Ocean red tide early warning method based on remote sensing inversion and time-space map convolutional network

Abstract

The invention provides an ocean red tide early warning method based on remote sensing inversion and a time-space map convolutional network. The ocean red tide early warning method comprises the following steps: collecting hyperspectral data and multispectral data; respectively carrying out data quality inspection on the collected hyperspectral data and multispectral data; preprocessing the hyperspectral data and the multispectral data after the quality inspection is qualified to respectively obtain a hyperspectral result image and a digital orthoimage; analyzing the hyperspectral result image and the digital orthoimage to obtain remote sensing inversion data of water temperature, salinity, total nitrogen, total phosphorus and chlorophyll a, and generating a gradient distribution thematic map of each index; according to water temperature, salinity, total nitrogen, total phosphorus, chlorophyll a and historical monitoring data, in combination with a change trend of key environmental parameters, introducing a space-time diagram convolutional network model, simulating a red tide diffusion situation, setting a risk threshold, and performing risk prediction on red tide occurrence through a red tide occurrence risk prediction model; and updating the red tide occurrence risk prediction model based on buoy monitoring and field monitoring data.

Description

Technical Field

[0001] The present invention relates to the technical field of marine red tide monitoring, and specifically relates to a marine red tide early warning method based on remote sensing inversion and spatio-temporal graph convolutional network. Background Technique

[0002] Red tide, also known as red tide or algal bloom, is an ecological anomaly that causes the discoloration of local water bodies under specific environmental conditions, resulting from the explosive proliferation or high aggregation of certain phytoplankton, protozoa or bacteria in the water body. The occurrence of red tide is often closely related to environmental factors such as water pollution, excess nutrients, rising water temperature, and sufficient light.

[0003] Real-time monitoring and early warning of red tides mainly utilize the national marine environmental monitoring network composed of satellite remote sensing, unmanned aerial vehicles, ships, buoys and shore stations. By monitoring the occurrence, development and spread of red tides, early warning information is released in a timely manner. In recent years, China has achieved remarkable results in red tide prevention and control. By strengthening monitoring, controlling pollution sources, improving water quality and other measures, the occurrence and harm of red tide disasters have been effectively reduced. However, red tide early warning still faces many challenges, such as complex red tide causes, limited monitoring means, and imperfect early warning technologies. Summary of the Invention

[0004] The present invention provides a marine red tide early warning method based on remote sensing inversion and spatio-temporal graph convolutional network to solve the above technical problems.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A marine red tide early warning method based on remote sensing inversion and spatio-temporal graph convolutional network includes the following steps: collecting hyperspectral data and multispectral data; respectively performing data quality inspections on the collected hyperspectral data and multispectral data. If the data quality inspection is unqualified, re-collect the hyperspectral data and multispectral data until the data quality inspection is qualified; preprocess the hyperspectral data and multispectral data after passing the quality inspection to obtain hyperspectral result images and digital orthophotos respectively; analyze the hyperspectral result images and the digital orthophotos to obtain remote sensing inversion data of water temperature, salinity, total nitrogen and total phosphorus, and chlorophyll a, and generate thematic maps of the gradient distribution of each index; according to the water temperature, salinity, total nitrogen and total phosphorus, chlorophyll a and historical monitoring data, combined with the change trend of key environmental parameters, introduce a spatio-temporal graph convolutional network model to simulate the red tide diffusion trend, set a risk threshold, and predict the risk of red tide occurrence through a red tide occurrence risk prediction model; update the red tide occurrence risk prediction model based on buoy monitoring and on-site monitoring data.

[0007] Further, the preprocessing of the hyperspectral data includes the following steps: performing band combination on the collected hyperspectral images; performing radiometric calibration using the absolute radiometric calibration coefficient to convert the brightness gray values of the hyperspectral images into absolute radiance values, eliminating the radiation bias of the sensor itself, and ensuring the radiation consistency of the images; performing atmospheric correction using an atmospheric correction algorithm to convert the apparent radiance data into surface reflectance data and removing interference factors; collecting control points and checking residuals, and performing geometric correction on the hyperspectral images after atmospheric correction; performing orthorectification on the hyperspectral images, using the digital elevation model DEM to correct the geometric distortion caused by the terrain, and eliminating the projection differences caused by terrain undulation and photo tilt; performing mosaicking on multiple hyperspectral images, and performing color homogenization operations, cropping the hyperspectral images, and extracting the hyperspectral image data of the target area; enhancing the contrast and clarity of the hyperspectral satellite images through image processing techniques, and finally obtaining the hyperspectral result images.

[0008] Further, the multispectral data includes panchromatic images and multispectral images, and its preprocessing includes the following steps: performing radiometric calibration on the panchromatic images to obtain surface reflectance data, and performing orthorectification in combination with DEM data and RPC files to obtain panchromatic orthoimages; performing radiometric calibration, atmospheric correction, and orthorectification on the multispectral images to obtain multispectral orthoimages; fusing the obtained panchromatic orthoimages and the multispectral orthoimages, and performing mosaicking on the fused images to achieve seamless splicing of adjacent images; performing geometric fine correction on the mosaicked images, and finally outputting digital orthoimages.

[0009] Further, the thematic maps of the gradient distributions of each index include the sea surface temperature distribution thematic map, the sea surface salinity distribution thematic map, and the total nitrogen, total phosphorus, and chlorophyll a concentration distribution thematic map. Among them, the generation of the sea surface temperature distribution thematic map includes the following steps: First, preprocess the multispectral data, including radiometric calibration and atmospheric correction; Second, calculate according to the normalized difference vegetation index, and use the normalized difference vegetation index NDVI to obtain the coverage image, and finally obtain the surface emissivity result; Then, use the radiance image of the near-infrared band, combine the surface emissivity and the atmospheric profile parameters, and calculate the blackbody radiance and the surface temperature. The atmospheric profile parameters include the atmospheric transmittance, the downward radiance, and the upward radiance; Finally, select the spectral bands according to the normalized difference water index model and perform the normalized difference water index calculation, extract the boundary information of the offshore water body, form the sea surface temperature gradient change, and form the sea surface temperature distribution thematic map.

[0010] Furthermore, the generation of the thematic map of sea surface salinity distribution includes the following steps: First, preprocess the multispectral data, including radiometric calibration and atmospheric correction; then, use the historical data of sea surface salinity stations as training samples, select the feature bands with strong correlation, and establish a non-linear relationship between the remote sensing reflectance and the sea surface salinity value through an artificial neural network to construct an inversion model; finally, according to the water body boundary range, use the remote sensing reflectance data of the sensitive feature bands as the input of the artificial neural network model, perform band calculation inversion to obtain the sea surface salinity gradient change result, and form the thematic map of sea surface salinity distribution.

[0011] Furthermore, the generation of the thematic maps of total nitrogen, total phosphorus, and chlorophyll a concentration distribution includes the following steps: First, preprocess the hyperspectral data and multispectral data, including geometric correction, radiometric calibration, and atmospheric correction; second, select spectral bands according to the normalized water body index model and calculate the normalized water body index NDWI to extract the information of the offshore water body boundary; then, conduct a correlation analysis by combining the historical data of water quality parameters and the water body spectral data, select the sensitive bands of total nitrogen, total phosphorus, and chlorophyll a, and establish empirical and semi-empirical models of the best bands or various band combinations; finally, use the established inversion model combined with the preprocessed remote sensing image, perform band calculation inversion to obtain the concentration gradient change results of total nitrogen, total phosphorus, and chlorophyll a, and form the corresponding concentration distribution thematic maps respectively.

[0012] Furthermore, the red tide risk prediction includes the following steps:

[0013] Collect data including water temperature, salinity, total nitrogen, total phosphorus, chlorophyll a, and historical monitoring data;

[0014] Construct a red tide occurrence risk prediction model, including the following steps:

[0015] Construct a spatio-temporal graph structure: Divide the ocean area into several grids, and each grid serves as a node of the spatio-temporal graph; construct the edges of the graph according to the geographical adjacency relationship and the sea current direction to form a spatio-temporal graph structure; extract the environmental parameters and time series data for each node as node features;

[0016] Construct a spatio-temporal graph convolutional network model:

[0017] Use the spatio-temporal graph convolutional network to capture spatial features. For each node i, its updated feature can be represented by the following formula:

[0018]

[0019] Among them, represents the feature representation of node i at the l-th layer, N(i) represents the neighbor set of node i, c ij is the normalization constant, W (l) and b (l)They are the weight matrix and bias term of the l-th layer respectively, and σ is the activation function;

[0020] The temporal convolutional network is adopted to capture the temporal features in the time series data. For each time step t, its output can be expressed by the following formula:

[0021] O t = TCN(X t )

[0022] In the formula, X t represents the input feature at time step t, and TCN represents the operation of the temporal convolutional network;

[0023] The attention mechanism is introduced to weight the key environmental factors, and its formula can be expressed as:

[0024]

[0025] In the formula, Qi, Ki, and Vi represent the query, key, and matrix value respectively, dk represents the dimension of the key, and the softmax function is used to calculate the attention weight;

[0026] To better fuse the spatial and temporal features, the outputs of the graph convolutional layer and the temporal convolutional layer are fused. Among them, the output of the graph convolutional layer is represented by h graph and the output of the temporal convolutional layer is represented by h time . Then the fused node feature h fusion is calculated by the following formula:

[0027] h fusion = αh graph + (1 - α)h time

[0028] In the formula, α is the weight parameter, which is a learnable weight parameter and its optimal value is determined through training to balance the contributions of the spatial and temporal features;

[0029] Red tide occurrence risk prediction:

[0030] The fully connected layer is used to map the fused node features to the red tide occurrence risk prediction model. The formula is

[0031] P = α(W fc h fusion + b fc )

[0032] In the formula, P represents the predicted risk value, W fc and b fc represent the weight and bias term of the fully connected layer respectively, and σ is the activation function;

[0033] Set the risk threshold as y. When the predicted risk value P exceeds the threshold y, it is considered that there is a risk of red tide occurrence in this area.

[0034] The predicted risk value P is divided into the following three intervals to represent the red tide occurrence risk level:

[0035] When 0 ≤ P < 0.3, the risk level is considered low;

[0036] When 0.3 ≤ P < 0.6, the risk level is considered medium;

[0037] When 0.6 ≤ P ≤ 1, the risk level is considered high.

[0038] The update of the red tide occurrence risk prediction model includes the following steps:

[0039] Divide the buoy monitoring and on-site monitoring data into a training set, a validation set, and a test set;

[0040] Use the cross-entropy loss function to measure the difference between the prediction result and the true label. The formula is:

[0041]

[0042] In the formula, N represents the number of samples, y i represents the true label of sample i, where the presence of red tide is 1 and the absence is 0, and P i represents the predicted risk probability of sample i;

[0043] Use the Adam optimizer to update the parameters of the red tide occurrence risk prediction model to minimize the loss function. The formula for updating the parameters is:

[0044] θ t+1 = θ t - η t m t

[0045] In the formula, θ t represents the parameter at time t, η t represents the learning rate, and m t represents the update amount based on momentum.

[0046] The beneficial effects of the present invention:

[0047] The ocean red tide early warning method and system based on remote sensing inversion and spatio-temporal graph convolutional network processes the collected hyperspectral and multispectral data to ensure the integrity, accuracy, consistency, reliability, and timeliness of the data. A spatio-temporal graph neural network is used to construct a red tide occurrence risk prediction model based on the processed hyperspectral and multispectral data, and early warning of red tide occurrence is achieved by fusing satellite remote sensing, buoy monitoring, and historical data. An attention mechanism is introduced to focus on the change trends of key environmental factors to improve prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of the ocean red tide early warning method based on remote sensing inversion and spatio-temporal graph convolutional network according to an embodiment of the present invention;

[0049] Figure 2 It is a flowchart of the hyperspectral data preprocessing according to an embodiment of the present invention;

[0050] Figure 3 It is a flowchart of the multispectral data preprocessing according to an embodiment of the present invention;

[0051] Figure 4 It is a flowchart of the generation of the thematic map of sea surface temperature distribution according to an embodiment of the present invention;

[0052] Figure 5 It is a flowchart of the generation of the thematic map of sea surface salinity distribution according to an embodiment of the present invention;

[0053] Figure 6 It is a flowchart of the generation of the thematic map of total nitrogen, total phosphorus, and chlorophyll a concentration distribution. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0055] As Figure 1 shown, an ocean red tide early warning method based on remote sensing inversion and spatio-temporal graph convolutional network according to an embodiment of the present invention includes the following steps:

[0056] S1: Collect hyperspectral data and multispectral data.

[0057] Due to the often excessive cloud cover over the sea, the clouds obscure the ocean surface, resulting in the inability to reflect the characteristic bands of sea surface temperature, sea surface salinity, total nitrogen, total phosphorus, and chlorophyll a in remote sensing data, which affects the integrity and accuracy of the data. The presence of clouds not only obscures some sea area information, causing data loss, but may also interfere with the accurate monitoring and analysis of ocean environmental parameters because clouds absorb and scatter electromagnetic waves, changing the radiation signals received by the sensors, thereby affecting data quality.

[0058] In addition, the optical properties of water bodies are complex and variable, and different water quality parameters may produce similar spectral characteristics, which brings difficulties to spectral analysis. At the same time, there are mixed pixels at the water-land boundary, and the same pixel may contain multiple types of information, increasing the difficulty of inverting water quality parameters. For example, in the coastal area, the water body contains various components such as sediment and plankton, and their spectral characteristics are intertwined, making it difficult to accurately distinguish and analyze, thus affecting the inversion accuracy of parameters such as chlorophyll a.

[0059] Therefore, for the above reasons, in the hyperspectral data acquisition of the present invention, multi-spectral remote sensing satellite data and cloud detection algorithms are used for cloud identification and removal. By analyzing the information of different spectral bands, the position and range of clouds can be accurately judged. At the same time, satellite programming photography is used to revisit the monitoring area at high frequencies, increasing the opportunity to obtain cloud-free or less-cloudy images to acquire image data that meets the monitoring requirements. And the multi-temporal data fusion technology is adopted to fill the data in the cloud-missing area, fuse the image data obtained at different times, and use the data in the cloud-free area to supplement the information in the cloud-covered area, improving the integrity and usability of the data. A hyperspectral sensor and an optimized spatial resolution are adopted. The hyperspectral sensor can capture radiation of different wavelengths, and the data of its 32 bands can capture richer spectral information, which helps to distinguish the spectral differences brought about by changes in different water quality parameters, thereby effectively improving the accuracy of water quality monitoring. And the 10-meter spatial resolution can effectively solve the problem of mixed pixels of medium-resolution satellites, can provide more detailed surface coverage information, better distinguish the water-land boundary, reduce the influence of mixed pixels, improve the accurate identification of water bodies and land information in the coastal sea area, and further improve the accuracy of inverting parameters such as chlorophyll a.

[0060] Similar to hyperspectral data acquisition, the cloud cover and shadows over the sea will obscure multi-spectral images, affecting data integrity and accuracy, resulting in the loss of information in some areas and being unable to accurately reflect the true distribution of information such as sea surface temperature and salinity. Excessive cloud cover may completely obscure some sea areas in the image, and the shadow part may change the radiation characteristics of ground objects, bringing difficulties to data processing and analysis.

[0061] To more accurately describe the spatio-temporal distribution characteristics of sea surface temperature and salinity in the sea area in a timely manner, it is necessary to ensure that the data accuracy, resolution, etc. meet the requirements. However, in the actual acquisition process, affected by various factors such as sensor performance and atmospheric conditions, it may be difficult to obtain this information completely accurately.

[0062] For the above reasons, during the acquisition of multi-spectral data, the present invention strictly controls the cloud cover of the overall data in a single period to be less than 15%, and the cloud cover and shadow of the original per-scene data do not cover the main monitoring area. By selecting appropriate acquisition times and satellite shooting angles, the influence of cloud cover and shadow is reduced. For the acquired multi-spectral data, a cloud detection algorithm is used for cloud identification and processing, and the shadow part is corrected and compensated to improve the data quality and ensure that information such as sea surface temperature and salinity can be accurately obtained.

[0063] The multi-spectral data of the land imager is used, and its technical indicators such as band settings and spatial resolution can better meet the requirements for describing the spatio-temporal distribution characteristics of sea surface temperature and salinity. For example, its thermal infrared band can be used for sea surface temperature monitoring, and different band combinations can be used to analyze information related to salinity. By reasonably using these data resources and combining accurate data processing technologies such as radiometric calibration and atmospheric correction, the accuracy of sea surface temperature and salinity monitoring is improved, so as to more accurately describe their spatio-temporal distribution characteristics.

[0064] S2: Perform data quality inspections on the acquired hyperspectral data and multi-spectral data respectively. If the data quality inspection is unqualified, re-acquire the hyperspectral data and multi-spectral data until the data quality inspection is qualified.

[0065] Among them, the acquired hyperspectral data and multi-spectral data need to be complete, accurate, consistent, reliable and timely. For the integrity of the data, it is necessary to ensure that the acquired satellite image data, hyperspectral and multi-spectral data are collected completely in each period without missing areas. The offshore chlorophyll a concentration, sea surface temperature, and sea surface salinity are completely obtained within the specified monitoring area, and the corresponding monitoring reports, thematic maps, raster results and other data are complete without missing any key information.

[0066] For the accuracy of the data, the technical indicators such as the spatial resolution, number of bands, and spectral range of the satellite image need to meet the requirements. For example, the spatial resolution of the hyperspectral satellite image is better than 10 meters, 32 bands, and the spectral range is 400 - 1000 nm. The multi-spectral satellite image has a spatial resolution better than 1000 meters, 11 bands, and the spectral range is 400 - 200 nm. During the acquisition process, the cloud cover is strictly controlled, with the cloud cover of a single scene image less than 20% and the overall area less than 15%, reducing the influence of clouds on the accuracy of the image data and ensuring that the image can clearly and accurately reflect the surface characteristics of the sea area.

[0067] For data consistency, similar data collected at different times (such as hyperspectral or multispectral image data in different quarters) should be consistent in terms of spatial resolution, spectral range, etc., in order to conduct effective time series analysis and comparative research. The various thematic data should be coordinated with each other. For example, the distribution data of chlorophyll a concentration, sea surface temperature, and sea surface salinity should conform to the actual marine environmental conditions, without contradictions or conflicts, and jointly provide a reliable basis for marine ecological environment assessment and research.

[0068] For data reliability, the satellite data sources used for data collection have good stability and reliability, and can continuously and stably provide high-quality data. The technical methods and models adopted in the data processing and thematic production processes have been verified and optimized, with high reliability, and can ensure the authenticity and effectiveness of the data, enabling the project results to truly reflect the actual situation of the sea area and providing credible data support for marine resource management and environmental protection.

[0069] For data reliability, it is updated quarterly to ensure that the data can timely reflect the dynamic changes of the sea area. The chlorophyll a concentration, sea surface temperature, and sea surface salinity in the offshore area need to be generated and provided quarterly on time, ensuring the continuity and timeliness of the data in terms of time, so as to timely grasp the changing trends of the marine environment and provide timely and effective data basis for relevant decisions.

[0070] S3: Preprocess the qualified hyperspectral data and multispectral data after quality inspection to obtain hyperspectral result images and digital orthophotos respectively.

[0071] In the embodiment of the present invention, as Figure 2 shown, the preprocessing of the hyperspectral data includes the following steps:

[0072] Combine the collected hyperspectral images by bands.

[0073] Perform radiometric calibration using the absolute radiometric calibration coefficient to convert the brightness gray values of the hyperspectral images into absolute radiance values, eliminate the radiation deviation of the sensor itself, and ensure the radiation consistency of the images.

[0074] Perform atmospheric correction using an atmospheric correction algorithm (correction method based on the 6S model) to convert the apparent radiance data into surface reflectance data and remove interference factors such as atmospheric scattering.

[0075] Collect control points and check residuals, and perform geometric correction on the hyperspectral images after atmospheric correction.

[0076] Perform orthorectification on the hyperspectral images, and use the digital elevation model DEM to correct the geometric distortion caused by the terrain, and eliminate the projection differences caused by terrain undulation and photo tilt.

[0077] Mosaic multiple of the hyperspectral images, perform color homogenization operation, crop the hyperspectral images, and extract the hyperspectral image data of the target area.

[0078] Enhance the contrast and clarity of the hyperspectral satellite images through image processing techniques, and finally obtain the hyperspectral result images.

[0079] In an embodiment of the present invention, as Figure 3 shown, the multispectral data includes panchromatic images and multispectral images, and its preprocessing includes the following steps:

[0080] Perform radiometric calibration on the panchromatic images to obtain surface reflectance data, and perform orthorectification in combination with DEM data and RPC files to obtain panchromatic orthoimages; perform radiometric calibration, atmospheric correction, and orthorectification on the multispectral images to obtain multispectral orthoimages.

[0081] Fuse the obtained panchromatic orthoimages and the multispectral orthoimages, and perform mosaic processing on the fused images to achieve seamless splicing of adjacent images.

[0082] Perform geometric fine correction on the mosaicked images, and finally output digital orthoimages.

[0083] S4: Analyze the hyperspectral result images and the digital orthoimages to obtain remote sensing inversion data of water temperature, salinity, total nitrogen and total phosphorus, and chlorophyll a, and generate thematic maps of the gradient distribution of each index.

[0084] Specifically, perform correlation analysis and regression analysis through ENVI and SPSS software to obtain remote sensing inversion data of water temperature, salinity, total nitrogen and total phosphorus, and chlorophyll a.

[0085] Among them, the thematic maps of the gradient distribution of each index include the sea surface temperature distribution thematic map, the sea surface salinity distribution thematic map, and the total nitrogen, total phosphorus and chlorophyll a concentration distribution thematic map.

[0086] In an embodiment of the present invention, as Figure 4 shown, the generation of the sea surface temperature distribution thematic map includes the following steps:

[0087] First, preprocess the multispectral data, including radiometric calibration and atmospheric correction;

[0088] Secondly, calculate according to the normalized difference vegetation index, and use the normalized difference vegetation index NDVI to obtain the coverage image, and finally obtain the surface emissivity result;

[0089] Then, using the radiance image in the near-infrared band and combining with the land surface emissivity and atmospheric profile parameters, calculate the blackbody radiance and land surface temperature. The atmospheric profile parameters include atmospheric transmittance, downward radiance, and upward radiance;

[0090] Finally, select spectral bands according to the normalized water index model and calculate the normalized water index to extract the offshore water body boundary information, form the sea surface temperature gradient change, and form a thematic map of sea surface temperature distribution.

[0091] The above steps are further described as follows:

[0092] (1) Data preprocessing:

[0093] Geometric correction (ENVI): Prepare reference data, select a correction method, collect control points, and use the least squares method to calculate the correction model and resampling method to generate the corrected image;

[0094] Radiometric calibration (ENVI): Obtain calibration coefficients, use tools to convert DN values, and generate the calibrated image in the required output format as needed;

[0095] Atmospheric correction (ENVI): Prepare relevant information, perform FLAASH correction, check the spectral curve for evaluation and adjust parameters to ensure accuracy;

[0096] Geometric fine correction (ArcGIS): Ensure that information such as projection is correct, select a correction method, collect control points and check the residuals, and perform operations to generate a high-precision image.

[0097] (2) Water body extraction: Load the image in ENVI to calculate the normalized difference water index NDWI, use the decision tree tool to distinguish water bodies from non-water bodies, export the binary result in vector format and then convert it to shpfile format, and import and clip the cropping area data in ArcGIS to obtain water body vector data.

[0098] (3) Blackbody radiance calculation: Load the image in ENVI to calculate the blackbody radiance, and use bands such as atmospheric upward radiance, atmospheric downward radiance, radiance, and atmospheric transmittance to calculate and obtain the blackbody radiance at the same temperature.

[0099] (4) Sea surface temperature inversion: Load the blackbody radiance obtained in the previous step in ENVI, use the Planck formula to invert and obtain the sea surface temperature, and convert the Kelvin unit to Celsius temperature.

[0100] In an embodiment of the present invention, as Figure 5 shown, the generation of the thematic map of sea surface salinity distribution includes the following steps:

[0101] First, preprocess the multispectral data, including radiometric calibration and atmospheric correction;

[0102] Then, using the historical data of sea surface salinity sites as training samples, select the feature bands with strong correlation, and through an artificial neural network, establish a non-linear relationship between remote sensing reflectance and sea surface salinity values to construct an inversion model.

[0103] Finally, according to the water body boundary range, use the remote sensing reflectance data of sensitive feature bands as the input of the artificial neural network model, perform band calculation inversion to obtain the sea surface salinity gradient change result, and form a thematic map of sea surface salinity distribution.

[0104] The above steps are further described as follows:

[0105] (1) Data preprocessing:

[0106] Radiometric calibration (Seadas): Obtain the calibration coefficient, convert the DN value, and generate the calibrated image with the output format set as required; Atmospheric correction (Seadas): Prepare relevant information, perform atmospheric correction, remove the influence of atmospheric scattering and absorption on the image, obtain the true reflectance of the ground object, and check the spectral curve to evaluate and adjust the parameters to ensure accuracy.

[0107] (2) Land masking: Use the l2_flag generated during the atmospheric correction process for land masking to generate the remote sensing reflectance result without land.

[0108] (3) Model establishment: Using correlation analysis, combined with sea surface salinity site data, select the feature bands with strong correlation. Based on the feature bands, through an artificial neural network (Artificial Neural Network, ANN), establish a non-linear relationship between remote sensing reflectance and remote sensing reflectance to construct a sea surface salinity inversion model. ANN is composed of a large number of artificial neurons connected to each other. Each neuron receives input signals, processes these inputs through an activation function, and generates output signals. The connections between neurons have weights, and the weights determine the strength and influence of signal transmission between neurons. The weights can be adjusted automatically through training. ANN includes an input layer, a hidden layer, and an output layer. The input layer receives the original data, the hidden layer performs multi-layer abstraction and processing on the data, and the output layer generates the final result.

[0109] Extract the sensitive bands for salinity inversion through correlation analysis as the input bands of the artificial neural network. Use the sea surface salinity data of offshore monitoring stations over the years as training data, and form one-to-one corresponding training samples by matching remote sensing data with the measured stations. Use the training sample set to train the artificial neural network, and further verify the inversion accuracy of the network using the measured data. According to the water body boundary range, process the daily VIIRS multispectral reflectance data covering the sea area, use the remote sensing reflectance of the 410, 443, 486, 551, 671, 745, and 862 nm bands as the input of the artificial neural network, establish the numerical relationship between the remote sensing reflectance and the sea surface salinity, and output the daily sea surface salinity (PSU), and perform monthly mean synthesis to obtain the monthly mean sea surface salinity distribution.

[0110] In one embodiment of the present invention, as Figure 6 shown, the generation of the thematic maps of the total nitrogen, total phosphorus and chlorophyll a concentration distributions includes the following steps:

[0111] First, preprocess the hyperspectral data and multispectral data, including geometric correction, radiometric calibration, and atmospheric correction;

[0112] Secondly, select spectral bands according to the normalized water body index model and calculate the normalized water body index NDWI to extract the offshore water body boundary information;

[0113] Then, conduct correlation analysis by combining the historical data of water quality parameters and the water body spectral data, select the sensitive bands of total nitrogen, total phosphorus and chlorophyll a, and establish empirical and semi-empirical models of the best bands or various band combinations;

[0114] Finally, use the established inversion model combined with the preprocessed remote sensing image to perform band calculation inversion to obtain the concentration gradient change results of total nitrogen, total phosphorus and chlorophyll a, and respectively form the corresponding concentration distribution thematic maps.

[0115] The above steps are further described as follows:

[0116] (1) Data preprocessing:

[0117] Geometric correction (ENVI): Prepare reference data, select the correction method, collect control points, calculate the correction model using the least squares method and generate the corrected image using the resampling method; Radiometric calibration (ENVI): Obtain the calibration coefficients, convert the DN values using the tool, and generate the calibrated image in the required output format; Atmospheric correction (ENVI): Prepare relevant information, perform FLAASH correction, check the spectral curve for evaluation and adjust the parameters to ensure accuracy; Geometric precise correction (ArcGIS): Ensure that the projection and other information are correct, select the correction method, collect control points and check the residuals, and perform the operation to generate a high-precision image.

[0118] (2) Water body extraction: Load the image in ENVI to calculate the Normalized Difference Water Index (NDWI), use the decision tree tool to distinguish water bodies from non-water bodies, export the binarized result in vector format and then convert it to shpfile format, and import and clip the cropping area data in ArcGIS to obtain water body vector data.

[0119] (3) Model establishment: Conduct spatial registration in ArcGIS, extract the corresponding remote sensing data of the sampling points in ENVI, integrate the data and then import it into SPSS, select an appropriate regression method to construct a linear regression model, and analyze the model parameters. Then, based on the analysis results of the sensitive bands, establish respective inversion models through the point data of total nitrogen, total phosphorus, and chlorophyll a to obtain the concentration distributions of total nitrogen, total phosphorus, and chlorophyll a.

[0120] S5: According to water temperature, salinity, total nitrogen, total phosphorus, chlorophyll a, and historical monitoring data, combined with the changing trends of key environmental parameters, introduce a spatio-temporal graph convolutional network model to simulate the spreading trend of red tides, set risk thresholds, and conduct risk prediction of red tide occurrence through the red tide occurrence risk prediction model.

[0121] In the embodiment of the present invention, the red tide risk prediction includes the following steps:

[0122] Collect data including water temperature, salinity, total nitrogen, total phosphorus, chlorophyll a, and historical monitoring data. Among them, the historical data includes records of past red tide events and historical environmental data, such as temperature, salinity, wind speed, etc.

[0123] In addition, the collected data can also be subjected to data cleaning and data standardization. Remove outliers and missing values through data cleaning, and perform necessary interpolation processing. Standardize the data from different sources through data standardization to ensure the consistency and comparability of the data.

[0124] Construct a red tide occurrence risk prediction model, which specifically includes the following steps:

[0125] Construct a spatio-temporal graph structure: Divide the ocean area into several grids, and each grid serves as a node of the spatio-temporal graph; construct the edges of the graph according to geographical adjacency relationships and ocean current directions to form a spatio-temporal graph structure; extract environmental parameters and time series data for each node as node features;

[0126] Construct a spatio-temporal graph convolutional network model:

[0127] Use the spatio-temporal graph convolutional network to capture spatial features. For each node i, its updated feature can be represented by the following formula:

[0128]

[0129] Among them, denotes the feature representation of node i at the l-th layer, N(i) denotes the set of neighbors of node i, and c ij is the normalization constant, W (l) and b (l) are the weight matrix and bias term of the l-th layer respectively, and σ is the activation function;

[0130] The temporal convolutional network is adopted to capture the temporal features in the time series data. For each time step t, its output can be expressed by the following formula:

[0131] O t = TCN(X t )

[0132] where X t denotes the input feature at time step t, and TCN denotes the operation of the temporal convolutional network;

[0133] The attention mechanism is introduced to weight the key environmental factors, and its formula can be expressed as:

[0134]

[0135] where Qi, Ki, and Vi denote the query, key, and matrix value respectively, dk denotes the dimension of the key, and the softmax function is used to calculate the attention weights;

[0136] To better fuse the spatial and temporal features, the outputs of the graph convolutional layer and the temporal convolutional layer are fused. Among them, the output of the graph convolutional layer is denoted by h graph and the output of the temporal convolutional layer is denoted by h time . Then the fused node feature h fusion is calculated by the following formula:

[0137] h fusion = αh graph + (1 - α)h time

[0138] where α is the weight parameter, which is a learnable weight parameter and its optimal value is determined through training to balance the contributions of the spatial and temporal features;

[0139] Red tide occurrence risk prediction:

[0140] The fully connected layer is used to map the fused node features to the red tide occurrence risk prediction model, and the formula is

[0141] P = α(W fc h fusion + b fc )

[0142] where P denotes the predicted risk value, W fc and bfc respectively represent the weights and bias terms of the fully connected layer, and σ is the activation function;

[0143] Set the risk threshold as y. When the predicted risk value P exceeds the threshold y, it is considered that there is a risk of red tide occurrence in this area.

[0144] In an embodiment of the present invention, in order to more finely evaluate the risk level, the probability interval can be divided into multiple levels. Specifically, the predicted risk value P is divided into the following three intervals to represent the red tide occurrence risk level:

[0145] When 0 ≤ P < 0.3, the risk level is considered low;

[0146] When 0.3 ≤ P < 0.6, the risk level is considered medium;

[0147] When 0.6 ≤ P ≤ 1, the risk level is considered high.

[0148] S6: Update the red tide occurrence risk prediction model based on buoy monitoring and on-site monitoring data.

[0149] In the embodiment of the present invention, the update of the red tide occurrence risk prediction model includes the following steps:

[0150] Divide the buoy monitoring and on-site monitoring data into a training set, a validation set, and a test set;

[0151] Use the cross-entropy loss function to measure the difference between the prediction result and the true label. The formula is:

[0152]

[0153] In the formula, N represents the number of samples, y i represents the true label of sample i, where the presence of red tide is 1 and the absence is 0, and P i represents the predicted risk probability of sample i;

[0154] Use the Adam optimizer to update the parameters of the red tide occurrence risk prediction model to minimize the loss function. The formula for updating the parameters is:

[0155] θ t+1 = θ t - η t m t

[0156] In the formula, θ t represents the parameter at time t, η t represents the learning rate, and m t represents the update amount based on momentum.

[0157] During the training process, the training set data is input into the red tide occurrence risk prediction model, the cross-entropy loss function is calculated, and the parameters of the red tide occurrence risk prediction model are updated through the Adam optimizer. This process is continuously repeated until the performance of the red tide occurrence risk prediction model on the validation set no longer improves.

[0158] In addition, according to the predicted red tide occurrence location and intensity, combined with environmental factors such as ocean currents and wind speeds, the propagation path of the red tide can be simulated, and the simulated red tide diffusion situation can be displayed in the form of animations or static images.

[0159] Specifically, the prediction method of the red tide path is as follows:

[0160] Data collection: First, collect data on environmental factors such as the red tide occurrence location, intensity, ocean currents, and wind speeds.

[0161] Model construction: Simulate the red tide propagation path through a particle tracking model.

[0162] Parameter setting: According to the collected data, set the parameters of the particle tracking model, such as the initial position, velocity, diffusion coefficient, etc. of the particles. These parameters need to be adjusted according to the actual situation to ensure the accuracy of the simulation results.

[0163] Simulation run: Run the set particle tracking model to simulate the red tide propagation path. During the simulation process, the impacts of environmental factors such as ocean currents and wind speeds on the red tide propagation need to be considered.

[0164] In the particle tracking model, the red tide organisms are regarded as particles, and the displacement of the particles is calculated by the superposition of the advection displacement and the diffusion displacement. Among them, the advection displacement refers to the displacement of the particles moving with the ocean current, and the diffusion displacement refers to the displacement of the particles generated due to the diffusion effect. The advection displacement formula is:

[0165] ΔX_adv = u_c * Δt

[0166] In the formula, ΔX_adv represents the advection displacement, u_c represents the ocean current velocity, and Δt represents the time interval.

[0167] The diffusion displacement formula is:

[0168]

[0169] In the formula, Δ X _diff represents the diffusion displacement, D represents the diffusion coefficient, and rand() represents the random number function, which is used to simulate the randomness of diffusion.

[0170] Therefore, the total displacement of the particles is: Δ X _adv + Δ X _diff.

[0171] The velocity of the particles can be calculated by the superposition of the ocean current velocity and the wind speed. However, it should be noted that the influence of the wind speed on the spread of the red tide is usually achieved by changing the flow of the sea surface layer. Therefore, the influence of the wind speed on the ocean current may need to be considered in actual calculations. Among them, the particle velocity formula is:

[0172] v_p = u_c + u_w * α

[0173] In the formula, v_p represents the velocity of the particles, u_c represents the ocean current velocity, u_w represents the wind speed, and α represents the coefficient of the influence of the wind speed on the ocean current, which needs to be set according to the actual situation.

[0174] According to the ocean red tide early warning method based on remote sensing inversion and spatio-temporal graph convolutional network of the embodiments of the present invention, by processing the collected hyperspectral and multispectral data, the integrity, accuracy, consistency, reliability, and timeliness of the data are ensured. The spatio-temporal graph neural network is used to construct a red tide occurrence risk prediction model based on the processed hyperspectral data and multispectral data, and through the fusion of satellite remote sensing, buoy monitoring, and historical data, the early warning of the red tide occurrence is realized. The attention mechanism is introduced to pay attention to the change trend of key environmental factors, and the prediction accuracy is improved.

[0175] In the description of the present invention, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. The meaning of "plurality" is two or more, unless otherwise specifically defined.

[0176] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed", etc. should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the internal connection of two components or the interaction relationship between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0177] In the present invention, unless otherwise clearly defined and limited, the first feature being "on" or "under" the second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, the first feature being "above", "over" and "on top of" the second feature may mean that the first feature is directly above or obliquely above the second feature, or simply indicates that the horizontal height of the first feature is higher than that of the second feature. The first feature being "under", "below" and "beneath" the second feature may mean that the first feature is directly below or obliquely below the second feature, or simply indicates that the horizontal height of the first feature is less than that of the second feature.

[0178] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not have to be directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0179] Any process or method description shown in the flowchart or described in other ways herein can be understood to represent a module, segment or part of code including one or more executable instructions for implementing a specific logical function or process, and the scope of the preferred embodiments of the present invention includes additional implementations, where the functions can be executed in a way that is not in the order shown or discussed, including in a substantially simultaneous manner or in the reverse order according to the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0180] The logic and / or steps represented in the flowchart or otherwise described herein can, for example, be considered as a definable sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in conjunction with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection portion with one or more wirings (electronic device), a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or other appropriate processing as necessary, and then storing it in a computer memory.

[0181] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0182] Those of ordinary skill in the art of this technology can understand that all or part of the steps carried by the method of the above embodiments can be completed by a program instructing relevant hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiments.

[0183] In addition, each functional unit in various embodiments of the present invention may be integrated into one processing module, may exist separately as individual units physically, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. When the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0184] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A marine red tide early warning method based on remote sensing inversion and spatiotemporal graph convolutional network, characterized in that: The following steps are involved: Collect hyperspectral and multispectral data; Performing data quality checks on the collected hyperspectral data and multispectral data respectively, and if the data quality checks fail, recollecting the hyperspectral data and multispectral data until the data quality checks pass; Preprocessing the hyperspectral data and the multispectral data after the quality inspection to obtain a hyperspectral result image and a digital orthophoto respectively; Analyze the hyperspectral image and the digital orthophoto to obtain remote sensing inversion data of water temperature, salinity, total nitrogen, total phosphorus and chlorophyll a, and generate thematic maps of gradient distribution of each indicator; Based on water temperature, salinity, total nitrogen and phosphorus, chlorophyll a and historical monitoring data, combined with the changing trends of key environmental parameters, a spatiotemporal graph convolutional network model was introduced to simulate the spread of red tides, set risk thresholds, and predict the risk of red tides through the red tide occurrence risk prediction model; Update the red tide risk prediction model based on buoy monitoring and on-site monitoring data.

2. The marine red tide early warning method based on remote sensing inversion and spatiotemporal graph convolutional network according to claim 1 is characterized in that: The preprocessing of hyperspectral data includes the following steps: Combine the bands of the acquired hyperspectral images; The absolute radiation calibration coefficient is used to perform radiation calibration, and the brightness gray value of the hyperspectral image is converted into an absolute radiation brightness value, so as to eliminate the radiation deviation of the sensor itself and ensure the radiation consistency of the image; The atmospheric correction algorithm is used to perform atmospheric correction, converting the apparent radiance data into surface reflectance data to remove interference factors; Collecting control points and checking residuals, and performing geometric correction on the hyperspectral image after atmospheric correction; Orthorectification is performed on the hyperspectral image, and the geometric distortion caused by the terrain is corrected using the digital elevation model DEM, so as to eliminate the projection difference caused by the terrain undulation and the tilt of the image; Mosaicing the multiple hyperspectral images, performing a color uniformity operation, cropping the hyperspectral images, and extracting the hyperspectral image data of the target area; The contrast and clarity of hyperspectral satellite images are enhanced through image processing technology, and finally a hyperspectral result image is obtained.

3. The marine red tide early warning method based on remote sensing inversion and spatiotemporal graph convolutional network according to claim 2 is characterized in that: Multispectral data includes panchromatic images and multispectral images. The preprocessing includes the following steps: The panchromatic image is calibrated by radiation to obtain surface reflectance data, and orthorectified by combining DEM data and RPC file to obtain a panchromatic orthophoto; the multispectral image is calibrated by radiation, atmospherically corrected, and orthorectified to obtain a multispectral orthophoto; The obtained panchromatic orthophoto is fused with the multispectral orthophoto, and mosaic processing is performed on the fused images to achieve seamless splicing of adjacent images; The mosaicked images are geometrically corrected and finally digital orthophotos are output.

4. The marine red tide early warning method based on remote sensing inversion and spatiotemporal graph convolutional network according to claim 3 is characterized in that: The gradient distribution thematic maps of each indicator include the sea surface temperature distribution thematic map, the sea surface salinity distribution thematic map, the total nitrogen, total phosphorus and chlorophyll a concentration distribution thematic map. The generation of the sea surface temperature distribution thematic map includes the following steps: First, the multispectral data are preprocessed, including radiation calibration and atmospheric correction; Secondly, the coverage image is obtained based on the calculation of the normalized vegetation index (NDVI), and finally the surface emissivity result is obtained; Then, the blackbody radiation brightness and surface temperature are calculated by using the radiation brightness image in the near-infrared band, combined with the surface emissivity and atmospheric profile parameters. The atmospheric profile parameters include atmospheric transmittance, downgoing radiation brightness, and upgoing radiation brightness. Finally, according to the normalized water index model, the spectral bands are selected and the normalized water index is calculated to extract the boundary information of the nearshore water body, form the sea surface temperature gradient change and form the sea surface temperature distribution thematic map.

5. The marine red tide early warning method based on remote sensing inversion and spatiotemporal graph convolutional network according to claim 4 is characterized in that: The generation of the sea surface salinity distribution thematic map includes the following steps: First, the multispectral data are preprocessed, including radiation calibration and atmospheric correction; Then, using the historical data of sea surface salinity stations as training samples, the characteristic bands with strong correlation were selected, and the nonlinear relationship between remote sensing reflectivity and sea surface salinity values ​​was established through artificial neural networks to construct an inversion model; Finally, according to the boundary range of the water body, the remote sensing reflectance data of the sensitive characteristic bands are used as the input of the artificial neural network model, and the band calculation and inversion are performed to obtain the results of the sea surface salinity gradient changes, forming a thematic map of sea surface salinity distribution.

6. The marine red tide early warning method based on remote sensing inversion and spatiotemporal graph convolutional network according to claim 5 is characterized in that: The generation of thematic maps of total nitrogen, total phosphorus and chlorophyll a concentration distribution includes the following steps: First, the hyperspectral data and multispectral data are preprocessed, including geometric correction, radiation calibration, and atmospheric correction; Secondly, according to the normalized water index model, the spectral bands are selected and the normalized water index NDWI is calculated to extract the boundary information of the offshore water body; Then, the historical data of water quality parameters and water spectrum data were combined for correlation analysis, and the sensitive bands of total nitrogen, total phosphorus and chlorophyll a were selected to establish the empirical or semi-empirical model of the best band or various band combinations; Finally, the established inversion model was combined with the preprocessed remote sensing images to perform band calculation inversion to obtain the concentration gradient changes of total nitrogen, total phosphorus and chlorophyll a, and form corresponding concentration distribution thematic maps respectively.

7. The marine red tide early warning method based on remote sensing inversion and spatiotemporal graph convolutional network according to claim 6 is characterized in that: Red tide risk prediction includes the following steps: Collect water temperature, salinity, total nitrogen, total phosphorus, chlorophyll a and historical monitoring data; The construction of the red tide risk prediction model includes the following steps: Constructing a spatiotemporal graph structure: Divide the ocean area into several grids, with each grid as a node in the spatiotemporal graph; construct the edges of the graph based on the geographical proximity and the direction of the ocean currents to form a spatiotemporal graph structure; extract environmental parameters and time series data for each node as node features; Constructing a spatiotemporal graph convolutional network model: The spatiotemporal graph convolutional network is used to capture spatial features. For each node i, its updated features can be expressed by the following formula: in, represents the feature representation of node i at layer l, N(i) represents the neighbor set of node i, c ij is the normalization constant, W (l) and b (l) are the weight matrix and bias term of the lth layer, σ is the activation function; A temporal convolutional network is used to capture the temporal features in time series data. For each time step t, its output can be expressed by the following formula: O t =TCN(X t ) Where, X t represents the input features of time step t, and TCN represents the operation of the temporal convolutional network; The attention mechanism is introduced to perform weighted processing on key environmental factors, and its formula can be expressed as: Where Qi, Ki, and Vi represent query, key, and matrix value, respectively, dk represents the dimension of the key, and the softmax function is used to calculate the attention weight; In order to better integrate spatial and temporal features, the outputs of the graph convolution layer and the temporal convolution layer are fused, where the output of the graph convolution layer is represented by h graph Indicates that the output of the temporal convolutional layer is represented by h time Represents that the fused node feature h fusion Calculated by the following formula: h fusion =ah graph +(1-a)h time In the formula, α is a weight parameter, which is a learnable weight parameter. Its optimal value is determined through training to balance the contribution of spatial features and temporal features. Red tide risk prediction: Use the fully connected layer to map the fused node features to the red tide risk prediction model. The formula is: P=α(W fc h fusion +b fc ) Where P represents the predicted risk value, W fc and b fc Represent the weight and bias term of the fully connected layer respectively, and σ is the activation function; The risk threshold is set to y. When the predicted risk value P exceeds the threshold y, it is considered that there is a risk of red tide occurring in the area.

8. The marine red tide early warning method based on remote sensing inversion and spatiotemporal graph convolutional network according to claim 7 is characterized in that: The predicted risk value P is divided into the following three intervals to indicate the risk level of red tide occurrence: When 0≤P<0.3, the risk level is considered low; When 0.3≤P<0.6, the risk level is considered to be medium; When 0.6≤P≤1, the risk level is considered high.

9. The marine red tide early warning method based on remote sensing inversion and spatiotemporal graph convolutional network according to claim 8 is characterized in that: The update of the red tide risk prediction model includes the following steps: The buoy monitoring and field monitoring data are divided into training set, validation set and test set; The cross entropy loss function is used to measure the difference between the predicted result and the true label. The formula is: In the formula, N represents the number of samples, y i represents the true label of sample i, where the presence of red tide is 1 and the absence is 0, P i represents the predicted risk probability of sample i; The Adam optimizer is used to update the parameters of the red tide risk prediction model to minimize the loss function. The formula for updating the parameters is: i t+1 =θ t -or t m t In the formula, θ t represents the parameter at time t, η t represents the learning rate, m t represents the momentum-based update amount.

Images