Chlorophyll a concentration prediction method, device, equipment and storage medium thereof

Abstract

The application discloses a chlorophyll-a concentration prediction method, device and equipment and a storage medium thereof, and belongs to the technical field of marine environment research. The chlorophyll-a concentration prediction method comprises the following steps: obtaining a remote sensing data set of a target sea area, wherein the remote sensing data set comprises first remote sensing data with chlorophyll-a concentration and second remote sensing data without chlorophyll-a concentration; filling the chlorophyll-a concentration of the second remote sensing data by using the first remote sensing data to obtain a continuous spatio-temporal sequence data set; training a preset to-be-trained model by using the spatio-temporal sequence data set to obtain a chlorophyll-a prediction model; and inputting to-be-predicted sea area data into the chlorophyll-a prediction model for prediction to obtain a chlorophyll-a concentration prediction result. The application solves or at least partially solves the technical problem of low prediction accuracy of chlorophyll-a concentration in the sea.

Description

Technical Field

[0001] This application relates to the field of marine environmental research technology, and in particular to a method, apparatus, device and storage medium for predicting chlorophyll a concentration. Background Technology

[0002] Algal blooms are a global phenomenon, typically characterized by a rapid increase or accumulation of algae in freshwater or marine systems, and are a significant indicator of nearshore marine environmental degradation. Algal blooms lead to increased algal pigments in the water, causing turbidity and resulting in destructive ecological, social, and economic impacts. They not only affect water transparency and dissolved oxygen levels but also secrete toxins, leading to mass mortality of fish and other aquatic organisms or ecological collapse. Therefore, harmful algal blooms pose a significant threat to aquaculture. Furthermore, through the bioaccumulation by filter-feeding marine organisms (such as oysters), harmful algal blooms increase algal toxin levels in seafood, affecting human health. In addition, the algae that bloom in an outbreak usually die shortly after emergence, sinking to the seabed where they are further decomposed by bacteria, reducing dissolved oxygen levels and causing water acidification and hypoxia, which can severely destroy the entire nearshore ecosystem. Algal blooms are closely related to chlorophyll a concentration in the water; therefore, chlorophyll a concentration is a commonly used indicator for assessing eutrophication and algal blooms.

[0003] Currently, the main methods for obtaining chlorophyll a concentration include field sampling surveys and remote sensing satellite observations. Among these, field sampling yields the most accurate chlorophyll a data, but it requires both field sampling and laboratory analysis, resulting in high costs, low efficiency, and limitations imposed by sea conditions and other factors, thus producing very limited data. Satellite remote sensing technology can perform large-scale spatial chlorophyll a retrieval, but its data quality is easily affected by atmospheric cloud cover and other parameters, limiting chlorophyll a coverage and failing to provide reliable forecasts of chlorophyll information.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, device, and storage medium for predicting chlorophyll a concentration, which aims to solve or at least partially solve the technical problem of low accuracy in predicting chlorophyll a concentration in the ocean.

[0006] To achieve the above objectives, this application provides a method for predicting chlorophyll a concentration, the method comprising:

[0007] Obtain a remote sensing dataset of the target sea area, wherein the remote sensing dataset includes: first remote sensing data with chlorophyll a concentration and second remote sensing data without chlorophyll a concentration;

[0008] The chlorophyll a concentration in the second remote sensing data is filled in using the first remote sensing data to obtain a continuous spatiotemporal sequence dataset;

[0009] The spatiotemporal sequence dataset is used to train the preset model to be trained, and a chlorophyll a prediction model is obtained.

[0010] The data of the sea area to be predicted is input into the chlorophyll a prediction model to obtain the chlorophyll a concentration prediction result.

[0011] Optionally, the step of using the first remote sensing data to fill in the chlorophyll a concentration in the second remote sensing data to obtain a continuous spatiotemporal series dataset includes:

[0012] Based on the preset mapping relationship between pixel position and physical position, the first physical position corresponding to each pixel point contained in each of the second remote sensing data is matched and obtained;

[0013] Based on the rate of change of chlorophyll a concentration at each of the first physical locations over time, a second physical location and a third physical location are determined from the first physical locations, wherein the rate of change of chlorophyll a concentration at the second physical location is higher than the rate of change of chlorophyll a concentration at the third physical location.

[0014] Based on the LSTM model, the missing data at the second physical location is predicted to obtain the first predicted data;

[0015] Based on the XGboost model, the missing data at the third physical location is predicted to obtain the second predicted data;

[0016] Based on the first and second prediction data, missing data is filled into the second remote sensing data to obtain the spatiotemporal sequence dataset.

[0017] Optionally, the step of determining the second and third physical locations from the first physical locations based on the rate of change of chlorophyll a concentration over time at each of the first physical locations includes:

[0018] The importance of each of the first physical locations is determined based on the chlorophyll a concentration change rate, wherein the chlorophyll a concentration change rate is directly proportional to the importance.

[0019] Based on the importance of each of the first physical locations, a predetermined proportion of the first physical locations are selected as the second physical locations, and the remaining first physical locations are selected as the third physical locations.

[0020] Optionally, the remote sensing dataset further includes: environmental influencing factors, which include at least one of: ocean surface temperature, aerosol index, and particulate backscattering coefficient.

[0021] Optionally, the step of training a preset model to obtain a chlorophyll a prediction model using the spatiotemporal sequence dataset includes:

[0022] The spatiotemporal sequence corresponding to each physical location in the spatiotemporal sequence dataset is divided into multiple subsequences with a preset first duration.

[0023] The data of the first second preset duration in each of the sub-sequences is used as the model input, and the data of the remaining duration is used as the label to generate training data;

[0024] The training data is used to train the model to be trained, and the chlorophyll a prediction model is obtained.

[0025] Optionally, the step of training a preset model to obtain a chlorophyll a prediction model using the spatiotemporal sequence dataset includes:

[0026] Identify the fourth physical location contained in the spatiotemporal sequence dataset;

[0027] Determine the prediction complexity of each of the fourth physical locations, and set the training weights of each of the fourth physical locations based on the prediction complexity;

[0028] Based on the training weights of each of the fourth physical locations, and using the spatiotemporal sequence dataset to train the model to be trained, the chlorophyll a prediction model is obtained.

[0029] Optionally, before the step of inputting the data of the sea area to be predicted into the chlorophyll a prediction model for prediction, the method further includes:

[0030] A composite loss function is constructed based on image reconstruction loss and matrix loss;

[0031] The chlorophyll a prediction model is optimized based on the composite loss function.

[0032] Optionally, the chlorophyll a prediction model includes:

[0033] A deep feature encoder is used to extract high-level features from input data and map the high-level features to a latent space;

[0034] PhyCell cells are used to simulate dynamic changes in physical processes and predict chlorophyll a concentration based on differential operator modeling.

[0035] ConvLSTM units are used to combine convolutional neural networks and long short-term memory networks and capture the spatial dependencies in the input data to generate the spatiotemporal information of the input data.

[0036] A state decoder is used to recombine the processed high-level features and backmap them back to the original data space through the network hierarchy to generate the chlorophyll a concentration prediction result.

[0037] This application also provides a chlorophyll a concentration prediction device, the device comprising:

[0038] The acquisition module is used to acquire a remote sensing dataset of the target sea area, wherein the remote sensing dataset includes: first remote sensing data with chlorophyll a concentration and second remote sensing data without chlorophyll a concentration;

[0039] The filling module is used to fill the chlorophyll a concentration in the second remote sensing data with the first remote sensing data to obtain a continuous spatiotemporal sequence dataset;

[0040] The training module is used to train the preset model to be trained using the spatiotemporal sequence dataset to obtain the chlorophyll a prediction model.

[0041] The prediction module is used to input the data of the sea area to be predicted into the chlorophyll a prediction model to make predictions and obtain the chlorophyll a concentration prediction results.

[0042] This application also provides an electronic device, the electronic device comprising: a memory, a processor, and a chlorophyll a concentration prediction program stored in the memory and executable on the processor, the chlorophyll a concentration prediction program being configured to implement the steps of the chlorophyll a concentration prediction method described above.

[0043] This application also provides a storage medium, which is a computer-readable storage medium, on which a chlorophyll a concentration prediction program is stored. The chlorophyll a concentration prediction program is executed by a processor to implement the steps of the above-described chlorophyll a concentration prediction method.

[0044] This application discloses a method for predicting chlorophyll a concentration. The method involves acquiring a remote sensing dataset of a target sea area, wherein the dataset includes: first remote sensing data with chlorophyll a concentration and second remote sensing data lacking chlorophyll a concentration; then, the first remote sensing data is used to fill in the chlorophyll a concentration in the second remote sensing data, resulting in a continuous spatiotemporal sequence dataset; then, the spatiotemporal sequence dataset is used to train a preset model to obtain a chlorophyll a prediction model; finally, the sea area data to be predicted is input into the chlorophyll a prediction model for prediction, yielding a chlorophyll a concentration prediction result. This application utilizes a remote sensing dataset containing chlorophyll a concentrations within the target sea area, providing strong support for constructing training data for a more comprehensive model. Addressing the data gaps caused by cloud cover in marine remote sensing data, this application fills in the missing chlorophyll a concentration data in the second remote sensing dataset, achieving data completion in both temporal and spatial dimensions, thus obtaining a continuous and comprehensive spatiotemporal sequence dataset. The chlorophyll a prediction model trained based on this spatiotemporal sequence dataset can accurately predict chlorophyll a concentration changes in various sea areas, effectively improving the prediction accuracy of chlorophyll a concentration in the ocean. Furthermore, the chlorophyll a prediction model can be directly applied to predict chlorophyll a concentrations in various geographical regions without requiring repeated training, thereby significantly improving the prediction efficiency of chlorophyll a concentrations in different regions. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the structure of an electronic device in the hardware operating environment involved in the embodiments of this application;

[0046] Figure 2 This is a flowchart illustrating the chlorophyll a concentration prediction method involved in the embodiments of this application;

[0047] Figure 3 This is a schematic diagram illustrating the chlorophyll a concentration prediction results involved in the embodiments of this application;

[0048] Figure 4 for Figure 2 A detailed flowchart of an embodiment of step S20;

[0049] Figure 5 This is a schematic diagram of the chlorophyll a prediction model involved in the embodiments of this application;

[0050] Figure 6 This is a schematic diagram of the frame structure of the chlorophyll a prediction device involved in the embodiments of this application.

[0051] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0052] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0053] Furthermore, the use of terms such as "first" and "second" in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the term "and / or" throughout the text includes three solutions; taking A and / or B as an example, it includes technical solution A, technical solution B, and a technical solution that simultaneously satisfies A and B. Furthermore, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of a person skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0054] Reference Figure 1 , Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of this application.

[0055] like Figure 1 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0056] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0057] like Figure 1As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and a chlorophyll a concentration prediction program.

[0058] exist Figure 1 In the illustrated electronic device, the network interface 1004 is mainly used for data communication with other devices; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the electronic device of this application can be disposed in the electronic device, and the electronic device calls the chlorophyll a concentration prediction program stored in the memory 1005 through the processor 1001 and performs the following operations:

[0059] Obtain a remote sensing dataset of the target sea area, wherein the remote sensing dataset includes: first remote sensing data with chlorophyll a concentration and second remote sensing data without chlorophyll a concentration;

[0060] The chlorophyll a concentration in the second remote sensing data is filled in using the first remote sensing data to obtain a continuous spatiotemporal sequence dataset;

[0061] The spatiotemporal sequence dataset is used to train the preset model to be trained, and a chlorophyll a prediction model is obtained.

[0062] The data of the sea area to be predicted is input into the chlorophyll a prediction model to obtain the chlorophyll a concentration prediction result.

[0063] Furthermore, the operation of using the first remote sensing data to fill in the chlorophyll a concentration in the second remote sensing data to obtain a continuous spatiotemporal series dataset includes:

[0064] Based on the preset mapping relationship between pixel position and physical position, the first physical position corresponding to each pixel point contained in each of the second remote sensing data is matched and obtained;

[0065] Based on the rate of change of chlorophyll a concentration at each of the first physical locations over time, a second physical location and a third physical location are determined from the first physical locations, wherein the rate of change of chlorophyll a concentration at the second physical location is higher than the rate of change of chlorophyll a concentration at the third physical location.

[0066] Based on the LSTM model, the missing data at the second physical location is predicted to obtain the first predicted data;

[0067] Based on the XGboost model, the missing data at the third physical location is predicted to obtain the second predicted data;

[0068] Based on the first and second prediction data, missing data is filled into the second remote sensing data to obtain the spatiotemporal sequence dataset.

[0069] Further, the operation of determining the second and third physical locations from the first physical locations based on the rate of change of chlorophyll a concentration over time at each of the first physical locations includes:

[0070] The importance of each of the first physical locations is determined based on the chlorophyll a concentration change rate, wherein the chlorophyll a concentration change rate is directly proportional to the importance.

[0071] Based on the importance of each of the first physical locations, a predetermined proportion of the first physical locations are selected as the second physical locations, and the remaining first physical locations are selected as the third physical locations.

[0072] Furthermore, the remote sensing dataset also includes environmental influencing factors, which include at least one of ocean surface temperature, aerosol index, and particulate backscattering coefficient.

[0073] Furthermore, the operation of training the preset model to obtain the chlorophyll a prediction model using the spatiotemporal sequence dataset includes:

[0074] The spatiotemporal sequence corresponding to each physical location in the spatiotemporal sequence dataset is divided into multiple subsequences with a preset first duration.

[0075] The data of the first second preset duration in each of the sub-sequences is used as the model input, and the data of the remaining duration is used as the label to generate training data;

[0076] The training data is used to train the model to be trained, and the chlorophyll a prediction model is obtained.

[0077] Furthermore, the processor 1001 can call the chlorophyll a concentration prediction program stored in the memory 1005 and also perform the following operations:

[0078] The operation of training a preset model to obtain a chlorophyll a prediction model using the spatiotemporal sequence dataset includes:

[0079] Identify the fourth physical location contained in the spatiotemporal sequence dataset;

[0080] Determine the prediction complexity of each of the fourth physical locations, and set the training weights of each of the fourth physical locations based on the prediction complexity;

[0081] Based on the training weights of each of the fourth physical locations, and using the spatiotemporal sequence dataset to train the model to be trained, the chlorophyll a prediction model is obtained.

[0082] Furthermore, the processor 1001 can call the chlorophyll a concentration prediction program stored in the memory 1005 and also perform the following operations:

[0083] Before the step of inputting the data of the sea area to be predicted into the chlorophyll a prediction model for prediction, the method further includes:

[0084] A composite loss function is constructed based on image reconstruction loss and matrix loss;

[0085] The chlorophyll a prediction model is optimized based on the composite loss function.

[0086] Furthermore, the chlorophyll a prediction model includes:

[0087] A deep feature encoder is used to extract high-level features from input data and map the high-level features to a latent space;

[0088] PhyCell cells are used to simulate dynamic changes in physical processes and predict chlorophyll a concentration based on differential operator modeling.

[0089] ConvLSTM units are used to combine convolutional neural networks and long short-term memory networks and capture the spatial dependencies in the input data to generate the spatiotemporal information of the input data.

[0090] A state decoder is used to recombine the processed high-level features and backmap them back to the original data space through the network hierarchy to generate the chlorophyll a concentration prediction result.

[0091] Based on the above structure, various embodiments of the chlorophyll a concentration prediction method are proposed.

[0092] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the chlorophyll a concentration prediction method of this application.

[0093] In this embodiment, the executing entity of the chlorophyll a concentration prediction method can be an electronic device, which can be a local device or a network device. No limitation is imposed in this embodiment. For ease of description, the executing entity is omitted from the following description of each embodiment. In this embodiment, the chlorophyll a concentration prediction method includes:

[0094] Step S10: Obtain a remote sensing dataset of the target sea area, wherein the remote sensing dataset includes: first remote sensing data with chlorophyll a concentration and second remote sensing data without chlorophyll a concentration.

[0095] In one feasible embodiment, in order to achieve accurate prediction of chlorophyll a concentration in the ocean, model training is required first; then, a remote sensing dataset of the target sea area is obtained, wherein the data in the remote sensing dataset includes: first remote sensing data with complete chlorophyll a concentration (Chl-a) and second remote sensing data missing chlorophyll a concentration.

[0096] The remote sensing dataset includes multiple remote sensing data points. The corresponding chlorophyll a concentration can be obtained from each remote sensing data point. Furthermore, the remote sensing dataset can also obtain environmental influencing factors, namely marine environmental factors that affect chlorophyll a concentration.

[0097] For example, the remote sensing dataset could be JAXA's Himawari (Sunflower Weather Satellite) data.

[0098] In one feasible embodiment, the environmental impact factors include at least one of ocean surface temperature, aerosol index, and particulate backscattering coefficient.

[0099] Changes in ocean surface temperature (SST) are crucial for phytoplankton growth, as phytoplankton are extremely sensitive to temperature. Moderate increases in water temperature can usually promote phytoplankton metabolism and growth, leading to an increase in chlorophyll a concentration; however, excessively high water temperatures may inhibit the growth of these organisms, thereby reducing chlorophyll a concentration.

[0100] The particulate backscattering coefficient (bbp) is an indicator of particulate matter (including phytoplankton and other suspended matter) in seawater. A high value usually indicates a high concentration of particulate matter and further suggests a high concentration of chlorophyll a.

[0101] An increase in the aerosol index (ARAE) may reduce the amount of solar radiation reaching the ocean surface, thereby affecting the photosynthesis of phytoplankton in the ocean, and may indirectly affect ocean surface temperature and phytoplankton growth by influencing cloud formation, thus affecting the concentration of chlorophyll a in the ocean.

[0102] For example, the Himawari dataset provides data slices at each time point that are pixel matrices with a spatial resolution of 0.05° and a shape of (2401×2401), covering the period from July 2015 to the present. It also provides multiple time resolution options, including every 10 minutes / hour / day / week / month. Remote sensing datasets for the target sea area (110.00°E-173.75°E, 3.75°S-60.00°N) are obtained from the Himawari dataset. These datasets include spatial data slices for chlorophyll a concentration (Chl-a), sea surface temperature (SST), aerosol index (ARAE), and particulate backscattering coefficient (bbp) at each time point. Each spatial data slice contains 64×64 pixels. The remote sensing dataset is defined as X∈R. T×C×H×W Where T is the time length, C represents different data features, namely Chl-a, SST, ARAE, and bbb, and (H×W) is the image size; the tensor representation at time t is defined as X. t ∈R C×H×W X t c，i，j Let be the value at the t-th time of pixel (c, i, j), where t = 1, 2, ..., T; c = 1, 2, 3, 4; i = 1, 2, ..., H; j = 1, 2, ..., W.

[0103] In this embodiment, chlorophyll a concentration is integrated with environmental factors that can influence it, providing strong support for constructing more comprehensive model training data. If only historical chlorophyll a concentrations are used as training data, it may be difficult to predict abrupt changes in chlorophyll a concentration evolution in the ocean. Furthermore, for cloud-covered areas where satellites cannot directly observe chlorophyll a concentration, the model will be unable to accurately predict the chlorophyll a concentration in those areas. This embodiment, by introducing directly observable factors such as ocean surface temperature, aerosol index, and particulate backscattering coefficient, can uncover the correlation between environmental factors and chlorophyll a, thereby accurately capturing potential trends in chlorophyll a concentration changes. This improves the accuracy of chlorophyll a prediction in complex areas, enabling precise prediction of chlorophyll a concentration in various sea areas during practical use.

[0104] Step S20: Use the first remote sensing data to fill in the chlorophyll a concentration of the second remote sensing data to obtain a continuous spatiotemporal sequence dataset;

[0105] In one feasible embodiment, since satellite observations may be affected by factors such as cloud cover, there may be a large amount of missing data (i.e., second remote sensing data) in the remote sensing dataset that cannot obtain chlorophyll a concentration; therefore, it is necessary to fill in the missing data in the second remote sensing dataset to obtain a continuous spatiotemporal sequence dataset.

[0106] Step S30: Use the spatiotemporal sequence dataset to train the preset model to be trained to obtain the chlorophyll a prediction model;

[0107] In one feasible embodiment, a pre-defined training model is trained using a complete, continuous spatiotemporal sequence dataset, so that the model learns about the changes in chlorophyll a concentration based on the training data, and finally obtains a trained chlorophyll a prediction model.

[0108] Step S40: Input the data of the sea area to be predicted into the chlorophyll a prediction model for prediction, and obtain the chlorophyll a concentration prediction result.

[0109] In one feasible embodiment, during actual prediction, the data of the sea area to be predicted is input into the trained chlorophyll a prediction model for prediction, and the chlorophyll a prediction model outputs the chlorophyll a concentration prediction result.

[0110] For example, the prediction results of chlorophyll a concentration are visualized, referring to... Figure 3 This paper presents a comparison between the predicted chlorophyll a concentration and the actual observed values ​​for the sea area to be predicted (110.00°E-173.75°E, 3.75°S-60.00°N). The horizontal axis represents longitude, and the vertical axis represents latitude. Figure 3 As can be seen, the error between the actual and predicted values ​​in this embodiment remains within a small range, approximately 5%. This low error rate clearly demonstrates that the chlorophyll a prediction model proposed in this application has high accuracy and reliability in capturing changes in chlorophyll a concentration. By accurately simulating and predicting chlorophyll a concentration, this application can provide strong support for marine ecological monitoring and management, showing great potential, especially in predicting and preventing algal blooms.

[0111] In this embodiment, the chlorophyll a concentration of the target sea area and environmental influencing factors are integrated, providing strong support for constructing more comprehensive model training data. Addressing the data gaps caused by cloud cover in marine remote sensing data, missing data is filled into the remote sensing dataset, achieving data perfection in both temporal and spatial dimensions, thus obtaining a continuous and comprehensive spatiotemporal sequence dataset. The chlorophyll a prediction model trained based on this spatiotemporal sequence dataset can accurately predict the changes in chlorophyll a concentration in various sea areas, effectively improving the prediction accuracy of chlorophyll a concentration in the ocean. Furthermore, the chlorophyll a prediction model can be directly applied to predict chlorophyll a concentration in various geographical regions without requiring repeated training, thereby greatly improving the prediction efficiency of chlorophyll a concentration in different regions.

[0112] Furthermore, based on the first embodiment described above, a second embodiment of the chlorophyll a concentration prediction method of this application is proposed. In this embodiment, reference is made to... Figure 4 Step S20, which involves using the first remote sensing data to fill in the chlorophyll a concentration in the second remote sensing data to obtain a continuous spatiotemporal sequence dataset, includes:

[0113] Step S21: Based on the preset mapping relationship between pixel position and physical position, the first physical position corresponding to each pixel point contained in each of the second remote sensing data is matched and obtained.

[0114] In one feasible embodiment, in order to accurately filter out spatial locations with significant physical meaning or environmental impact (i.e., physical locations), this application employs a spatial mapping method based on raw data, mapping each pixel point... Generate two-dimensional position point I m,n As input, Expanding yields the value V corresponding to the physical location. m As output, this reflects the spatial information (i.e., physical location) of each pixel, where the mapping relationship between pixel location and physical location is as follows:

[0115]

[0116]

[0117] Based on the preset mapping relationship between pixel positions in remote sensing data and actual physical positions, the first physical position corresponding to each pixel point contained in each second remote sensing data is obtained, wherein the pixel point is a pixel point lacking chlorophyll a concentration.

[0118] Step S22: Based on the chlorophyll a concentration change rate of each of the first physical locations over time, determine the second physical location and the third physical location from the first physical locations, wherein the chlorophyll a concentration change rate of the second physical location is higher than the chlorophyll a concentration change rate of the third physical location.

[0119] In one feasible embodiment, based on the rate of change of chlorophyll a concentration at each first physical location over time, the first physical locations are divided into second physical locations with a higher rate of change of chlorophyll a concentration and third physical locations with a lower rate of change of chlorophyll a concentration.

[0120] In one feasible embodiment, step S22, determining the second and third physical locations from the first physical locations based on the rate of change of chlorophyll a concentration over time at each of the first physical locations, includes:

[0121] Step S221: Determine the importance of each of the first physical locations based on the chlorophyll a concentration change rate, wherein the chlorophyll a concentration change rate is directly proportional to the importance.

[0122] In one feasible embodiment, the importance of each first physical location is determined based on the rate of change of chlorophyll a concentration, wherein the importance is used to measure the significance of the physical location during model training, and the rate of change of chlorophyll a concentration is directly proportional to the importance.

[0123] Optionally, the gradient boosting decision tree (GBDT) algorithm is used to calculate the importance of each first physical position, and the first physical positions are sorted based on this to filter out those positions with significant changes in chlorophyll a concentration over time (second physical positions) and those with less change (third physical positions).

[0124] For example, the importance of position j is measured by the average importance of position j in the gradient boosting decision tree:

[0125]

[0126] Where M is the number of gradient boosting decision trees. The importance of position j in a single tree is:

[0127]

[0128] Where L is the number of leaf nodes in the gradient boosting decision tree, L-1 is the number of non-leaf nodes in the gradient boosting decision tree, and v t These are features associated with nodes. It is the reduction in squared loss after node t splits.

[0129] Step S222: Based on the importance of each of the first physical locations, select a preset proportion of the first physical locations to serve as the second physical locations, and use the remaining first physical locations as the third physical locations.

[0130] In one feasible embodiment, the first physical locations are sorted based on their importance, and the first physical locations with a higher priority are selected as the second physical locations, while the remaining first physical locations are selected as the third physical locations.

[0131] In this embodiment, since the physical locations with larger chlorophyll a concentration change rates have richer temporal information and are more critical during model training, the importance of each first physical location is determined by the chlorophyll a concentration change rate. By classifying the physical locations, the second physical locations with significant chlorophyll a concentration changes over time and the third physical locations with less change are obtained, thereby enabling targeted data filling.

[0132] Step S23: Based on the LSTM model, predict the missing data of the second physical location to obtain the first predicted data;

[0133] In one feasible embodiment, the chlorophyll a concentration at the second physical location has a large rate of change and a high degree of change over time, which is a scenario rich in time-series information. In this case, an LSTM model is used to predict missing data to obtain the first predicted data. The LSTM model can learn the time dependence of the data and capture the time dependence of regions that show significant fluctuations or trends over time, thereby achieving accurate and coherent data filling.

[0134] Step S24: Based on the XGboost model, predict the missing data at the third physical location to obtain the second predicted data;

[0135] In one feasible embodiment, the chlorophyll a concentration at the third physical location has a low rate of change, i.e., the region with less change in the time dimension. This may be because the actual rate of change of chlorophyll a concentration in this region is low, or because it is difficult to obtain the chlorophyll a concentration at each moment due to cloud cover. Then, the XGBoost model is used to predict the missing data to obtain the second predicted data. The XGBoost model can accurately capture the spatial information of the region and effectively handle data with little change in the time dimension.

[0136] Step S25: Based on the first predicted data and the second predicted data, fill in the missing data in the second remote sensing data to obtain the spatiotemporal sequence dataset.

[0137] In one feasible embodiment, missing data is filled into the remote sensing dataset based on the first and second prediction data, i.e., the second remote sensing data is filled in to obtain a continuous spatiotemporal sequence dataset.

[0138] Alternatively, a spatiotemporal sequence dataset can be represented as:

[0139] features agg ={Chl-a, SST, ARAE, bbp}

[0140] Where Chl-a represents chlorophyll a concentration, SST represents ocean surface temperature, ARAE represents the aerosol index, and bbb represents the particulate backscattering coefficient, all with dimensions T×H×W; features are concatenated from the first dimension to obtain the fused feature matrix. agg The dimensions of the (i.e., spatiotemporal sequence dataset) are T×C×H×W, where C=4 represents the feature dimension.

[0141] In this embodiment, by combining the XGBoost model and the LSTM model, not only is the accuracy of spatial features improved, but the coherence of the temporal dimension is also enhanced, enabling accurate filling of missing values ​​of chlorophyll a concentration in the spatiotemporal dimensions, which is conducive to building a more comprehensive training dataset for the model.

[0142] Furthermore, based on the first and / or second embodiments described above, a third embodiment of the chlorophyll a concentration prediction method of this application is proposed. In this embodiment, step S30, which involves training a preset model to be trained using the spatiotemporal sequence dataset to obtain a chlorophyll a prediction model, includes:

[0143] Step S31: Divide the spatiotemporal sequence corresponding to each physical location in the spatiotemporal sequence dataset into multiple subsequences of a preset first duration;

[0144] In one feasible embodiment, in order to transform model training into supervised learning, it is necessary to divide the spatiotemporal sequence dataset into features and labels; then, the spatiotemporal sequence corresponding to each physical location in the spatiotemporal sequence dataset is divided into multiple subsequences of a preset first duration.

[0145] Optionally, to prevent the data scale from being too large and making the model difficult to train, the data in the spatiotemporal sequence dataset is normalized.

[0146] Step S32: Take the data of the first second preset duration in each of the sub-sequences as the model input, take the data of the remaining duration as the label, and generate training data;

[0147] Step S33: Use the training data to train the model to be trained to obtain the chlorophyll a prediction model.

[0148] In one feasible embodiment, data from the first second preset duration of each subsequence is used as model input, and data from the remaining duration is used as labels to generate training data. Data is classified by duration so that the model can learn the temporal variation pattern of chlorophyll a concentration based on the training data. The constructed training data is then used to train the model to be trained, resulting in the chlorophyll a prediction model.

[0149] For example, a 20-day sliding window is set, and one time step is slid along the time dimension each time, thus dividing the spatiotemporal sequence corresponding to each physical location in the spatiotemporal sequence dataset into multiple subsequences with a duration of 20 days. The first 10 days are used as model input, and the last 10 days are used as labels. The transformed data can be represented as X∈R N×L×H×W Where N is the number of samples, L = 10 represents the time length, and (H, W) represents the spatial grid size of each time slice, i.e., height and width.

[0150] In one feasible implementation, before step S40, which involves inputting the data of the sea area to be predicted into the chlorophyll a prediction model for prediction, the method further includes:

[0151] Step S41: Construct a composite loss function based on image reconstruction loss and matrix loss;

[0152] Step S42: Optimize the chlorophyll a prediction model based on the composite loss function.

[0153] In one feasible embodiment, a composite loss function is constructed based on image reconstruction loss and matrix loss to ensure that the differential operator of the PhyCell unit conforms to physical laws. The composite loss function is:

[0154]

[0155] For image reconstruction loss, This is the matrix loss. Furthermore, the expression for the image reconstruction loss is:

[0156]

[0157] Where D = {u} 1:N The constructed training set is also the input to the model. u(i) is the input sequence of the i-th training sample in the training set, and PhyDNet(u(i); Θ) represents the output of the PhyDNet model using parameter Θ, where ||·||2 is the L2 norm.

[0158] Furthermore, the expression for matrix loss is:

[0159]

[0160] Among them, w p,i,j These are the learning filter parameters of PhyCell, M(w) p,i,j ) is the matrix moment representing the differential order of the filter, Δ k,i,j It is the moment of the target matrix, ||·|| F It is the Frobenius norm. This constraint is applied to w. p,i,j This enables it to have the ability to estimate difference operators. The ability to assess the degree to which the target is satisfied is achieved by comparing the matrix of the target.

[0161] Optionally, the loss function, which is the optimization objective, is defined by the constraint: minimizing the loss function.

[0162]

[0163] The first loss term is the image reconstruction loss. The second loss is the physical constraint loss. The weights λ1 and λ2 of the two losses are... n The values ​​are fixed at λ1 = 1 and λ2 = 2.

[0164] Optionally, in order to optimize and update the model parameters, this application employs the gradient descent method, the mathematical expression of which is as follows:

[0165]

[0166] in, This refers to the set of training parameters for the k-th iteration, w p w represents the trainable parameters in a PhyCell unit. r represents the trainable parameters of the spatiotemporal prediction ConvLSTM unit, while w s These are the trainable parameters in the entire model's encoding and decoding. ξ is the learning rate of the neural network, which uses a decay strategy: it decreases as the number of training iterations increases during training. During training, this application uses the Adam optimizer with an initial learning rate lr of 0.001, and exponential decay rates β1 and β2 for the first and second moment estimates are 0.9 and 0.999, respectively. The learning rate adjustment strategy uses ReduceLROnPlateau, the loss change mode is minimization (min), the patience value is set to 2, and the learning rate reduction factor is set to 0.1.

[0167] In this embodiment, the model training was precisely tuned by combining image reconstruction loss and matrix moment loss. This approach, which combines physical laws with deep learning algorithms, not only enhances the model's understanding of complex physical processes but also significantly improves prediction accuracy. Furthermore, the physics-driven approach helps the model better understand and capture the complex environmental factors affecting chlorophyll a concentration, thus improving the prediction accuracy of chlorophyll a concentration in the ocean.

[0168] In one feasible implementation, step S30, training the preset model to be trained using the spatiotemporal sequence dataset to obtain the chlorophyll a prediction model, includes:

[0169] Step S34: Identify the fourth physical location contained in the spatiotemporal sequence dataset;

[0170] Step S35: Determine the prediction complexity of each of the fourth physical locations, and set the training weights of each of the fourth physical locations based on the prediction complexity.

[0171] Step S36: Based on the training weights of each of the fourth physical locations, and using the spatiotemporal sequence dataset, train the model to be trained to obtain the chlorophyll a prediction model.

[0172] In this embodiment, adversarial adaptive training of the model is introduced. For the prediction of chlorophyll a concentration at each location, an adaptive training weight is introduced. The core idea is to give more attention and weight to complex training points; unlike conventional self-attention mechanisms, which require a large number of parameters and training weights, this application proposes a sub-network-based attention mechanism, the optimization objective of which is:

[0173]

[0174]

[0175] The parameter φ controls the weights. Furthermore, this application introduces a subnet for the weight function, and scales the output of the neural network as follows to obtain the final weights:

[0176]

[0177] Where, N fThis represents the number of samples; it is clear from this weight that it is bounded to ensure training stability. Furthermore, the loss optimization objective is transformed into the above minmax form because this application aims to assign more weight to the difficult-to-learn chlorophyll a prediction region, which will also lead to an increase in the training loss value, thus resulting in an internal maximum constraint on φ. At the same time, solving this internal constraint is expensive, requiring many iterations. Therefore, this application employs a gradient ascent estimation method, namely:

[0178]

[0179] This method can approximate the optimization requirement for maximizing φ in minmax optimization, achieving the same training effect while reducing the training burden.

[0180] In this embodiment, an adaptive training mechanism prioritizes and allocates more computational resources to data points that significantly impact prediction performance. This strategy not only maintains the overall prediction accuracy of the model but also significantly enhances its ability to predict complex or challenging data points. This adaptability enables the model to exhibit higher accuracy and robustness when dealing with chlorophyll a concentration prediction in complex marine environments.

[0181] Furthermore, based on the first, second, and / or third embodiments described above, a fourth embodiment of the chlorophyll a concentration prediction method of this application is proposed. In this embodiment, the chlorophyll a prediction model is a PhyDNet model, which can utilize convolutional filters to approximate partial differential equations (PDEs) that conform to physical laws, and cleverly integrates the above equations into a deep learning model. It should be understood that this application fully utilizes the correlation between different features to incorporate various other variables (h) that affect chlorophyll a concentration. o This approach is comprehensively applied to the prediction of chlorophyll a. Furthermore, this application fully utilizes prior physical knowledge, coupling it to physical information, and successfully captures the physical dynamics (h) affecting chlorophyll a concentration and other variables. p h o ) and unknown dynamics (h r This significantly improves the model's generalization ability, enabling it to more effectively learn the dynamic characteristics of chlorophyll a concentration and achieve accurate predictions. The chlorophyll a prediction model includes:

[0182] A deep feature encoder is used to extract high-level features from input data and map the high-level features to a latent space, wherein the input data is the data input to the chlorophyll a prediction model;

[0183] PhyCell units are used to simulate physical dynamics in physical processes and to accurately predict chlorophyll a concentration based on differential operator modeling.

[0184] ConvLSTM units are used to combine convolutional neural networks and long short-term memory networks and capture the spatial dependencies in the input data to generate the spatiotemporal information of the input data.

[0185] A state decoder is used to recombine the processed high-level features and backmap them back to the original data space through the network hierarchy to generate the chlorophyll a concentration prediction result.

[0186] In one feasible embodiment, the input data is mapped to the latent space by a deep convolutional encoder (E). A potential state (h) is generated. This potential state is then decomposed into the physical dynamics of chlorophyll a concentration (ha). p ), physical dynamics of other variables (h) o ), and unknown dynamics (h r The PhyCell unit, by simulating differential operators, predicts the physical dynamics of chlorophyll a concentration and other variables, effectively utilizing physical laws to improve prediction accuracy. The ConvLSTM unit can combine information from previous moments regarding chlorophyll a concentration to capture residual unknown dynamics (i.e., residual dynamics) affecting chlorophyll a concentration. The updated h... p h o and h r The chlorophyll a concentration is then decoded and remapped using a sigmoid activation function to generate the final prediction. The PhyDNet model integrates PhyCell modules and other physical feature information to comprehensively predict chlorophyll a concentration, while also considering the interaction of multiple environmental variables, thus providing more accurate predictions in complex environments.

[0187] Optionally, the physical dynamics of chlorophyll a concentration and other variables predicted by the PhyCell unit are as follows:

[0188]

[0189]

[0190] Where DEC stands for State Decoder. and These represent the physical dynamics affecting chlorophyll a concentration and other variables affecting chlorophyll concentration at the next time step, respectively. and The inputs to the unit are the physical dynamics affecting chlorophyll a concentration at the current moment and the physical dynamics affecting other variables affecting chlorophyll concentration, respectively.

[0191] Optionally, the residual dynamics affecting chlorophyll a concentration predicted by the ConvLSTM unit are as follows:

[0192]

[0193] Where DEC stands for State Decoder. This is to determine the residual dynamics affecting chlorophyll a concentration at the next moment. The residual dynamics affecting chlorophyll a concentration at the current moment.

[0194] In one feasible embodiment, reference is made to Figure 5 The chlorophyll a prediction model includes:

[0195] Downsampling block (conv-block): When constructing the downsampling layer, a convolutional operation is performed first, followed by a group normalization layer to normalize the output, and finally LeakyReLU is used as a non-linear activation function to enable the network to learn complex feature maps.

[0196] Deep Convolutional Encoder (E): Employs a cascaded structure of three downsampling blocks to perform depthwise convolution on the input sequence, progressively extracting features, and then mapping the input data to the latent space.

[0197] Status encoder (E) p E o and E r The two-layer downsampling cascaded structure is used to decouple the potential states and input them to the PhyCell unit and the ConvLSTM unit respectively.

[0198] PhyCell element: a type of physically constrained recursive element that predicts (Φ(h)). p )) and correction (C(h p E(u))) captures physical dynamics. The PhyCell cell uses spatial derivatives to simulate linear partial differential equations, achieving forward Euler discretization of the latent prediction. This prediction relies solely on the hidden representation (h(t)). p The correction step guides assimilation (C(h)). p The new hidden state (h(t+Δ)) is calculated using the Kalman gain (K(t)). p The specific expression is:

[0199]

[0200]

[0201] in, For the prediction term, h t+1 For the correction term, (K(t)) is the gating factor used to control the trade-off between prediction and correction; K is the Kalman gain, which is defined as:

[0202]

[0203] When K(t) = 0, it means there is no input contribution, while when K(t) = 1, the entire prediction is discarded and the hidden state is reset according to the input.

[0204] The ConvLSTM unit is a recurrent neural network (RNN) for spatiotemporal sequence prediction, employing convolution operations in both state-to-state and input-to-state transitions. This structure enables it to more effectively capture spatial dependencies, helping to better integrate spatial information into the network dynamics. In this embodiment, the ConvLSTM unit is used to extract spatiotemporal features from the latent space, process residual dynamics, and accurately predict unknown dynamics h. r The computation process inside a ConvLSTM cell includes:

[0205] it=σ(W xi *X t +W hi *H t-1 +W ci ⊙C t-1 +b i )

[0206] ft=σ(W xf *X t +W hf *H t-1 +W cf ⊙C t-1 +b f )

[0207] C t =ft⊙C t-1 +it⊙tanh(W xc *X t +W hc *H t-1 +b c )

[0208] ot=σ(W xo *X t +W ho *H t-1 +W co ⊙C t +b o )

[0209] H t =ot⊙tanh(Ct )

[0210] Among them, X t H represents the input for time step t. t-1 It is the hidden state of the previous time step, C t-1 This represents the cell output at the previous time step, and σ is the sigmoid activation function. The symbols * and ⊙ represent convolution and Hadamard product (element-by-element multiplication), respectively.

[0211] Upsampling blocks (UpConv-blocks) are the parts of convolutional neural networks (CNNs) used to map feature maps from small resolution to large resolution, improving the model's ability to recognize spatial information. First, a transposed convolutional layer is used, then the output is normalized using a group normalization layer, and finally, LeakyReLU is applied as a non-linear activation function to generate higher-level features in the latent space.

[0212] GroupNorm is a novel normalization method that first divides the channels into multiple groups and normalizes each group, reshaping the feature dimensions from [N,C,H,W] to [N,G,C / / G,H,W], with the normalized dimension being [C / / G,H,W]. Between the convolutional layers, a GroupNormlayer is introduced to normalize the group normalization layer of the output. LeakyReLU enables the network to learn non-linear activation functions for complex feature maps, as shown in the following formula:

[0213]

[0214] Here, negative_slope represents the coefficient when x is negative, used to control the angle of negative slope, with a default value of 0.01.

[0215] State Decoder D p D o and D r A dual upsampling block cascaded structure is adopted. The unwrapped state is mapped back to the latent space to construct the state decoder.

[0216] Decoder D: This is a module for generating images, containing two upsampling blocks and a transposed convolutional layer (upconv). It progressively transforms the output features into image form through upsampling operations, achieving a mapping from the latent space to the image.

[0217] For both downsampling and upsampling blocks, this application employs a 3×3 convolutional kernel. When the stride is 1, the kernel performs a matrix shape identity mapping (H,W)→(H,W); when the stride is 2, it performs a compression mapping (H,W)→(H / 2,W / 2) to encode the features and generate a higher-dimensional feature representation. For the deconvolutional layer, when the stride is 1, it performs a matrix shape identity mapping (H,W)→(H,W); when the stride is 2, it performs an expansion mapping (H,W)→(2H,2W) to restore the features.

[0218]

[0219]

[0220] The encoder (DEC) consists of three downsampling blocks (upconv-blocks) used to construct the latent space. Conversely, the decoder employs two upsampling blocks (upconv-blocks) and a transposed convolutional layer (upconv) to implement the image reconstruction mapping. To achieve state encoding and decoding, this application designs six key modules: encoder_Ep, encoder_Eo, encoder_Er, decoded_Dp, decoded_Do, and decoded_Dr. These modules achieve mutual transformation between the state mapping and the latent space through the superposition of upsampling blocks (conv-blocks) and downsampling blocks (upconv-blocks). In the EncoderRNN encoding and decoding network for chlorophyll a concentration prediction, the model structure can be summarized as follows:

[0221]

[0222]

[0223]

[0224] Furthermore, embodiments of this application also provide a chlorophyll a prediction device, referring to... Figure 6 The chlorophyll a prediction device is applied to an electronic device, and the chlorophyll a prediction device includes:

[0225] The acquisition module 10 is used to acquire a remote sensing dataset of the target sea area, wherein the remote sensing dataset includes: first remote sensing data with chlorophyll a concentration and second remote sensing data without chlorophyll a concentration;

[0226] The filling module 20 is used to fill the chlorophyll a concentration of the second remote sensing data with the first remote sensing data to obtain a continuous spatiotemporal sequence dataset;

[0227] Training module 30 is used to train a preset model to be trained using the spatiotemporal sequence dataset to obtain a chlorophyll a prediction model.

[0228] The prediction module 40 is used to input the data of the sea area to be predicted into the chlorophyll a prediction model to make a prediction and obtain the chlorophyll a concentration prediction result.

[0229] The specific implementation of the television content recognition device in this application is basically the same as the embodiments of the television content recognition method described above, and will not be repeated here.

[0230] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0231] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0232] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for predicting chlorophyll a concentration, characterized in that, The method includes the following steps: Obtain a remote sensing dataset of the target sea area, wherein the remote sensing dataset includes: first remote sensing data with chlorophyll a concentration and second remote sensing data without chlorophyll a concentration; Based on the preset mapping relationship between pixel position and physical position, the first physical position corresponding to each pixel point contained in each of the second remote sensing data is matched and obtained; The importance of each of the first physical locations is determined based on the chlorophyll a concentration change rate, wherein the chlorophyll a concentration change rate is directly proportional to the importance. Based on the importance of each of the first physical locations, a predetermined proportion of the first physical locations are selected as second physical locations, and the remaining first physical locations are selected as third physical locations. The chlorophyll a concentration change rate of the second physical location is higher than that of the third physical location. Based on the LSTM model, the missing data at the second physical location is predicted to obtain the first predicted data; Based on the XGboost model, the missing data at the third physical location is predicted to obtain the second predicted data; Based on the first and second prediction data, missing data is filled into the second remote sensing data to obtain a spatiotemporal sequence dataset. The spatiotemporal sequence dataset is used to train the preset model to be trained, and a chlorophyll a prediction model is obtained. The data of the sea area to be predicted is input into the chlorophyll a prediction model to make a prediction and obtain the chlorophyll a concentration prediction result. The chlorophyll a prediction model includes: A deep feature encoder is used to extract high-level features from input data and map the high-level features to a latent space; PhyCell cells are used to simulate dynamic changes in physical processes and predict chlorophyll a concentration based on differential operator modeling. ConvLSTM units are used to combine convolutional neural networks and long short-term memory networks and capture the spatial dependencies in the input data to generate the spatiotemporal information of the input data. A state decoder is used to recombine the processed high-level features and backmap them back to the original data space through the network hierarchy to generate the chlorophyll a concentration prediction result.

2. The chlorophyll a concentration prediction method as described in claim 1, characterized in that, The remote sensing dataset also includes environmental influencing factors, which include at least one of ocean surface temperature, aerosol index, and particulate backscattering coefficient.

3. The chlorophyll a concentration prediction method as described in claim 1, characterized in that, The step of training a preset model to obtain a chlorophyll a prediction model using the spatiotemporal sequence dataset includes: The spatiotemporal sequence corresponding to each physical location in the spatiotemporal sequence dataset is divided into multiple subsequences with a preset first duration. The data of the first second preset duration in each of the sub-sequences is used as the model input, and the data of the remaining duration is used as the label to generate training data; The training data is used to train the model to be trained, and the chlorophyll a prediction model is obtained.

4. The chlorophyll a concentration prediction method as described in claim 1, characterized in that, The step of training a preset model to obtain a chlorophyll a prediction model using the spatiotemporal sequence dataset includes: Identify the fourth physical location contained in the spatiotemporal sequence dataset; Determine the prediction complexity of each of the fourth physical locations, and set the training weights of each of the fourth physical locations based on the prediction complexity; Based on the training weights of each of the fourth physical locations, and using the spatiotemporal sequence dataset to train the model to be trained, the chlorophyll a prediction model is obtained.

5. The chlorophyll a concentration prediction method as described in claim 1, characterized in that, Before the step of inputting the data of the sea area to be predicted into the chlorophyll a prediction model for prediction, the method further includes: A composite loss function is constructed based on image reconstruction loss and matrix loss; The chlorophyll a prediction model is optimized based on the composite loss function.

6. A chlorophyll a concentration prediction device, characterized in that, The device includes: The acquisition module is used to acquire a remote sensing dataset of the target sea area, wherein the remote sensing dataset includes: first remote sensing data with chlorophyll a concentration and second remote sensing data without chlorophyll a concentration; The imputation module is used to: match and obtain the first physical location corresponding to each pixel in each of the second remote sensing data according to a preset mapping relationship between pixel location and physical location; determine the importance of each first physical location according to the chlorophyll a concentration change rate, wherein the chlorophyll a concentration change rate is proportional to the importance; select a preset proportion of first physical locations from the first physical locations based on the importance of each first physical location to serve as second physical locations, and use the remaining first physical locations as third physical locations, wherein the chlorophyll a concentration change rate of the second physical locations is higher than that of the third physical locations; predict the missing data of the second physical locations based on an LSTM model to obtain first predicted data; predict the missing data of the third physical locations based on an XGboost model to obtain second predicted data; and impute the missing data of the second remote sensing data according to the first predicted data and the second predicted data to obtain a spatiotemporal sequence dataset. The training module is used to train the preset model to be trained using the spatiotemporal sequence dataset to obtain the chlorophyll a prediction model. The prediction module is used to input the data of the sea area to be predicted into the chlorophyll a prediction model to make predictions and obtain the chlorophyll a concentration prediction results. The chlorophyll a prediction model includes: A deep feature encoder is used to extract high-level features from input data and map the high-level features to a latent space; PhyCell cells are used to simulate dynamic changes in physical processes and predict chlorophyll a concentration based on differential operator modeling. ConvLSTM units are used to combine convolutional neural networks and long short-term memory networks and capture the spatial dependencies in the input data to generate the spatiotemporal information of the input data. A state decoder is used to recombine the processed high-level features and backmap them back to the original data space through the network hierarchy to generate the chlorophyll a concentration prediction result.

7. An electronic device, characterized in that, The device includes: a memory, a processor, and a chlorophyll a concentration prediction program stored in the memory and executable on the processor, the chlorophyll a concentration prediction program being configured to implement the steps of the chlorophyll a concentration prediction method as described in any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and the storage medium stores a chlorophyll a concentration prediction program, which, when executed by a processor, implements the steps of the chlorophyll a concentration prediction method as described in any one of claims 1 to 5.

Images