Offshore chlorophyll-a concentration prediction method and system based on spatio-temporal feature fusion

Abstract

The present application belongs to the technical field of deep learning, and discloses a near-shore chlorophyll-a concentration prediction method and system based on space-time feature fusion. First, original chlorophyll-a remote sensing data, meteorological remote sensing data and buoy chlorophyll-a monitoring data are acquired, and data preprocessing is performed to obtain continuous space-time sequence data. Then, based on spatial heterogeneity, the sea area to be predicted is divided into several sub-zones. For each sub-zone, time-based input data and space-based input data are prepared, and the space-time attention feature fusion prediction network is used to obtain the time evolution feature and the space evolution feature of chlorophyll-a in each sub-zone. Finally, the time evolution feature map and the space evolution feature map of each sub-zone are fused to obtain the sub-zone prediction result, all sub-zone prediction results are combined to obtain a long-time continuous near-shore chlorophyll-a prediction map, and the problems of insufficient space-time feature mining and uneven distribution of near-shore chlorophyll-a concentration data are solved.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, and specifically relates to a method and system for predicting nearshore chlorophyll a concentration based on spatiotemporal feature fusion. Background Technology

[0002] Chlorophyll a (Chl-a) concentration is one of the important variables used to study the environmental effects of marine dynamic processes. As an indicator of changes in aquatic ecosystems and an important measure of marine primary productivity, changes in chlorophyll a concentration indirectly reflect changes in eutrophication. Therefore, predicting the spatiotemporal variations of Chl-a concentration is of great significance for preventing eutrophication in nearshore waters, utilizing fishery resources, and understanding the global carbon cycle.

[0003] With the increasing volume and variety of monitoring data, machine learning, especially deep learning, has been used to predict Chl-a concentration. Essentially, this is a spatiotemporal sequence prediction problem, that is, predicting future Chl-a concentration evolution based on historical Chl-a concentrations and exogenous factors (such as temperature and salinity). Currently, most machine learning-based Chl-a concentration prediction methods are point-source predictions based on monitoring data. For example, early Chl-a concentration prediction methods based on backpropagation neural networks and support vector machines achieved some success compared to earlier traditional methods, but these networks did not consider the temporal dependencies between multiple historical inputs. Recurrent neural networks, which emerged later, incorporate the output of the previous historical time step into the next hidden layer for training. For example, Long Short-Term Memory (LSTM) networks, by introducing additional gating mechanisms to control feature flow and loss, address the long-term dependencies between multiple memory input sequences and have been widely applied to Chl-a concentration prediction.

[0004] However, existing methods only focus on the simple application of deep neural network models, predicting Chl-a concentrations based on a single point source or multiple independent point sources, and only studying temporal-scale dependent changes. Moreover, this method of predicting Chl-a concentrations from point sources using monitoring data is difficult to achieve large-scale spatiotemporal continuous prediction. The evolution trend of Chl-a concentration is spatiotemporally dynamic, and the occurrence of harmful algal blooms also has a certain regional range. Studying only independent point sources cannot achieve refined prediction of the region. Therefore, it is more meaningful to use spatiotemporal grid data for regional area source Chl-a concentration prediction. In addition, existing methods have the following problems:

[0005] First, the spatiotemporal correlation mining is incomplete, making accurate long-term time-series predictions difficult. Firstly, the neighborhood correlation of stations is not considered. Most prediction methods only address a single point source or process multiple point sources within a region independently, considering only its own time-series evolution characteristics but neglecting the influence of its spatial neighbors, thus hindering large-scale spatiotemporal predictions. Secondly, there is a lack of spatiotemporal feature fusion and correlation. The future changes of each station are influenced by multiple factors, including its own temporal evolution and the spatial neighborhood. However, current chlorophyll a concentration methods do not consider spatiotemporal fusion and correlation modeling, resulting in insufficient mining of spatiotemporal evolution correlation features and low prediction accuracy.

[0006] Second, the uneven distribution of Chl-a concentration data in nearshore areas was not considered, and the spatiotemporal distribution characteristics of chlorophyll a in nearshore areas were not studied in depth. The Chl-a concentration values ​​vary widely across different regions or stations, and the frequencies of different concentration values ​​are highly unbalanced. Sparse high-concentration sample values ​​are not sufficiently learned during model training, leading to larger prediction errors. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for predicting nearshore chlorophyll a concentration based on spatiotemporal feature fusion, solving the problems of insufficient spatiotemporal feature mining and uneven distribution of nearshore chlorophyll a concentration data.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A method for predicting nearshore chlorophyll a concentration based on spatiotemporal feature fusion includes the following steps:

[0010] Step 1: Acquire raw chlorophyll a remote sensing data, meteorological remote sensing data, and buoy chlorophyll a monitoring data;

[0011] Step 2: Data preprocessing to complete the chlorophyll a continuous spatiotemporal sequence data;

[0012] Step 3, Spatial Heterogeneity Partitioning: Based on spatial heterogeneity, the sea area to be predicted is divided into several sub-partitions according to the chlorophyll a continuous spatiotemporal sequence data obtained in Step 2.

[0013] Step 4: Construct a spatiotemporal attention feature fusion prediction network. For each sub-partition after step 3, prepare time-based input data and space-based input data, and use the spatiotemporal attention feature fusion prediction network to obtain the temporal and spatial evolution features of chlorophyll a in each sub-partition.

[0014] The spatiotemporal attention feature fusion prediction network includes a temporal feature extraction module based on an attention mechanism, a spatial feature extraction module based on local and global features, and a spatiotemporal feature fusion prediction module. The temporal feature extraction module obtains the temporal evolution feature map of chlorophyll a, and the spatial feature extraction module obtains the spatial evolution feature map of chlorophyll a.

[0015] Step 5: Use the spatiotemporal feature fusion prediction module to perform post-fusion of the temporal evolution feature map and spatial evolution feature map of each sub-partition to obtain the chlorophyll a prediction map of the next preceding time.

[0016] Step 6: Iterate through the regional predictions, merge all sub-region prediction results, and obtain a long-term continuous chlorophyll a prediction map of the entire nearshore region.

[0017] Furthermore, in step 2, the specific data preprocessing method is as follows: First, the corresponding latitude and longitude grid point data are filled using effective buoy chlorophyll a monitoring data; then, the regional spatiotemporal missing values ​​are filled using the empirical orthogonal function interpolation method; finally, the data with filled values ​​less than 0 are filled using four-way geographic embedding average imputation, as shown in the following formula:

[0018]

[0019] In the formula, q is 4, indicating that four-directional geographic embedding is selected; i represents the specific embedding direction; and x... i x' represents the chlorophyll a concentration value at the grid point in the i-direction. o This represents the value after the final filling.

[0020] Furthermore, in step 3, based on the spatial heterogeneity of the nearshore area, multiple sub-regions are constructed, including nearshore, central, and far-from-shore areas, and each sub-region is trained and predicted separately.

[0021] Furthermore, in step 4, the methods for creating time-based and space-based input data are as follows:

[0022] For each sub-partition, obtain the chlorophyll a, radial wind, and zonal wind variation maps for T consecutive days. Construct a data sequence of three feature inputs using a sliding window mechanism, with a size of T*3*m*n, where m represents the number of grid points in the longitude direction and n represents the number of grid points in the latitude direction. First, perform a dimension transformation on this input sequence to construct a two-dimensional time input sequence.

[0023] Secondly, the input sequence of size T*3*m*n and the original region input sequence are used as spatial input data to construct a three-dimensional spatial input sequence.

[0024] Furthermore, the temporal feature extraction module includes an LSTM network and a multi-head self-attention mechanism. First, a multi-layer LSTM is used to mine the temporal scale dependency features of the input sequence in the sub-region. Then, the multi-head self-attention mechanism is used to automatically learn the contribution of the latent variable outputs of the LSTM at multiple time steps to the prediction. Specifically, the latent variable outputs of the LSTM at multiple time steps are multiplied by the query, key, and value parameter weight matrices of the multi-head self-attention mechanism, and then subjected to subsequent Softmax and linear transformation to obtain an output matrix with the same dimension as the input of the multi-head self-attention mechanism but containing the contribution values.

[0025] Furthermore, the spatial feature extraction module includes a local spatial feature extractor and a global spatial feature extractor. The local spatial feature extractor uses multi-layer convolution and activation function operations to enable the model to learn the chlorophyll a concentration evolution law of each sub-region. The global spatial feature extractor is completed by downsampling and convolution operations to obtain a feature map of the same size and dimension as the sub-region. Finally, the chlorophyll a at the current time, the local spatial features, and the global spatial features are fused into the last convolutional layer to complete the extraction of spatial features.

[0026] Furthermore, the spatiotemporal feature fusion prediction module performs fusion and prediction on the feature maps from the spatial feature extraction module and the temporal feature extraction module, as follows:

[0027] First, the input sequence is reconstructed by concatenating temporal and spatial evolution feature maps. Then, a convolutional layer is used to extract the spatiotemporal fusion feature relationships from the reconstructed input sequence. Finally, a linear layer and an activation layer are used to filter out negative values. Finally, the prediction results of multiple sub-regions are concatenated to predict the chlorophyll a concentration of the entire region.

[0028] Furthermore, the loss function adopts a weighted average root mean square loss function based on physical prior knowledge, as follows:

[0029] 1) Introduce the weighted root mean square loss function BMSE.

[0030]

[0031] The weighting range is set according to the characteristics of the chlorophyll data: the weights change according to the chlorophyll a concentration, and the unit of y is mg / m³. 3 :

[0032]

[0033] Where L represents the predicted duration, l represents a specific time point, (i,j) represents a specific station within the region, and ω l,i,jLet represent the weight of station (i,j) at the prediction time l, m represent the number of grid points in the longitude direction, n represent the number of grid points in the latitude direction, and y represent the weight of station (i,j). l,i,j This represents the predicted value of chlorophyll a concentration. This represents the true value of chlorophyll a concentration;

[0034] 2) Construct an input data region mask M0 to filter out land locations: If there is a chlorophyll a concentration value at the location, set the value of M0 at the given location to 1; if the chlorophyll a concentration is not available, such as if the location is on land, set the value to 0; when calculating the final loss, by multiplying the predicted value by M0, the training results of only the sea location points can be considered for optimization.

[0035] 3) The root mean square loss function for weighted equilibrium based on prior physical knowledge is:

[0036]

[0037] Furthermore, this invention also provides a nearshore chlorophyll a concentration prediction system based on spatiotemporal feature fusion, realizing the nearshore chlorophyll a concentration prediction method based on spatiotemporal feature fusion as described above. The system includes a raw data acquisition module, a raw data preprocessing module, a spatiotemporal input preprocessing module, a temporal feature extraction module based on an attention mechanism, a spatial feature extraction module based on local and global features, a spatiotemporal feature fusion prediction module, and an output module.

[0038] The raw data acquisition module is used to acquire raw chlorophyll a remote sensing data, meteorological remote sensing data, and buoy chlorophyll a monitoring data.

[0039] The raw data preprocessing module is used to complete the raw data to obtain chlorophyll a continuous spatiotemporal sequence data.

[0040] The spatiotemporal input preprocessing module is used to divide the sea area to be predicted into several sub-regions, and for each sub-region, to generate time-based input data and space-based input data.

[0041] The time feature extraction module is used to obtain the time evolution feature map of chlorophyll a.

[0042] The spatial feature extraction module is used to obtain the spatial evolution feature map of chlorophyll a.

[0043] The spatiotemporal feature fusion prediction module is used to perform post-fusion of the temporal evolution feature map and the spatial evolution feature map of each sub-partition to obtain the chlorophyll a prediction map of the next preceding time.

[0044] The output module merges the prediction results of all sub-regions to obtain a long-term continuous chlorophyll a prediction map of the entire nearshore area and outputs it.

[0045] Compared with the prior art, the advantages of this invention are:

[0046] (1) Given the high rate of missing data in nearshore areas, remote sensing data of chlorophyll a and related variables, as well as buoy chlorophyll a monitoring data, were used simultaneously. Based on the spatiotemporal heterogeneity of the region, a multi-step prediction method for regional non-source chlorophyll a concentration based on multivariate remote sensing data was designed. Through data preprocessing and spatiotemporal heterogeneous region division, the quality of remote sensing data was improved, and an input with spatial characteristics was constructed.

[0047] (2) A spatiotemporal attention feature fusion prediction network was constructed, which can learn the evolutionary features of multiple spatiotemporal interactions and retain the differences between the time domain and the spatial domain. This achieves the complementarity of spatiotemporal correlation features from multiple perspectives, fully explores the spatiotemporal correlation features, and improves the prediction of the evolutionary law of time-dependent features based on LSTM network. First, LSTM based on attention mechanism is introduced as a time feature extraction module to establish the difference in the dependence of output at multiple time points on different time latent variable features, so as to improve the utilization of time latent variable features, that is, to improve the time prediction mining ability of LSTM. Second, a spatial feature extraction module based on local and global features is introduced to fully explore the regional spatial dynamic correlation evolution law ignored by LSTM. Finally, the time evolution feature map and the spatial evolution feature map are fed into the spatiotemporal feature fusion prediction module for spatiotemporal correlation feature prediction modeling. In this way, it can learn the evolutionary features of multiple spatiotemporal interactions and retain the differences between the time domain and the spatial domain, and maximize the exploration of the spatiotemporal evolution law of chlorophyll a concentration.

[0048] (3) To address the imbalanced distribution of chlorophyll a concentration data, a weighted average root mean square loss function based on prior physical knowledge was designed. First, an improved average root mean square loss function using a weighted balancing mechanism was adopted, and weight ranges were set according to the characteristics of the chlorophyll data itself. This solved the problem of insufficient learning for sparse high-concentration sample values ​​and minimized the prediction error of high-concentration sample values. Second, a land mask design was introduced to filter out land locations and automatically encode marine physical knowledge into the modeling process, thereby improving the robustness of the model. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1This is a flowchart of the method of the present invention.

[0051] Figure 2 This is a structural diagram of the prediction system of the present invention.

[0052] Figure 3 This is a diagram of the time feature extraction module of the present invention.

[0053] Figure 4 This is a diagram of the spatial feature extraction module of the present invention. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0055] Example 1

[0056] like Figure 1 As shown, this embodiment provides a method for predicting nearshore chlorophyll a concentration based on spatiotemporal feature fusion, including the following steps:

[0057] Step 1: Obtain raw chlorophyll a remote sensing data, meteorological remote sensing data, and buoy chlorophyll a monitoring data.

[0058] Step 2: Data preprocessing to complete the chlorophyll a continuous spatiotemporal sequence data.

[0059] The specific data preprocessing method is as follows: First, the corresponding latitude and longitude grid points are filled using effective buoy chlorophyll a monitoring data; then, the regional spatiotemporal missing values ​​are filled using the empirical orthogonal function interpolation method (DINEOF); ​​finally, the data with filled values ​​less than 0 are filled using four-way geographic embedding average imputation, as shown in the following formula:

[0060]

[0061] In the formula, q is 4, indicating that four-way geographic embedding is selected, i indicates which specific embedding direction, i indicates the chlorophyll a concentration value of the grid point in i direction, and i indicates the final value after filling.

[0062] Step 3, Spatial Heterogeneity Partitioning: Based on spatial heterogeneity, the sea area to be predicted is divided into several sub-partitions according to the chlorophyll a continuous spatiotemporal sequence data obtained in Step 2.

[0063] Specifically, due to significant differences in spatial relationships across different ocean regions, the point values ​​in open ocean subregions are similar with relatively small fluctuations. This shows a strong positive correlation between nearby and distant points. Conversely, in nearshore regions, the values ​​differ, with significant seasonal variations. These different relationships across subregions indicate the existence of spatial heterogeneity. Using the entire original region as input for prediction, insufficient sample size prevents a full exploration of the spatial heterogeneity within subregions.

[0064] Therefore, this invention, based on the spatial heterogeneity of nearshore areas, constructs multiple sub-regions: nearshore, central sea area, and far offshore, and trains predictions for each sub-region separately. This approach avoids the catastrophic forgetting that can occur with simultaneous training of all sub-regions, thus improving prediction accuracy.

[0065] Step 4: Construct a spatiotemporal attention feature fusion prediction network. For each sub-region after step 3, generate time-based input data and space-based input data, and use the spatiotemporal attention feature fusion prediction network to obtain the temporal and spatial evolution features of chlorophyll a in each sub-region.

[0066] This embodiment Figure 2 The three sub-partitions shown The time- and space-based input data for the three partitions are TI1, SI1, TI2, SI2, and TI3, SI3.

[0067] Combined Figure 2 As shown, the spatiotemporal attention feature fusion prediction network includes a temporal feature extraction module based on an attention mechanism, a spatial feature extraction module based on local and global features, and a spatiotemporal feature fusion prediction module. The temporal feature extraction module obtains the temporal evolution feature map of chlorophyll a, and the spatial feature extraction module obtains the spatial evolution feature map of chlorophyll a.

[0068] Specifically, the methods for creating time-based and space-based input data in step 4 are as follows:

[0069] For each sub-region, chlorophyll a, radial wind, and zonal wind variation maps for consecutive days are obtained. A sliding window mechanism is used to construct a data sequence of three feature inputs, with a size of T*3*m*n, where m represents the number of grid points in the longitude direction and n represents the number of grid points in the latitude direction. In the prediction, the chlorophyll a concentration of T days is used to predict the chlorophyll a concentration of the next L days. First, this input sequence is dimensionally transformed to construct a two-dimensional time input sequence; this is then fed into the time feature extraction module of the spatiotemporal attention feature fusion prediction network for time feature extraction. Second, the sub-region sequence of size T*3*m*n and the original region input sequence are used as spatial input data to construct a three-dimensional spatial input sequence. This is then fed into the spatial feature extraction module of the spatiotemporal attention feature fusion prediction network for spatial feature extraction.

[0070] The time feature extraction module (TAFE Module), such as Figure 3As shown, the algorithm includes an LSTM network and a multi-head attention mechanism. First, a two-layer LSTM is used to mine the temporal scale dependencies of the input sequence in the sub-region. Then, the multi-head attention mechanism automatically learns the contribution of the latent variable outputs at multiple time steps of the LSTM to the prediction. In other words, the weights of different latent variable outputs are redistributed, fully considering the temporal interdependencies between different latent variables and treating them differently based on their importance. This ensures that the network focuses its attention on the most important time steps in terms of temporal features. Specifically, the input to the attention mechanism (i.e., the latent variable outputs of the LSTM at multiple time steps) is multiplied by the query, key, and value parameter weight matrices of the multi-head attention mechanism, and then subjected to softmax and linear transformations to obtain an output matrix with the same dimensions as the input to the multi-head attention mechanism but containing the contribution values.

[0071] The spatial feature extraction module (SAFE Module), such as Figure 4 As shown, the model includes a local spatial feature extractor and a global spatial feature extractor. The local spatial feature extractor, through multi-layer convolution and activation function operations, enables the model to better learn the chlorophyll a concentration evolution pattern of each sub-region. However, due to the fluidity of the water body, the neighborhood association between the local sub-region and the global spatial region is affected. Therefore, a global spatial feature extractor is added to address the influence of spatial neighborhood on the sub-region. The global spatial feature extractor is completed by downsampling (pooling) and convolution operations (CNN) to obtain a feature map of the same size and dimension as the sub-region. The chlorophyll a at the current time step is more relevant to the future chlorophyll a at the next time step, meaning it has the greatest impact on the next time step. Therefore, the chlorophyll a at the current time step, the local spatial features, and the global spatial features are finally fused into the last convolutional layer to complete the extraction of spatial features.

[0072] Step 5: Use the Spatiotemporal Feature Fusion Prediction Module (STFP Module) to fuse the temporal evolution feature map and spatial evolution feature map of each sub-partition to obtain the chlorophyll a prediction map for the next preceding time.

[0073] The spatiotemporal feature fusion prediction module (STFP Module) performs fusion and prediction on the feature maps from the spatial feature extraction module and the temporal feature extraction module. The specific method is as follows:

[0074] First, the input sequence is reconstructed by concatenating temporal and spatial evolution feature maps. Then, a convolutional layer is used to extract the spatiotemporal fusion feature relationships from the reconstructed input sequence. Finally, a linear layer and an activation layer are used to filter out negative values. Finally, the prediction results of multiple sub-regions are concatenated to predict the chlorophyll a concentration of the entire region.

[0075] Step 6: Iterate through the regional predictions, merge all sub-region prediction results, and obtain a long-term continuous chlorophyll a prediction map of the entire nearshore region.

[0076] Furthermore, based on the aforementioned concentration prediction method, this invention also proposes a weighted balanced root mean square loss function based on prior physical knowledge. On one hand, by introducing a land mask to filter out land locations, ocean physical knowledge is automatically encoded into the modeling process, thereby improving model robustness. On the other hand, an improved root mean square loss function using a weighted balancing mechanism is employed, and weight ranges are set according to the characteristics of the chlorophyll data itself, solving the problem of insufficient learning for sparse high-concentration sample values ​​and minimizing the prediction error of high-concentration sample values. Specifically:

[0077] 1) Introduce the weighted root mean square loss function BMSE.

[0078]

[0079] The weights change based on the chlorophyll a concentration, and the unit of y is mg / m³. 3 :

[0080]

[0081] Where L represents the predicted duration, l represents a specific time point, (i,j) represents a specific station within the region, and ω l,i,j y represents the weight of station (i,j) at the prediction time l; m represents the number of grid points in the longitude direction, and n represents the number of grid points in the latitude direction; l,i,j This represents the predicted value of chlorophyll a concentration. This represents the true value of chlorophyll a concentration.

[0082] 2) Construct an input data region mask M0 to filter out land locations: If there is a chlorophyll a concentration value at the location, set the value of M0 at the given location to 1; if the chlorophyll a concentration is not available, such as if the location is on land, set the value to 0; when calculating the final loss, by multiplying the predicted value by M0, the training results of only the sea location points can be considered for optimization.

[0083] 3) The root mean square loss function for weighted equilibrium based on prior physical knowledge is:

[0084]

[0085] Example 2

[0086] This embodiment provides a nearshore chlorophyll a concentration prediction system based on spatiotemporal feature fusion, including a raw data acquisition module, a raw data preprocessing module, a spatiotemporal input preprocessing module, a temporal feature extraction module based on an attention mechanism, a spatial feature extraction module based on local and global features, a spatiotemporal feature fusion prediction module, and an output module.

[0087] The raw data acquisition module is used to acquire raw chlorophyll a remote sensing data, meteorological remote sensing data, and buoy chlorophyll a monitoring data.

[0088] The raw data preprocessing module is used to complete the raw data to obtain chlorophyll a continuous spatiotemporal sequence data.

[0089] The Spatiotemporal Input Preprocessing Module (STIP Module) divides the sea area to be predicted into several sub-regions based on the characteristics of nearshore, central and deep sea areas. For each sub-region, it generates time-based and space-based input data and performs separate training and prediction for each sub-region.

[0090] The time feature extraction module is used to obtain the time evolution feature map of chlorophyll a.

[0091] The spatial feature extraction module is used to obtain the spatial evolution feature map of chlorophyll a.

[0092] The spatiotemporal feature fusion prediction module is used to perform post-fusion of the temporal evolution feature map and the spatial evolution feature map of each sub-partition to obtain the chlorophyll a prediction map of the next preceding time.

[0093] The output module merges the prediction results of all sub-regions to obtain a long-term continuous chlorophyll a prediction map of the entire nearshore area and outputs it.

[0094] The functions and implementation methods of each module of this system are the same as those described in Part 1 of Embodiment 1, and will not be repeated here.

[0095] In summary, this invention, considering the high missing rate of nearshore data, utilizes remote sensing data of chlorophyll a and related variables, as well as buoy chlorophyll a monitoring data. Based on the spatiotemporal heterogeneity of the region, it designs a multi-step prediction method for regional non-source chlorophyll a concentration based on multivariate remote sensing data. This method fully explores the spatiotemporal correlation features and improves the prediction method based solely on LSTM networks to extract the evolution of time-dependent features. Specifically, firstly, an attention-based LSTM is introduced as a temporal feature extraction module to establish the differences in the dependence of outputs at multiple time points on different temporal latent variable features, thereby improving the utilization of temporal latent variable features and enhancing the temporal prediction mining capability of LSTM. Secondly, a spatial feature extraction module based on local and global features is introduced to fully explore the regional spatial dynamic correlation evolution patterns ignored by LSTM. Finally, the temporal evolution feature map and the spatial evolution feature map are fed into the spatiotemporal feature fusion prediction module to perform spatiotemporal correlation feature prediction modeling. This can not only learn the evolution features of multiple spatiotemporal interactions, but also retain the differences between the time domain and the spatial domain, and maximize the exploration of the spatiotemporal evolution law of chlorophyll a concentration.

[0096] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should be protected by the present invention.

Claims

1. A method for predicting nearshore chlorophyll a concentration based on spatiotemporal feature fusion, characterized in that, Includes the following steps: Step 1: Acquire raw chlorophyll a remote sensing data, meteorological remote sensing data, and buoy chlorophyll a monitoring data; Step 2: Data preprocessing to complete the chlorophyll a continuous spatiotemporal sequence data; Step 3, Spatial Heterogeneity Partitioning: Based on spatial heterogeneity, the sea area to be predicted is divided into several sub-partitions according to the chlorophyll a continuous spatiotemporal sequence data obtained in Step 2. Specifically, based on the spatial heterogeneity of the nearshore area, multiple sub-regions are constructed, including nearshore, central, and far-shore areas, and each sub-region is trained and predicted separately. Step 4: Construct a spatiotemporal attention feature fusion prediction network. For each sub-partition after step 3, prepare time-based input data and space-based input data, and use the spatiotemporal attention feature fusion prediction network to obtain the temporal and spatial evolution features of chlorophyll a in each sub-partition. The spatiotemporal attention feature fusion prediction network includes a temporal feature extraction module based on an attention mechanism, a spatial feature extraction module based on local and global features, and a spatiotemporal feature fusion prediction module. The temporal feature extraction module obtains the temporal evolution feature map of chlorophyll a, and the spatial feature extraction module obtains the spatial evolution feature map of chlorophyll a. Step 5: Use the spatiotemporal feature fusion prediction module to perform post-fusion of the temporal evolution feature map and spatial evolution feature map of each sub-partition to obtain the chlorophyll a prediction map of the next preceding time. The spatiotemporal feature fusion and prediction module performs fusion and prediction on the feature maps from the spatial feature extraction module and the temporal feature extraction module. The specific method is as follows: First, the input sequence is reconstructed by concatenating temporal and spatial evolution feature maps. Then, a convolutional layer is used to extract the spatiotemporal fusion feature relationships from the reconstructed input sequence. Finally, a linear layer and an activation layer are used to filter out negative values. Finally, the prediction results of multiple sub-regions are concatenated to predict the chlorophyll a concentration of the entire region. Step 6: Iterate through the regional predictions, merge all sub-region prediction results, and obtain a long-term continuous chlorophyll a prediction map of the entire nearshore region.

2. The nearshore chlorophyll a concentration prediction method based on spatiotemporal feature fusion according to claim 1, characterized in that, In step 2, the specific data preprocessing method is as follows: First, the corresponding latitude and longitude grid point data are filled using effective buoy chlorophyll a monitoring data; then, the regional spatiotemporal missing values ​​are filled using the empirical orthogonal function interpolation method; finally, the data with filled values ​​less than 0 are filled using four-way geographic embedding average imputation, as shown in the following formula: (1)； In the formula, the value of q is 4, indicating that four-directional geographic embedding is selected, and i represents the specific embedding direction. This represents the chlorophyll a concentration value at the grid point in the i-direction. This represents the value after the final filling.

3. The nearshore chlorophyll a concentration prediction method based on spatiotemporal feature fusion according to claim 1, characterized in that, In step 4, the methods for creating time-based and space-based input data are as follows: For each sub-partition, obtain the chlorophyll a, radial wind, and zonal wind variation maps for T consecutive days. Construct a data sequence of three feature inputs using a sliding window mechanism, with a size of T*3*m*n, where m represents the number of grid points in the longitude direction and n represents the number of grid points in the latitude direction. First, perform a dimension transformation on this input sequence to construct a two-dimensional time input sequence. Secondly, the input sequence of size T*3*m*n and the original region input sequence are used as spatial input data to construct a three-dimensional spatial input sequence.

4. The nearshore chlorophyll a concentration prediction method based on spatiotemporal feature fusion according to claim 1, characterized in that, The temporal feature extraction module includes an LSTM network and a multi-head self-attention mechanism. First, a multi-layer LSTM is used to mine the temporal scale dependency features of the input sequence in the sub-region. Then, the multi-head self-attention mechanism is used to automatically learn the contribution of the latent variable outputs of the LSTM at multiple time steps to the prediction. Specifically, the latent variable outputs of the LSTM at multiple time steps are multiplied by the query, key, and value parameter weight matrices of the multi-head self-attention mechanism, and then subjected to subsequent Softmax and linear transformation to obtain an output matrix with the same dimension as the input of the multi-head self-attention mechanism but containing the contribution values.

5. The nearshore chlorophyll a concentration prediction method based on spatiotemporal feature fusion according to claim 1, characterized in that, The spatial feature extraction module includes a local spatial feature extractor and a global spatial feature extractor. The local spatial feature extractor uses multi-layer convolution and activation function operations to enable the model to learn the chlorophyll a concentration evolution pattern of each sub-region. The global spatial feature extractor is completed by downsampling and convolution operations to obtain a feature map with the same size and dimension as the sub-region. Finally, the chlorophyll a at the current time, the local spatial features, and the global spatial features are fused into the last convolutional layer to complete the extraction of spatial features.

6. The nearshore chlorophyll a concentration prediction method based on spatiotemporal feature fusion according to claim 1, characterized in that, The loss function is a weighted average root mean square loss function based on physical prior knowledge, as follows: 1) Introduce the weighted root mean square loss function BMSE. (2)； The weighting range is set according to the characteristics of the chlorophyll data: the weights change according to the chlorophyll a concentration, and the unit of y is mg / m³. 3 : (3)； Where L represents the predicted duration, l represents a specific time point, and (i,j) represents a specific station within the region. Let represent the weight of station (i,j) at the prediction time l, m represent the number of grid points in the longitude direction, and n represent the number of grid points in the latitude direction. This represents the predicted value of chlorophyll a concentration. This represents the true value of chlorophyll a concentration; 2) Construct an input data region mask M0 to filter out land locations: If there is a chlorophyll a concentration value at the location, set the value of M0 at the given location to 1; if the chlorophyll a concentration is not available, such as if the location is on land, set the value to 0; when calculating the final loss, by multiplying the predicted value by M0, the training results of only the sea location points can be considered for optimization. 3) The root mean square loss function for weighted equilibrium based on prior physical knowledge is: (4) 。 7. A nearshore chlorophyll a concentration prediction system based on spatiotemporal feature fusion, implementing the nearshore chlorophyll a concentration prediction method based on spatiotemporal feature fusion as described in any one of claims 1-6, characterized in that, It includes a raw data acquisition module, a raw data preprocessing module, a spatiotemporal input preprocessing module, a temporal feature extraction module based on an attention mechanism, a spatial feature extraction module based on local and global features, a spatiotemporal feature fusion and prediction module, and an output module. The raw data acquisition module is used to acquire raw chlorophyll a remote sensing data, meteorological remote sensing data, and buoy chlorophyll a monitoring data. The raw data preprocessing module is used to complete the raw data to obtain chlorophyll a continuous spatiotemporal sequence data. The spatiotemporal input preprocessing module is used to divide the sea area to be predicted into several sub-regions, and for each sub-region, to generate time-based input data and space-based input data. The time feature extraction module is used to obtain the time evolution feature map of chlorophyll a. The spatial feature extraction module is used to obtain the spatial evolution feature map of chlorophyll a. The spatiotemporal feature fusion prediction module is used to fuse the temporal evolution feature map and spatial evolution feature map of each sub-partition to obtain the chlorophyll a prediction map of the next preceding time. The output module merges the prediction results of all sub-regions to obtain a long-term continuous chlorophyll a prediction map of the entire nearshore area and outputs it.

Images