorophyll a concentration is recovered using a daily accumulation method: .in, This represents the chlorophyll a concentration predicted by the SH-GTAN model on day t+k. This represents the chlorophyll a concentration predicted by the SH-GTAN model on day t+k-1. This represents the change in chlorophyll a concentration predicted by the SH-GTAN model from day t+k-1 to day t+k.

[0069] The performance of the SH-GTAN model is evaluated below, verifying its numerical accuracy, temporal stability, forecast lead decay characteristics, and forecast reliability under complex conditions in the 7-day forecast task, and comparing it with baseline models. A Convolutional Long Short-Term Memory (ConvLSTM) model and a persistent forecast model are set as comparison baselines. The ConvLSTM model and the SH-GTAN model use the same training and validation set partitioning method. The main difference is that ConvLSTM does not introduce spatial heterogeneous partitioning, graph structures, or adversarial constraints, but extracts the local spatiotemporal features of chlorophyll a and SST through the LSTM structure. The persistent forecast model does not undergo model training; it directly uses the chlorophyll a concentration field on the first day of forecasting as the forecast result for each future day, used to measure the improvement effect of the deep learning model compared to the simple persistent assumption. Under the same hyperparameter conditions, the SH-GTAN model and the ConvLSTM model are trained and forecasted in parallel three times each to understand the average performance and stability of the models.

[0070] The model's forecasting performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE), correlation coefficient (CC), and percentage bias (PBIAS). The calculation formulas are as follows: ; ; ; Where n represents the number of valid sea area samples used in the calculation; This represents the actual concentration of chlorophyll a. This indicates the predicted chlorophyll a concentration; This represents the average of the true values; This represents the average value of the forecast.

[0071] Model forecast performance under different historical input days, such as Figure 2As shown in the figure. Within the 1-21 day range, the SH-GTAN model achieved better results when the input historical data was 17 consecutive days, while the ConvLSTM model achieved better results when the input historical data was 21 consecutive days. Therefore, in subsequent comparative experiments, the SH-GTAN model used 17 consecutive days of historical input, and the ConvLSTM model used 21 consecutive days of historical input. During model training, the RMSE of the validation set changed with the number of training epochs as shown in the figure. Figure 3 As shown, the SH-GTAN model exhibits relatively large fluctuations in RMSE on the validation set during the early training stages due to the introduction of spatial heterogeneity modeling and adversarial constraints; however, the RMSE gradually decreases and stabilizes as the number of training epochs increases. The 7-day average forecast performance for each reporting day during the study period is shown in the figure. Figure 4 As shown, the SH-GTAN model outperformed persistent forecast models on most lead times, indicating that the SH-GTAN model can effectively predict short-term changes in chlorophyll a concentration. However, its forecast performance was relatively weaker during certain periods of spring and winter. Model performance for different forecast lead times 1-7 days is shown below. Figure 5 As shown, the forecast performance of all three models deteriorates with increasing forecast lead time. Figure 5 As shown in PBIAS (d), the SH-GTAN model exhibits significantly better stability than the ConvLSTM model. To further compare the performance of the three models under adverse forecasting conditions, the forecast results for the entire year during the study period were ranked by performance percentile, as shown below. Figure 6 As shown in the figure. For R² and CC, the focus is on comparing the lower percentile intervals; for the absolute values ​​of RMSE and PBIAS, the focus is on comparing the higher percentile intervals. The results show that the advantage of the SH-GTAN model over the ConvLSTM model is mainly reflected in the extremely low performance or high error intervals, that is, on the difficult-to-predict start date, the SH-GTAN model can reduce errors and improve forecast stability. The scatter distribution of the forecast values ​​and the actual values ​​is shown in the figure. Figure 7 As shown. Figure 7 In the SH-GTAN and ConvLSTM models, the scatter points are more concentrated overall. The proximity of the reference line indicates that both deep learning models can reproduce the spatial variation of chlorophyll a concentration well. Combining the R², RMSE, CC, and PBIAS results, the SH-GTAN model shows lower overall error and better consistency between predicted and actual values.

[0072] This invention achieves objective spatial partitioning by using chlorophyll a behavior characteristics for clustering, avoiding the subjectivity and coarse boundaries of manual partitioning. It enhances adaptability to complex nearshore waters by using GCN to handle complex coastlines and heterogeneous nodes, enabling the model to consider both shared spatial information across the entire sea area and localized variations. The spatial rationality of forecast results is improved by constraining the generator output through region discriminators and region relationship discriminators, achieving a balance between consistency within sub-regions and inter-regional linkage features. In summary, this invention solves the problems of subjective handling of spatial heterogeneity, insufficiently refined partition boundaries, and significant differences in statistical characteristics within regions caused by manual partitioning.

Claims

1. A chlorophyll a prediction method based on a spatially heterogeneous sensing map temporal adversarial network, characterized in that, Includes the following steps: S1 Data Acquisition and Processing: Acquire diurnal chlorophyll a concentration remote sensing data of the study sea area, and perform spatial cropping, land-sea masking, invalid grid removal, logarithmic transformation and missing data reconstruction on the chlorophyll a concentration remote sensing data to form a temporally continuous and spatially complete chlorophyll a concentration field. S2 Spatial Heterogeneous Partitioning: Based on the chlorophyll a concentration field obtained in step S1, the effective sea area grid within the study area is spatially heterogeneous to obtain several sub-regions with similar chlorophyll a temporal variation characteristics. Based on the spatial heterogeneous partitioning results obtained in step S2, S3 constructs a chlorophyll a short-term forecast GCN-TCN generator that integrates GCN and TCN; Based on the spatial heterogeneous partitioning results obtained in step S2, S4 constructs a discriminator to constrain the spatial structure rationality and inter-regional relationship rationality of the generator output results, constructs a region discrimination module and a region relationship discrimination module, and constructs a chlorophyll a short-term forecasting model that includes spatial heterogeneous partitioning, GCN-TCN generator, region discrimination module, region relationship module and staged adversarial training strategy, namely the spatial heterogeneous perception map temporal adversarial network SH-GTAN. S5 performs sample construction, normalization, and inference restoration on the SH-GTAN model constructed in step S4, restoring the normalized daily variation of chlorophyll a concentration output by the generator to the future chlorophyll a concentration field in the real concentration space.

2. The chlorophyll a prediction method based on spatial heterogeneous sensing graph temporal adversarial network according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.1 Using chlorophyll a concentration remote sensing products provided by the European Space Agency's Climate Change Initiative Ocean Color Project as samples, and spatially cropping the samples according to the study area, and extracting effective sea area grids by combining land-sea masking; S1.2 Removing grid points from the samples that have no chlorophyll a concentration remote sensing data throughout the entire study period; S1.3 Performing a base-10 logarithmic transformation on the chlorophyll a concentration remote sensing data; S1.4 Using the DINEOF data interpolation method based on empirical orthogonal functions to reconstruct missing data from the chlorophyll a concentration remote sensing data, and recovering the missing grid values ​​through empirical orthogonal function decomposition and iterative filling process, thereby obtaining a temporally continuous and spatially complete chlorophyll a concentration field; dividing the reconstructed chlorophyll a concentration field into a validation set and a training set; introducing sea surface temperature (SST) data obtained from numerical simulation as an environmental auxiliary variable to characterize the impact of the sea surface thermal environment on short-term changes in chlorophyll a.

3. The chlorophyll a prediction method based on spatial heterogeneous perception graph temporal adversarial network according to claim 2, characterized in that, Step S2 includes the following sub-steps: S2.1 Take the chlorophyll a concentration time series of each grid point of the effective sea area grid as a sample to be partitioned; S2.2 Perform a fast Fourier transform on the time series and divide it into multiple periodic frequency bands, and calculate the energy distribution of each training set in different periodic frequency bands. S2.3 normalizes the energy value of each grid point in each period frequency band into a frequency band energy composition vector, so that the sum of the square proportions of the energy in each frequency band of each grid point is 1; S2.4 Based on the frequency band energy composition vector obtained in step S2.3, K-means clustering analysis is used to partition the data, and the number of partitions is determined according to the prediction performance of the model validation set under different numbers of partitions.

4. The chlorophyll a prediction method based on spatial heterogeneous sensing graph temporal adversarial network according to claim 3, characterized in that, Step S3 includes the following sub-steps: S3.1 Treat each grid point of the effective sea area grid of the study area as a graph node, and map the chlorophyll a concentration time series and SST time series corresponding to the graph node as node features; S3.2 Calculate edge weights using a Gaussian function based on the spatial proximity relationships between graph nodes to construct a distance graph; S3.3 Construct a two-layer globally shared GCN encoding layer to extract the local spatial structure features of chlorophyll a concentration and SST; S3.4 Based on the spatial heterogeneous partitioning results obtained in step S2, and building upon the globally shared GCN results, set up region adaptation modules for different sub-regions to achieve lightweight region-specific transformations of node features in each sub-region; S3.5 Construct a four-layer TCN encoding module for each sub-region, with the convolution dilation coefficient of each layer increasing sequentially to gradually expand the temporal receptive field, and take the output of the last moment in the historical input window as the temporal state feature of the corresponding graph node to extract the temporal evolution features of each graph node within the historical input window; S3.

6. When the SST auxiliary variable is introduced, the generator performs independent spatiotemporal encoding on the chlorophyll a concentration sequence and the SST sequence respectively; S3.7 The spatial embedding feature corresponding to the last moment of the historical input window is concatenated with the temporal evolution feature obtained by TCN encoding, and then mapped to the fusion hidden state through a multi-layer fully connected network to fuse the spatial and temporal states of each graph node; S3.8 Regional forecast heads are set for different spatial sub-regions. Each regional forecast head outputs the daily change of chlorophyll a concentration for the next 7 days based on the fusion hidden state of the graph node of that sub-region; S3.9 After the generator outputs the daily change of chlorophyll a concentration for the next 7 days, it is compared with the label of the actual daily change of chlorophyll a concentration, and the mean squared error is used to construct the supervised error loss of the generator.

5. The chlorophyll a prediction method based on spatial heterogeneous sensing graph temporal adversarial network according to claim 4, characterized in that, In step S3.3, the GCN encoding layer retains both the local features and the neighborhood weighted average features for each graph node. The calculation formula is as follows: ;in, Indicates the first The output features of each graph node after GCN encoding; A learnable weight matrix representing its own characteristics; This represents the input features of graph node i; A learnable weight matrix representing neighborhood features; This indicates a weighted average aggregation of features from neighboring nodes. Represents graph nodes The set of neighboring nodes; The distance weight between graph node i and graph node j is represented. The input features of graph node j are represented by GCN. After processing, the model obtains the node spatial embedding features through normalization and nonlinear activation functions.

6. The chlorophyll a prediction method based on spatial heterogeneous sensing graph temporal adversarial network according to claim 5, characterized in that, Step S4 includes the following sub-steps: S4.1 Constructing the historical conditional input of the discriminator, the historical conditional input receives the same historical input sequence as the generator, including the chlorophyll a concentration sequence and SST sequence of the past several days; S4.2 Construct the future change input of the discriminator. The future change input receives the daily change sequence of chlorophyll a concentration for the next 7 days. Its input includes the real chlorophyll a concentration daily change label sequence as the real sample, and the daily change sequence of chlorophyll a concentration for the next 7 days output by the generator as the generated sample.

7. The chlorophyll a prediction method based on spatial heterogeneous sensing graph temporal adversarial network according to claim 6, characterized in that, Step S4 includes the following sub-steps: S4.3 Perform regional averaging or weighted aggregation on the node features belonging to the same spatial sub-region to obtain the regional embedding features of each sub-region; S4.4 constructs a region discrimination module, which performs true / false discrimination on the embedded features of each region, outputs the discrimination score of the corresponding sub-region, and summarizes the discrimination scores of multiple sub-regions to obtain the overall region discrimination result; S4.5 Based on the spatial adjacency relationship of graph nodes in the effective sea area grid on the latitude and longitude plane, it determines adjacent sub-region pairs, then combines the regional embedding features of each pair of sub-regions to form regional relationship features, constructs a regional relationship discrimination module, and outputs the corresponding regional relationship discrimination score; S4.6 The overall region discrimination result obtained in step S4.4 and the region relationship discrimination score obtained in step S4.5 are weighted and fused to obtain the final output score D of the discriminator. total : ;in, The weighting coefficients for the region discrimination term; Indicates the discrimination score within the region; These are the weighting coefficients for the regional relationship discrimination term; The score indicates the relationship between regions.

8. The chlorophyll a prediction method based on spatial heterogeneous sensing graph temporal adversarial network according to claim 7, characterized in that, Step S4 includes the following sub-step: S4.7 Introduces a gradient-penalized Wasserstein adversarial training strategy to train the discriminator. In each training round, the true daily variation of chlorophyll a concentration label sequence and the generator's output of the 7-day daily variation of chlorophyll a concentration are first input into the discriminator to obtain the true sample score and the generated sample score. Then, linear interpolation is performed between the true daily variation of chlorophyll a concentration label sequence and the generator's output of the 7-day daily variation of chlorophyll a concentration, and the gradient penalty term of the discriminator output on the interpolated sample is calculated to constrain the discriminator to satisfy Lipschitz continuity. The discriminator loss function is: ;in, This represents the discriminator loss function; Represents the mathematical expectation operator; This represents the discriminator function or the discriminator output score; This represents the sequence of future chlorophyll a changes output by the generator; This represents a sequence representing the actual future changes in chlorophyll a. Indicates the gradient penalty weights; S4.8 represents the gradient penalty term; based on the adversarial feedback from the discriminator, construct the adversarial constraint term for the generator. ;in, This represents the adversarial loss obtained from feedback from the discriminator.

9. The chlorophyll a prediction method based on a spatially heterogeneous sensing map temporal adversarial network according to claim 8, characterized in that, Step S4 includes the sub-step: S4.9 Constructing the complete loss function of the generator; During generator training, supervised error loss and adversarial loss are introduced simultaneously; The complete loss function of the generator is: ;in, These are the weighting coefficients for supervised error loss; This indicates that there is a supervised error loss; Weighting coefficients to counteract loss; supervised error loss. The calculation formula is: ;in, Indicates batch size; Indicates the number of days in the future forecast; N represents the number of valid sea area nodes; Indicates the first output of the generator The number of samples, the k-th forecast day, and the... Daily variation of chlorophyll a concentration at each node in the graph. This label represents the daily change in the actual chlorophyll a concentration.

10. The chlorophyll a prediction method based on a spatially heterogeneous sensing map temporal adversarial network according to claim 9, characterized in that, Step S5 includes the following sub-steps: S5.1 Constructing the model input variables and forecast target. The SH-GTAN model input includes the chlorophyll a concentration sequence and SST sequence for the past several days. The forecast target is set as the daily change in chlorophyll a concentration in the true concentration space. , ;in This represents the chlorophyll a concentration on day t+1. This represents the chlorophyll a concentration on day t. This represents the change in chlorophyll a concentration from day t to day t+1; S5.2 performs maximum-min normalization on the logarithmic chlorophyll a concentration remote sensing data variable and the SST variable: ;in, Represents the normalized variable value; This represents the variable to be normalized; Represents variables in the training set The minimum value; Represents variables in the training set The maximum value; for the daily variation label of chlorophyll a concentration, a symmetrical normalization using an inverse hyperbolic sine function is adopted to reduce the impact of abnormally large variations on the training process; in the inference stage of the SH-GTAN model, S5.3 transforms the normalized daily variation output by the generator back to the true concentration space to obtain the chlorophyll a concentration variation for each future day; then, using the true chlorophyll a concentration on the reporting date as the initial field, the future chlorophyll a concentration is recovered by daily accumulation: ;in, This represents the chlorophyll a concentration predicted by the SH-GTAN model on day t+k. This represents the chlorophyll a concentration predicted by the SH-GTAN model on day t+k-1. This represents the change in chlorophyll a concentration predicted by the SH-GTAN model from day t+k-1 to day t+k.

Images