Method and System for Predicting Sea Surface Chlorophyll Concentration Based on Remote Sensing and Deep Learning

Abstract

The present application discloses a method and system for predicting sea surface chlorophyll concentration based on remote sensing and deep learning, which relates to the technical field of marine ecological environment. The method includes inputting the marine chlorophyll a data of a target area into a trained marine chlorophyll concentration prediction model to obtain the marine chlorophyll concentration information of the target area; wherein, the prediction model is an improved marine chlorophyll a concentration prediction model based on a convolutional long short-term memory network; a periodic feature extraction module is used to extract periodic features related to the same historical period based on the chlorophyll concentration data of the same historical period, and a noise removal and fusion module is used to eliminate the noise in the outputs of the recent spatio-temporal feature extraction module and the periodic feature extraction module, and fuse the output results to obtain the prediction result of the marine chlorophyll a concentration data. The present application can more accurately predict the sea surface chlorophyll concentration.

Description

Technical Field

[0001] This application relates to the technical field of marine ecological environment, and particularly to a method and system for predicting sea surface chlorophyll concentration based on remote sensing and deep learning. Background Art

[0002] The concentration of marine chlorophyll a (Chlorophyll-a or Chl-a) is a key indicator for measuring marine primary productivity, which is directly related to the vitality and health status of the marine ecosystem. Its concentration change has a profound impact on the stability of the marine food web and the abundance of fishery resources. Accurately predicting the spatio-temporal distribution of Chl-a is crucial for deeply understanding the dynamic changes of the marine ecosystem, evaluating the health status of the marine environment, timely warning of ecological disasters such as red tides and algal blooms, studying climate change, and guiding fishery resource management.

[0003] However, although existing prediction modeling techniques have made progress, there are still difficulties in capturing the short-term changes and periodic trends of Chl-a, especially in coastal areas with more noise. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for predicting sea surface chlorophyll concentration based on remote sensing and deep learning, which can accurately predict the sea surface chlorophyll concentration.

[0005] To achieve the above purpose, this application provides the following solutions:

[0006] In the first aspect, this application provides a method for predicting sea surface chlorophyll concentration based on remote sensing and deep learning, including:

[0007] Obtain the marine chlorophyll a data of the target area.

[0008] Input the marine chlorophyll a data into the trained marine chlorophyll concentration prediction model to obtain the marine chlorophyll concentration information of the target area.

[0009] Among them, the training process of the marine chlorophyll concentration prediction model is:

[0010] Obtain the historical marine chlorophyll a data of the set area.

[0011] Perform data preprocessing on the historical marine chlorophyll a data to obtain the processed historical marine chlorophyll a data.

[0012] Construct a marine chlorophyll a data set according to the processed historical marine chlorophyll a data.

[0013] Based on the marine chlorophyll a dataset, train the marine chlorophyll concentration prediction model to obtain a trained marine chlorophyll concentration prediction model; the marine chlorophyll concentration prediction model is composed of a recent spatio-temporal feature extraction module, a periodic feature extraction module, and a noise removal and fusion module; the periodic feature extraction module is used to extract periodic features related to the historical same period of the month to be predicted based on the chlorophyll a concentration data of the historical same period; the noise removal and fusion module is used to eliminate the noise in the outputs of the recent spatio-temporal feature extraction module and the periodic feature extraction module, and fuse the output results of the recent spatio-temporal feature extraction module and the periodic feature extraction module to obtain the prediction result of the marine chlorophyll a concentration data.

[0014] In a second aspect, the present application provides a sea surface chlorophyll concentration prediction system based on remote sensing and deep learning, including:

[0015] A data acquisition module, configured to acquire marine chlorophyll a data of a target area.

[0016] A prediction module, configured to input the marine chlorophyll a data into the trained marine chlorophyll concentration prediction model to obtain the marine chlorophyll concentration information of the target area.

[0017] A training module, including:

[0018] A historical data acquisition unit, configured to acquire historical marine chlorophyll a data of a set area.

[0019] A preprocessing unit, configured to perform data preprocessing on the historical marine chlorophyll a data to obtain processed historical marine chlorophyll a data.

[0020] A dataset establishment unit, configured to construct a marine chlorophyll a dataset according to the processed historical marine chlorophyll a data.

[0021] A training unit, configured to train the marine chlorophyll concentration prediction model based on the marine chlorophyll a dataset to obtain a trained marine chlorophyll concentration prediction model; the marine chlorophyll concentration prediction model is composed of a recent spatio-temporal feature extraction module, a periodic feature extraction module, and a noise removal and fusion module; the periodic feature extraction module is used to extract periodic features related to the historical same period of the month to be predicted based on the chlorophyll a concentration data of the historical same period; the noise removal and fusion module is used to eliminate the noise in the outputs of the recent spatio-temporal feature extraction module and the periodic feature extraction module, and fuse the output results of the recent spatio-temporal feature extraction module and the periodic feature extraction module to obtain the prediction result of the marine chlorophyll a concentration data.

[0022] According to the specific embodiments provided by the present application, the following technical effects are disclosed in the present application:

[0023] The present application provides a method and system for predicting sea surface chlorophyll concentration based on remote sensing and deep learning. The marine chlorophyll a data of the target area is input into the trained marine chlorophyll concentration prediction model to obtain the marine chlorophyll concentration information of the target area. The marine chlorophyll concentration prediction model is composed of a recent spatio-temporal feature extraction module, a periodic feature extraction module, and a noise removal and fusion module. The periodic feature extraction module is used to extract periodic features related to the historical same period of the month to be predicted based on the chlorophyll a concentration data of the historical same period. The noise removal and fusion module is used to eliminate the noise in the outputs of the recent spatio-temporal feature extraction module and the periodic feature extraction module, and fuse the output results of the recent spatio-temporal feature extraction module and the periodic feature extraction module to obtain the prediction result of the marine chlorophyll a concentration data. The prediction model in the present application integrates the recent spatio-temporal feature extraction, periodic feature extraction, and denoising and fusion modules to effectively capture the short-term spatio-temporal changes and periodic changes of the Chl-a concentration. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0025] Figure 1 It is a schematic flowchart of a method for predicting sea surface chlorophyll concentration based on remote sensing and deep learning provided by an embodiment of the present application.

[0026] Figure 2 It is a structure of a chlorophyll concentration prediction model provided by an embodiment of the present application.

[0027] Figure 3 It is a flowchart of Chl-a prediction provided by an embodiment of the present application.

[0028] Figure 4 It is a graph showing the change trend of monthly average Chl-a on the test set and comparing the prediction results of three models provided by an embodiment of the present application.

[0029] Figure 5 It is a graph showing the analysis results of the performance of the Chl-a prediction model between different calendar months provided by an embodiment of the present application.

[0030] Figure 6 It is a spatial distribution map of satellite observation values and prediction values of three models in four seasons provided by an embodiment of the present application.

[0031] Figure 7Spatial distribution map of the mean absolute error (MAE) between the Chl-a values predicted by three models and the Chl-a values observed by satellites in four seasons provided by an embodiment of the present application.

[0032] Figure 8 Spatial distribution map of the MAE between the Chl-a values predicted by three models and the Chl-a values observed by satellites in four seasons within the offshore range provided by an embodiment of the present application.

[0033] Figure 9 Spatial distribution map of the root mean square error (RMSE) of Chl-a predicted by three models provided by an embodiment of the present application.

[0034] Figure 10 Spatial distribution map of the MAE of the future 3-step Chl-a prediction by the ocean chlorophyll concentration prediction model (ChlaPM) provided by an embodiment of the present application.

[0035] Figure 11 Schematic diagram of the linear transformation denoising coefficient matrix of the ocean chlorophyll concentration prediction model provided by an embodiment of the present application. Detailed implementation manners

[0036] Next, the technical solutions in the embodiments of the present application will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0037] The concentration of oceanic chlorophyll a (Chlorophyll-a or Chl-a) is a key indicator for measuring the primary productivity of the ocean, directly related to the vitality and health of the marine ecosystem. Its concentration change has a profound impact on the stability of the marine food web and the abundance of fishery resources. Accurately predicting the spatio-temporal distribution of Chl-a is crucial for deeply understanding the dynamic changes of the marine ecosystem, evaluating the health status of the marine environment, timely warning of ecological disasters such as red tides and algal blooms, studying climate change, and guiding fishery resource management.

[0038] The prediction methods of chlorophyll concentration can be divided into two categories: physics-based methods and data-driven methods. Physics-based methods consider various physical, chemical, and biological processes and use physical and ecological dynamics equations to design models for prediction, which can more comprehensively describe the spatio-temporal dynamics of Chl-a. For example, Hamilton et al. developed a coupled hydrodynamic and ecological water quality model, DYRESM-CAEDYM, to simulate the effects of multiple ecological and chemical variables on Chl-a and other water quality indicators in lakes and reservoirs. Los et al. established a coupled physical-chemical-ecological BLOOM / GEM model and verified its ability to predict ecological variables such as nutrient cycling, dissolved oxygen, and primary production in the Dutch North Sea through three-dimensional simulations. However, physical models are highly complex, and the performance of physical parameterization schemes may vary under different environmental conditions, affecting the applicability of the models. At the same time, it is difficult to obtain data on future boundary conditions such as hydrology, meteorology, and nutrient concentrations in the marine environment, further affecting the application effect of physical models in Chl-a prediction.

[0039] Data-driven methods infer the future distribution of Chl-a by learning the change characteristics of historical data, including traditional statistical methods, popular machine learning methods, and artificial intelligence methods. This method requires less prior knowledge and is not as complex as physics-based methods, and has been widely used in water quality prediction tasks. For example, Tian et al. optimized a traditional artificial neural network-based dynamic prediction model of chlorophyll a, significantly reducing the cost of in-situ monitoring of the water environment and improving the accuracy of algal bloom prediction. Yu et al. proposed an LSTM model combining wavelet threshold denoising and mean fusion to simulate and predict the spatio-temporal changes in the Chl-a concentration in Dianchi Lake, and the results showed a significant reduction in errors and strong model generalization ability. However, most traditional data-driven methods independently process the observed values at each location in a region without considering the influence of its surrounding locations, making it difficult to apply to the prediction of large-scale sea areas.

[0040] With the development of satellite remote sensing technology, the method of remotely sensing and inverting chlorophyll a concentration has become increasingly mature, enabling the acquisition of a large amount of spatio-temporal observation data, creating new opportunities for predicting and estimating the sea surface Chl-a concentration. Network models based on convolutional structures can mine local spatial correlation information and demonstrate good performance in the remote sensing data analysis of Chl-a. In addition, spatio-temporal prediction models that couple convolutional and recurrent structures can further mine spatio-temporal dynamic characteristics and are widely used in the spatio-temporal prediction of ocean remote sensing chlorophyll concentration. For example, Liu et al. used a Convolutional Neural Network-Long Short-Term Memory Network (CNN-LSTM) to effectively predict the Chl-a concentration over a period of 3 years, accurately capturing the spatio-temporal change trend of the Chl-a concentration. Yao et al. compared the prediction effects of four models, namely SA-Convlstm, CNN-LSTM, E3D-LSTM, and SA-Convlstm, on chlorophyll concentration, demonstrating the superior performance of these models in the prediction of Chl-a concentration. However, when convolutional operations extract spatial information, they cannot filter out the noise signals in non-ocean regions (such as land and islands), resulting in an increase in prediction errors around the noise regions. Although Ye et al. solved this problem through a spatio-temporal prediction model based on graph convolution, an additional adjacency matrix needs to be introduced to represent the spatial correlation of observation points, increasing the computational complexity and limiting its application in large-scale sea areas.

[0041] The purpose of this application is to provide a method and system for predicting sea surface chlorophyll concentration based on remote sensing and deep learning, which can accurately predict the sea surface chlorophyll concentration.

[0042] To make the above objects, features, and advantages of this application more obvious and understandable, the following further details this application in conjunction with the accompanying drawings and specific embodiments.

[0043] Embodiment 1

[0044] As Figure 1 shown, this embodiment provides a method for predicting sea surface chlorophyll concentration based on remote sensing and deep learning, including:

[0045] Step 101: Obtain the ocean chlorophyll a data of the target area.

[0046] Step 102: Input the ocean chlorophyll a data into the trained ocean chlorophyll concentration prediction model to obtain the ocean chlorophyll concentration information of the target area.

[0047] Among them, before performing Step 102, it also includes: training the ocean chlorophyll concentration prediction model, and the specific training process is as follows:

[0048] Step 201: Obtain the historical ocean chlorophyll a data of the set area.

[0049] Step 202: Perform data preprocessing on the historical ocean chlorophyll a data to obtain the processed historical ocean chlorophyll a data.

[0050] Step 203: Construct an ocean chlorophyll a dataset based on the processed historical ocean chlorophyll a data.

[0051] Step 204: Train an ocean chlorophyll concentration prediction model based on the ocean chlorophyll a dataset to obtain a trained ocean chlorophyll concentration prediction model; the ocean chlorophyll concentration prediction model is composed of a recent spatio-temporal feature extraction module, a periodic feature extraction module, and a noise removal and fusion module; the periodic feature extraction module is used to extract periodic features related to the historical same period of the month to be predicted based on the chlorophyll a concentration data of the historical same period; the noise removal and fusion module is used to eliminate the noise in the outputs of the recent spatio-temporal feature extraction module and the periodic feature extraction module, and fuse the output results of the recent spatio-temporal feature extraction module and the periodic feature extraction module to obtain the prediction result of the ocean chlorophyll a concentration data.

[0052] Among them, when performing step 201, specifically, it can be as follows:

[0053] The study area is located in some offshore regions. This area is affected by the East Asian monsoon, surface runoff, and the Kuroshio Current. The chlorophyll a concentration varies greatly from the coast to the open sea and has obvious seasonal variation characteristics. The coastline of the eastern sea area is long, and the water temperature and salinity in this area change significantly. Seasonal upwelling and downwelling provide rich nutrients for the marine ecosystem. In recent years, with climate change and the rapid development of coastal cities, the frequency of red tide occurrence has increased significantly, which not only pollutes the marine environment in this area but also causes huge economic losses.

[0054] The historical ocean chlorophyll a data is the historical monthly average surface ocean chlorophyll a concentration data of a set area. In this embodiment, the monthly average surface Chl-a concentration data provided by the European Space Agency (ESA) is used. This product integrates satellite data such as SeaWiFS, MODIS-Aqua, MODIS-Terra, MERIS, VIIRS-SNPP, OLCI-S3A&S3B, etc., and the spatial resolution is 4 km. The spatial range of the data used in this embodiment is 26N° - 35N°, 120E° - 126E°, and the time range is from January 1998 to December 2023.

[0055] Among them, when performing steps 202 - 203, specifically, it can be as follows:

[0056] As Figure 3 shown, perform study area cropping, missing value filling, and masking preprocessing on the historical ocean chlorophyll a data to obtain the processed historical ocean chlorophyll a data.

[0057] Specifically, the data preprocessing includes three steps: cropping and masking, missing value filling, and dataset construction.

[0058] The first step is cropping and land-sea masking. The original Chl-a concentration data is cropped according to the research area range to retain the data within the research scope. After cropping, the data is masked to remove the Chl-a concentration data in land rivers and lakes, and only the Chl-a concentration data in the ocean is retained.

[0059] The second step is missing value filling. Due to factors such as weather conditions and the pollution of high solar glint, there are some missing values in the satellite remote sensing sea surface chlorophyll products. The Data Interpolation Empirical Orthogonal Function (DINEOF) method can reconstruct large-area missing data with high accuracy and has been widely used for filling missing values in ocean remote sensing data. Therefore, in this embodiment, the Data Interpolation Empirical Orthogonal Function (DINEOF) method is selected to reconstruct the missing sea surface Chl-a concentration data.

[0060] The third step is dataset construction. The dataset is divided into a model training set, a validation set, and a prediction set according to the time period. In the time dimension, the monthly average Chl-a data from January 1998 to December 2023 includes 312 data. Among them, 276 data from January 1998 to December 2020 are used for model training and validation, and 36 data from January 2021 to December 2023 are used as the prediction set.

[0061] Among them, before performing step 204, it also includes constructing an ocean chlorophyll concentration prediction model, which can be specifically as follows:

[0062] The ocean chlorophyll concentration prediction model is an improved ocean chlorophyll a (Chl-a) concentration prediction model based on ConvLSTM, named ChlaPM. This model reduces the impact of land and island areas on the prediction accuracy by integrating a noise region filtering module. At the same time, it enhances its ability to capture the long-term dynamic changes of Chl-a through a periodic feature extraction module.

[0063] Among them, the Convolutional Long Short-Term Memory Network (ConvLSTM) is a deep learning model that combines the advantages of the Convolutional Neural Network (CNN) and the Long Short-Term Memory Network (LSTM). It can not only process sequential data but also effectively capture complex patterns in spatio-temporal data. The core idea of ConvLSTM is to introduce a convolutional structure based on LSTM, enabling the network to consider spatial correlations in the input-to-state and state-to-state transitions. This structure is particularly suitable for processing time series data with spatial correlations, such as videos and meteorological radar images.

[0064] The convolutional long short-term memory network consists of three parts: the forget gate, the input gate, and the output gate, and the formulas are as follows:

[0065] 。

[0066] 。

[0067] 。

[0068] 。

[0069] 。

[0070] Among them, 、 、 、 represent the forget gate, the input gate, the memory cell, and the output gate respectively, and represent the input and hidden state at the current time step respectively, and represent the memory cell and hidden state at the previous time step respectively. and represent the weight matrix and the bias vector respectively. and represent the sigmoid activation function and the hyperbolic tangent activation function respectively, represents convolution, represents the Hadamard product.

[0071] In this embodiment, an ocean chlorophyll concentration prediction model (ChlaPM) is established based on the ConvLSTM neural network. As Figure 2 shown, the model includes three modules, namely the recent spatio-temporal feature extraction module (RSTFE), the periodic feature extraction module (PFE), and the noise removal and fusion module (DNF). The recent spatio-temporal feature extraction module consists of four convolutional long short-term memory network layers and one convolutional layer.

[0072] The recent chlorophyll concentration will affect the future chlorophyll concentration. Therefore, in this embodiment, the RSTFE module is designed to utilize past historical data to mine future complex spatio-temporal correlations. The RSTFE stacks four convolutional long short-term memory network layers and one convolutional layer.

[0073] In view of the periodic variation characteristics of ocean chlorophyll concentration, a PFE module is integrated in parallel in the model, which is specifically responsible for extracting periodic features related to the same historical period of the month to be predicted. The structure of the PFE module is similar to that of the RSTFE module, but its input data is the chlorophyll concentration data of the same historical period to capture the seasonal variation pattern.

[0074] Considering that the convolution operation may be interfered by noise regions such as land and islands when extracting spatial features, resulting in an increase in prediction error, this embodiment also introduces a DNF module. The purpose of this module is to eliminate the noise in the outputs of the RSTFE and PFE and fuse the outputs of the two parts to generate the final prediction result. Specifically, the periodic feature extraction module makes up for the deficiency that the basic model (RSTFE) simply relies on the recent spatio-temporal trend for prediction by capturing the long-term variation pattern of chlorophyll-a concentration (such as annual cycle and seasonal features). The experimental results show that the RSTFE+PFE model can more accurately predict the inflection point of the change in Chl-a concentration, thus reducing the overestimation or underestimation phenomenon. Especially at the time nodes with drastic changes (such as season alternation or the peak period of chlorophyll concentration), this module shows excellent adaptability and effectively improves the prediction accuracy.

[0075] Specifically, for the outputs of the RSTFE and PFE and use linear transformation to remove the noise to obtain and , and then and are fused to output the final prediction result . The formula is as follows:

[0076] .

[0077] Among them, is the prediction result, is the output of the recent spatio-temporal feature extraction module obtained by using linear transformation to remove the noise; is the output of the periodic feature extraction module obtained by using linear transformation to remove the noise, ; ; and are the linear transformation weight matrices of the corresponding modules respectively. Each element corresponds to the weight of the spatial position and the feature channel. These matrices are optimized through network training so that the feature intensity can be adjusted according to the statistical distribution of historical data, thereby reducing the noise interference. Through the above transformation, the model can effectively reduce the impact of noise on the prediction result and improve the prediction accuracy.

[0078] Among them, in some embodiments, based on the marine chlorophyll a dataset, a marine chlorophyll concentration prediction model is trained to obtain a trained marine chlorophyll concentration prediction model, specifically including:

[0079] According to the absolute error index, root mean square error index, and mean absolute percentage error index, the performance of the marine chlorophyll concentration prediction model is evaluated. The formulas are as follows:

[0080] The formula expression of the absolute error index is: 。

[0081] The formula expression of the root mean square error is: 。

[0082] The formula expression of the mean absolute percentage error index is: 。

[0083] Among them, represents the model prediction value, represents the satellite observation value, and n is the number of data points used for verification.

[0084] Among them, the marine chlorophyll a dataset includes a training set, a test set, and a validation set divided in a set ratio.

[0085] In some embodiments, in order to discuss the influence of different modules on chlorophyll concentration prediction, this embodiment designs three prediction models, namely the RSTFE model that only uses recent historical data as input, the RSTFE+PFE model that combines periodic data input, and the ChlaPM model that adds a noise removal module. The length of the recent historical data is set to the previous six months, the periodic data input is the historical data of the same month in the previous three years, and the chlorophyll concentration in the next month is predicted.

[0086] As Figure 4 shown, the curve of the spatial average value of the Chl-a prediction results of RSTFE, RSTFE+PFE, and ChlaPM changing with time. Compared with the RSTFE+PFE model, the prediction results of ChlaPM are closer to the changes in the real Chl-a. After introducing the noise removal module, the ability of ChlaPM to learn the future Chl-a change trend by extracting the spatio-temporal information of the data is enhanced. In this embodiment, the average chlorophyll concentration reached the annual cycle peak in January 2022, and the chlorophyll concentration reached the annual cycle bottom in June 2023, while the RSTFE predicted the cycle peak and bottom to appear in February 22 and July 23 respectively, showing a "lag" phenomenon. The reason for this difference may be that RSTFE only learns the recent change trend, does not introduce periodic data, and does not have information on the inflection point of the change.

[0087] Figure 5The average Chl-a prediction values and the resulting average RMSE of RSTFE, RSTFE+PFE, and ChlaPM in different calendar months are shown. Among them, Figure 5 in (a) is the comparison of Chl-a prediction values of ChlaPM, RSTFE, and RSTFE+PFE for different calendar months; Figure 5 in (b) is the comparison of RMSE of Chl-a prediction of ChlaPM, RSTFE, and RSTFE+PFE for different calendar months. It can be seen that ChlaPM achieved the lowest RMSE among the three models in each month, especially in summer and autumn. In addition, compared with RSTFE, the average RMSE of each month of the RSTFE+PFE model and ChlaPM that introduced periodic data decreased significantly. This indicates that periodic characteristics are crucial in ocean Chl-a prediction.

[0088] In addition, the spatial distribution of chlorophyll a concentration in the study area has significant seasonal variations. Figure 6 The seasonal distributions of satellite observed values and predicted values of the three models of chlorophyll concentration are given. Among them, Figure 6 (a)-(d) in it represent the distributions of satellite observed chlorophyll in spring, summer, autumn, and winter. (e)-(h) represent the distributions of predicted values of the RSTFE model, (i)-(l) represent the distributions of predicted values of the RSTFE+PFE model, and (m)-(p) represent the distributions of predicted values of the ChlaPM model. In the nearshore high-value area, especially near the offshore, the predicted values of the RSTFE model are lower than the observed values, while the predicted values of the RSTFE+PFE model are higher than the observed values. The RSTFE+PFE model and ChlaPM can simulate seasonal variations more accurately, and the seasonal distribution of the predicted values of ChlaPM has higher accuracy in the nearshore and offshore. Figure 7 and Figure 8 show that the MAE between the predicted values of the RSTFE model and the observed values is large. The inaccuracy is mainly concentrated in the areas with higher chlorophyll concentration. Among them, Figure 7 in (a)-(d) represent the distributions of MAE between the predicted values of the RSTFE model and satellite observed values in spring, summer, autumn, and winter, (e)-(h) represent the distributions of MAE between the predicted values of the RSTFE+PFE model and satellite observed values, and (i)-(l) represent the distributions of MAE between the predicted values of the ChlaPM model and satellite observed values. Figure 8 in (a)-(d) represent the distributions of MAE between the predicted values of the RSTFE model and satellite observed values in spring, summer, autumn, and winter, (e)-(h) represent the distributions of MAE between the predicted values of the RSTFE+PFE model and satellite observed values, and (i)-(l) represent the distributions of MAE between the predicted values of the ChlaPM model and satellite observed values.

[0089] The spatial distribution of the prediction error also shows seasonal differences. In autumn and winter, the prediction error is relatively large in the areas near the coast, while in spring and summer, the prediction error in this area decreases. Compared with the RSTFE model, the prediction error of the RSTFE+PFE model has decreased, and there is no obvious seasonal difference in the spatial distribution of the prediction error. The MAE between the predicted value and the observed value of the ChlaPM model is small. Compared with RSTFE and RSTFE+PFE, the prediction performance in the coastal area has been significantly improved.

[0090] As Figure 9 shown, Figure 9 in (a) represents the distribution of RMSE between the predicted value of the RSTFE model and the satellite observed value, (b) represents the distribution of RMSE between the predicted value of the RSTFE+PFE model and the satellite observed value, and (c) represents the distribution of RMSE between the predicted value of the ChlaPM model and the satellite observed value. It can be seen that the RMSE of the prediction results of the RSTFE model is larger in the high and medium chlorophyll concentration regions. The RMSE distribution of the RSTFE+PFE model is highly consistent with that of the RSTFE model, but the absolute value of the error has decreased. Compared with the RSTFE model, the RMSE of the prediction results of the ChlaPM model has decreased significantly, and the RMSE of the prediction results is only relatively high in a small part of the nearshore area.

[0091] In some embodiments, this embodiment also focuses on the performance of the model at different prediction steps. By analyzing the prediction performance for the next 1 month, the next 3 months, and the next 6 months, the effectiveness of the model in short-term and long-term predictions can be evaluated. Such analysis not only helps this embodiment understand the stability and accuracy of the model on different time scales, but also is of great significance for optimizing the model structure and improving the prediction accuracy.

[0092] Table 1 shows the RMSE, MAE, and MAPE results obtained by RSTFE, RSTFE+PFE, and ChlaPM in Chl-a prediction. From the results in Table 1, it can be seen that when predicting the next 1 month, the model performance is the best, and then the prediction performance of the neural network decreases as the prediction step increases. The performance indicators of the RSTFE model at all prediction steps are inferior to those of the RSTFE+PFE model and the ChlaPM model. The ChlaPM model also outperforms the RSTFE+PFE model in terms of performance. Its performance gradually decreases as the prediction step increases, but the decline is relatively small. Therefore, for long-term predictions, the performance of the ChlaPM model is more stable.

[0093] Table 1 Chl-a prediction results

[0094]

[0095] Figure 10The spatial distribution of the 3-step prediction MAE for ChlaPM is shown. From left to right, they represent the MAE distributions of the prediction results for the next T+1 month, the next T+2 months, and the next T+3 months respectively. The average MAEs are 0.264, 0.275, and 0.276 respectively. The Chl-a values at the deep-water nodes in the offshore area far from the land are small and have little fluctuation. Therefore, as the prediction step size increases, the deviation of the prediction results does not change much.

[0096] The ChlaPM model further adopts a linear transformation technique to eliminate the prediction bias caused by the masked area and the marine spatial heterogeneity. Figure 11 Shows the linear transformation denoising coefficient matrices output by the RSTFE module and the PFE module in the ChlaPM model and . First of all, the coefficient matrix successfully captures the spatial distribution characteristics of the land-sea boundary area. In the land and island areas, the coefficient matrix values are low, indicating that the model effectively suppresses or ignores the output interference in these areas, thus avoiding the misleading of the zero-value mask on the prediction. Further, in the marine area at the land edge (such as the nearshore transition area), the denoising coefficient matrix has a high value, indicating that the model can moderately amplify the output in these key areas, thus correcting the underestimation bias caused by the land area. In addition, in the areas with high Chl-a concentration, the denoising coefficient matrix has a significantly high value, reflecting that the model makes fine adjustments in the areas with significant seasonal fluctuations and captures the complex characteristics of the concentration changes in these areas. In summary, the ChlaPM model effectively solves the influence of the masked area and the marine spatial heterogeneity on the prediction of Chl-a concentration by finely adjusting the linear transformation denoising coefficient matrix, thus achieving a significant performance improvement in simulating the seasonal changes of Chl-a. This discovery not only provides a new perspective for the prediction of ocean chlorophyll concentration, but also provides an important reference for the application of deep learning models in the field of ocean remote sensing.

[0097] Example 2

[0098] This embodiment provides a sea surface chlorophyll concentration prediction system based on remote sensing and deep learning, including: [[ID=Z0]]

[0099] A data acquisition module for acquiring ocean chlorophyll a data of a target area.

[0100] A prediction module for inputting the ocean chlorophyll a data into a trained ocean chlorophyll concentration prediction model to obtain the ocean chlorophyll concentration information of the target area.

[0101] A training module, including:

[0102] A historical data acquisition unit for acquiring historical ocean chlorophyll a data of a set area.

[0103] A preprocessing unit for preprocessing the historical ocean chlorophyll a data to obtain processed historical ocean chlorophyll a data.

[0104] A dataset establishment unit for constructing an ocean chlorophyll a dataset based on the processed historical ocean chlorophyll a data.

[0105] A training unit for training an ocean chlorophyll concentration prediction model based on the ocean chlorophyll a dataset to obtain a trained ocean chlorophyll concentration prediction model; the ocean chlorophyll concentration prediction model is composed of a recent spatio-temporal feature extraction module, a periodic feature extraction module, and a noise removal and fusion module; the periodic feature extraction module is used to extract periodic features related to the historical same period of the month to be predicted based on the chlorophyll a concentration data of the historical same period; the noise removal and fusion module is used to eliminate the noise in the outputs of the recent spatio-temporal feature extraction module and the periodic feature extraction module, and fuse the output results of the recent spatio-temporal feature extraction module and the periodic feature extraction module to obtain a prediction result of the ocean chlorophyll a concentration data.

[0106] Among them, the preprocessing unit specifically includes:

[0107] A preprocessing subunit for performing study area cropping, missing value filling, and masking preprocessing on the historical ocean chlorophyll a data to obtain processed historical ocean chlorophyll a data; the historical ocean chlorophyll a data is the historical monthly average surface ocean chlorophyll a concentration data of a set area; the ocean chlorophyll a dataset includes a training set, a test set, and a validation set divided in a set ratio; among them, the ocean chlorophyll concentration prediction model is an improved ocean chlorophyll a concentration prediction model based on a convolutional long short-term memory network; the recent spatio-temporal feature extraction module is composed of four convolutional long short-term memory network layers and one convolutional layer.

[0108] Among them, the training unit specifically includes:

[0109] An evaluation subunit for evaluating the performance of the ocean chlorophyll concentration prediction model according to the absolute error index, the root mean square error index, and the mean absolute percentage error index.

[0110] In summary, the present application has the following technical effects:

[0111] This application aims to improve the prediction ability of ocean chlorophyll a concentration by introducing the ChlaPM model based on the Convolutional Long Short-Term Memory (ConvLSTM) network. This model integrates modules for recent spatiotemporal feature extraction (RSTFE), periodic feature extraction (PFE), and denoise fusion (DNF) to effectively capture the short-term spatiotemporal variations and periodic changes in Chl-a concentration. In this application, monthly average Chl-a remote sensing data from 1998 to 2023 were used to evaluate the performance of the ChlaPM model in single-step and multi-step predictions. The results show that compared with the RSTFE model, the ChlaPM model significantly reduced the root mean square error (RMSE) by 53.84%, 53.58%, and 49.70% respectively when predicting the Chl-a concentration for the next 1 month, 3 months, and 6 months. These improvements highlight the model's ability to effectively capture the short-term variations and periodic trends in Chl-a concentration.

[0112] The large-scale and high-precision prediction of ocean chlorophyll a (Chl-a) concentration proposed in this application has important scientific and practical significance for preventing large-scale algal bloom events and alleviating marine environmental pollution. This application comprehensively considers the short-term spatiotemporal variability and periodic changes in Chl-a concentration and evaluates the potential impact of regions such as land and islands on prediction accuracy. To this end, three deep learning models were trained in this application: the recent spatiotemporal feature extraction model (RSTFE), the enhanced periodic feature extraction model (RSTFE+PFE), and the model integrating noise removal and feature fusion (ChlaPM), with the aim of improving the prediction accuracy of Chl-a concentration.

[0113] Due to its advantage of being able to capture both temporal dynamics and spatial features simultaneously, the ConvLSTM model has been widely applied to tasks such as sea surface temperature prediction and precipitation forecasting. Such models perform particularly well in dealing with data with complex spatiotemporal correlations in earth sciences.

[0114] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0115] In this application, specific examples are used to elaborate on the principles and implementation manners of the application. The description of the above embodiments is only used to help understand the method and its core idea of the application. At the same time, for those of ordinary skill in the art, according to the idea of the application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the application.

Claims

1. A method for predicting sea surface chlorophyll concentration based on remote sensing and deep learning, characterized in that, Including: Obtain the ocean chlorophyll a data of the target area; Input the ocean chlorophyll a data into the trained ocean chlorophyll concentration prediction model to obtain the ocean chlorophyll concentration information of the target area; Among them, the training process of the ocean chlorophyll concentration prediction model is: Obtain the historical ocean chlorophyll a data of the set area; Perform data preprocessing on the historical ocean chlorophyll a data to obtain the processed historical ocean chlorophyll a data; Construct an ocean chlorophyll a data set based on the processed historical ocean chlorophyll a data; Based on the ocean chlorophyll a data set, train the ocean chlorophyll concentration prediction model to obtain a trained ocean chlorophyll concentration prediction model; the ocean chlorophyll concentration prediction model is composed of a recent spatio-temporal feature extraction module, a periodic feature extraction module, and a noise removal and fusion module; the periodic feature extraction module is used to extract periodic features related to the historical same period of the month to be predicted based on the chlorophyll a concentration data of the historical same period; the noise removal and fusion module is used to eliminate the noise in the outputs of the recent spatio-temporal feature extraction module and the periodic feature extraction module, and fuse the output results of the recent spatio-temporal feature extraction module and the periodic feature extraction module to obtain the prediction result of the ocean chlorophyll a concentration data; The recent spatio-temporal feature extraction module is composed of four convolutional long short-term memory network layers and one convolutional layer; the input of the recent spatio-temporal feature extraction module is recent historical data; The periodic feature extraction module is composed of four convolutional long short-term memory network layers and one convolutional layer; the input of the periodic feature extraction module is historical same period data; The time length of the recent historical data is set to the previous six months, and the historical same period data input is the historical same month data of the previous three years; Fuse the output results of the recent spatio-temporal feature extraction module and the periodic feature extraction module to obtain the prediction result of the ocean chlorophyll a concentration data, specifically including: ； Among them, is the prediction result, is the output of the recent spatio-temporal feature extraction module obtained by removing noise using linear transformation; is the output of the periodic feature extraction module obtained by removing noise using linear transformation.

2. The method for predicting sea surface chlorophyll concentration based on remote sensing and deep learning according to claim 1, wherein, The ocean chlorophyll concentration prediction model is an improved ocean chlorophyll a concentration prediction model based on a convolutional long short-term memory network; the convolutional long short-term memory network is composed of a forgetting gate, an input gate, and an output gate.

3. The method for predicting sea surface chlorophyll concentration based on remote sensing and deep learning according to claim 2, wherein The formula expression of the convolutional long short-term memory network is: ； ； ； ； ； Among them, , , , represent the forget gate, input gate, memory cell, and output gate respectively, and represent the input and hidden state at the current time step respectively, and represent the memory cell and hidden state at the previous time step respectively; and represent the weight matrix and bias vector respectively; and represent the sigmoid activation function and hyperbolic tangent activation function respectively, represents convolution, represents the Hadamard product.

4. The method for predicting sea surface chlorophyll concentration based on remote sensing and deep learning according to claim 1, wherein Perform data preprocessing on the historical ocean chlorophyll a data to obtain the processed historical ocean chlorophyll a data, specifically including: Perform study area cropping, missing value filling, and masking preprocessing on the historical ocean chlorophyll a data to obtain the processed historical ocean chlorophyll a data; the historical ocean chlorophyll a data is the historical monthly average surface ocean chlorophyll a concentration data of the set area.

5. A method for predicting sea surface chlorophyll concentration based on remote sensing and deep learning according to claim 1, characterized in that, Based on the ocean chlorophyll a data set, train the ocean chlorophyll concentration prediction model to obtain a trained ocean chlorophyll concentration prediction model, specifically including: Evaluate the performance of the ocean chlorophyll concentration prediction model according to the absolute error index, root mean square error index, and mean absolute percentage error index; the ocean chlorophyll a data set includes a training set, a test set, and a validation set divided in a set ratio.

6. The method for predicting sea surface chlorophyll concentration based on remote sensing and deep learning according to claim 5, wherein, The formula expression of the absolute error index is as follows: ; The formula expression of the root mean square error is as follows: ; The formula expression of the mean absolute percentage error index is as follows: ; Among them, represents the model predicted value, represents the satellite observed value, and n is the number of data points used for verification.

7. A sea surface chlorophyll concentration prediction system based on remote sensing and deep learning, characterized in that, Including: A data acquisition module for obtaining the ocean chlorophyll a data of the target area; A prediction module for inputting ocean chlorophyll a data into a trained ocean chlorophyll concentration prediction model to obtain the ocean chlorophyll concentration information of the target area; A training module, including: A historical data acquisition unit for acquiring historical ocean chlorophyll a data of a set area; A preprocessing unit for preprocessing the historical ocean chlorophyll a data to obtain the processed historical ocean chlorophyll a data; A dataset establishment unit for constructing an ocean chlorophyll a dataset based on the processed historical ocean chlorophyll a data; A training unit for training an ocean chlorophyll concentration prediction model based on the ocean chlorophyll a dataset to obtain a trained ocean chlorophyll concentration prediction model; the ocean chlorophyll concentration prediction model is composed of a recent spatio-temporal feature extraction module, a periodic feature extraction module, and a noise removal fusion module; the periodic feature extraction module is used to extract periodic features related to the historical same period of the month to be predicted based on the chlorophyll a concentration data of the historical same period; the noise removal fusion module is used to eliminate the noise in the outputs of the recent spatio-temporal feature extraction module and the periodic feature extraction module, and fuse the output results of the recent spatio-temporal feature extraction module and the periodic feature extraction module to obtain the prediction result of the ocean chlorophyll a concentration data; The recent spatio-temporal feature extraction module is composed of four convolutional long short-term memory network layers and one convolutional layer; the input of the recent spatio-temporal feature extraction module is recent historical data; The periodic feature extraction module is composed of four convolutional long short-term memory network layers and one convolutional layer; the input of the periodic feature extraction module is historical same-period data; The time length of the recent historical data is set to the previous six months, and the historical same-period data input is the historical same-month data of the previous three years; Fusing the output results of the recent spatio-temporal feature extraction module and the periodic feature extraction module to obtain the prediction result of the ocean chlorophyll a concentration data, specifically including: ； Among them, is the prediction result, is the output of the recent spatio-temporal feature extraction module obtained by removing noise using linear transformation; is the output of the periodic feature extraction module obtained by removing noise using linear transformation.

8. The sea surface chlorophyll concentration prediction system based on remote sensing and deep learning according to claim 7, characterized in that, The preprocessing unit specifically includes: A preprocessing subunit for performing study area cropping, missing value filling, and masking preprocessing on the historical ocean chlorophyll a data to obtain the processed historical ocean chlorophyll a data; the historical ocean chlorophyll a data is the historical monthly average surface ocean chlorophyll a concentration data of a set area; the ocean chlorophyll a dataset includes a training set, a test set, and a validation set divided in a set ratio; among them, the ocean chlorophyll concentration prediction model is an improved ocean chlorophyll a concentration prediction model based on a convolutional long short-term memory network; the recent spatio-temporal feature extraction module is composed of four convolutional long short-term memory network layers and one convolutional layer.

9. The sea surface chlorophyll concentration prediction system based on remote sensing and deep learning according to claim 7, wherein The training unit specifically includes: An evaluation subunit for evaluating the performance of the ocean chlorophyll concentration prediction model according to the absolute error index, the root mean square error index, and the mean absolute percentage error index.

Images