High-precision sea surface chlorophyll concentration inversion method applying cross attention mechanism and space-spectrum fusion

Abstract

The invention provides a high-precision sea surface chlorophyll concentration inversion method applying a cross attention mechanism and space-spectrum fusion, and aims to solve the problem of balancing the spatial resolution and the time resolution of the sea surface chlorophyll concentration in satellite remote sensing. According to the method, a cross attention mechanism is introduced through a space-spectrum fusion technology, space and spectrum information with different resolutions can be fully mined and effectively fused, and sea surface reflectivity data with high spatial resolution and high time resolution can be made; and the precision and robustness of chlorophyll concentration inversion are improved by using a random forest fitting inversion method. The methods overcome the application limitation of a traditional sea surface chlorophyll concentration manufacturing method in a complex marine environment. According to the method, the spatial resolution and the time resolution of the sea surface chlorophyll concentration can be remarkably improved, and the method has relatively high adaptability and operability.

Description

Technical Field

[0001] The present invention belongs to the field of ocean color remote sensing, and in particular relates to a high-precision sea surface chlorophyll concentration inversion method using a cross-attention mechanism and space-spectrum fusion. Background Art

[0002] Due to the physical and technical limitations of sensors, a trade-off between spatial resolution, temporal resolution, and spectral resolution is necessary in satellite remote sensing technology. Improving spatial resolution usually means that the sensor needs to capture more detailed information, which will reduce its coverage, thereby increasing the time interval required to cover the same area twice, resulting in reduced temporal resolution. At the same time, for sea surface reflectance data, obtaining high-resolution data in multiple spectral bands simultaneously requires complex sensing technology and large data processing capabilities, which further limits the possibility of maintaining high spectral resolution while maintaining high spatial or temporal resolution. Therefore, according to different application requirements and mission characteristics, specific technical and design choices must be made between these three, and it is impossible to take into account high spatial resolution, high temporal resolution, and high spectral resolution. The inversion of high-precision sea surface water color products is often based on sea surface reflectance data with high spectral resolution. Its spatial resolution and temporal resolution are also limited by sea surface reflectance data, so high-quality water color product images are scarce. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a high-precision sea surface chlorophyll concentration inversion method using a cross-attention mechanism and spatial spectrum fusion. The specific technical solution is as follows:

[0004] A high-precision sea surface chlorophyll concentration inversion method using a cross-attention mechanism and spatial-spectral fusion includes the following steps:

[0005] Step 1: To obtain the spatial resolution SR of the predicted location p2 at the predicted time t3 H , time resolution is TR H The chlorophyll concentration data is constructed with a spatial resolution of SR H , time resolution is TR L The multi-band remote sensing image dataset is X1, and the spatial resolution is SR L , time resolution is TR H The multi-band remote sensing image dataset is X2; among them, the spatial resolution SR H >SR L , time resolution TR L <TR H ;

[0006] Select from X1 a multi-band remote sensing image X11 of date t1 of location p1 that is adjacent to or contains the location p2 to be predicted and is used for training the spatial-spectral fusion model;

[0007] Select from X1 a multi-band remote sensing image X10 of a location p1 that is adjacent to or contains the location p2 to be predicted and at date t0 that is close to t1 for training of the spatial-spectral fusion model;

[0008] Select from X2 a multi-band remote sensing image X21 of date t1 of location p1 that is adjacent to or contains the location p2 to be predicted and is used for training the spatial-spectral fusion model;

[0009] Select from X2 a multi-band remote sensing image X20 of a location p1 that is adjacent to or contains the location p2 to be predicted and at a date t0 that is adjacent to t1 for training of a spatial-spectral fusion model;

[0010] Select a multi-band remote sensing image X23 located at the location p2 to be predicted at the time t3 to be predicted from X2;

[0011] Select from X2 a multi-band remote sensing image X22 located at the location p2 to be predicted at t2 close to the time t3 to be predicted;

[0012] Select from X1 a multi-band remote sensing image X12 located at the location p2 to be predicted at t2 close to the time t3 to be predicted;

[0013] Step 2: Take the sea surface reflectance data images corresponding to the multi-band remote sensing images X11, X10, X21 and X20 to create a training data set for the spatial-spectral fusion model; build a spatial-spectral fusion model based on the cross-attention mechanism and complete the training of the spatial-spectral fusion model;

[0014] Step 3: Take the sea surface reflectance data images corresponding to the multi-band remote sensing images X12, X22 and X23, make a prediction data set, input the spatial spectrum fusion model trained in step 2, and predict the spatial resolution of time t3 to be SR H , time resolution is TR H Multi-band remote sensing image X3;

[0015] Step 4: Select several sea surface reflectance data images from the multi-band remote sensing image dataset X1, and the known chlorophyll concentrations in the same time and area as the selected sea surface reflectance data images, perform longitude and latitude matching, and obtain a training data set for the chlorophyll concentration inversion model;

[0016] Step 5: Construct a chlorophyll concentration inversion model based on random forest fitting, and use the training data set of the chlorophyll concentration inversion model in step 4 to train the chlorophyll concentration inversion model to obtain a trained chlorophyll concentration inversion model;

[0017] Step 6: The spatial resolution of time t3 obtained in step 3 is SR H , time resolution is TR H The multi-band remote sensing image X3 is input into the trained chlorophyll concentration inversion model obtained in step 5. The chlorophyll concentration inversion model outputs the spatial resolution of the specified area p2 at time t3 as SR H , time resolution is TR H chlorophyll concentration.

[0018] Furthermore, the bands selected by the multi-band remote sensing images X11, X10, X21, X20 and X12, X22, X23 need to select blue, green, red, and near-infrared bands, and the multi-band remote sensing images X11, X10, X12 and the multi-band remote sensing images X21, X2 0、 The wavelength ranges of X22 and X23 in corresponding bands are as close as possible.

[0019] Furthermore, the location p2 to be predicted is included in the location p1.

[0020] Furthermore, t0, t1, t2, and t3 are close to each other; at the same time, the time interval between t0 and t1 is the same as the time interval between t2 and t3.

[0021] Furthermore, the step 2 includes the following sub-steps:

[0022] S2.1: For the sea surface reflectance data images corresponding to the multi-band remote sensing images X11, X10, X21 and X20, perform spatial segmentation and land masking on location p1 to remove all land areas; then perform interpolation filling of missing values ​​and outliers; then upsample the filled X21 and X20 to the same spatial resolution as the multi-band remote sensing images X11 and X10;

[0023] S2.2: Cut the four sea surface reflectance data images with the same spatial resolution obtained after processing in step S2.1 into three-dimensional image blocks of the same size and quantity, use the sea surface reflectance data images corresponding to X10, X20 and X21 as the input of the spatial-spectral fusion model, and use the sea surface reflectance data image corresponding to X11 as the output of the spatial-spectral fusion model to produce a training data set;

[0024] S2.3: Construct a spatial-spectral fusion model based on convolutional neural network, implement the cross-attention mechanism, use the training data set of S2.2 to train the spatial-spectral fusion model, and obtain the trained spatial-spectral fusion model.

[0025] Furthermore, the step three includes the following sub-steps:

[0026] S3.1: Perform spatial cutting and land masking on the sea surface reflectance data images corresponding to X12, X22 and X23 for the predicted location p2 to remove all land areas; then perform interpolation filling of missing values ​​and outliers; then upsample the filled sea surface reflectance data images corresponding to X22 and X23 to the same spatial resolution as the multi-band remote sensing image X12;

[0027] S3.2: Cut the three sea surface reflectance data images with the same spatial resolution obtained after processing in S3.1 into three-dimensional image blocks of the same size as the training data set for the spatial-spectral fusion model prepared in step 2, and use the sea surface reflectance data images corresponding to X12, X22 and X23 as the input of the spatial-spectral fusion model trained in step 2 to prepare a prediction data set;

[0028] S3.3: Input the prediction data set prepared in S3.2 into the spatial-spectral fusion model trained in step 2, and obtain the spatial resolution of the prediction time t3 as SR H , time resolution is TR H Multi-band remote sensing image X3.

[0029] Furthermore, when performing longitude and latitude matching in step 4, when the known chlorophyll concentration in the same time and area is a grid image, and the spatial resolution of the grid image is different from the spatial resolution of the multi-band remote sensing image dataset X1, it is necessary to downsample the image with the higher spatial resolution of the two images to ensure that the longitude and latitude matching of the two is performed at the same spatial resolution.

[0030] The beneficial effects of the present invention are as follows:

[0031] The high-precision sea surface chlorophyll concentration inversion method using a cross-attention mechanism and spatial-spectral fusion provided by the present invention uses machine learning for fitting, effectively overcoming the nonlinear relationship problem encountered by traditional methods when dealing with the spatiotemporal variations of chlorophyll concentration in complex water bodies, learning complex nonlinear patterns from a large amount of data, and establishing a more accurate prediction model, thereby improving the accuracy and robustness of chlorophyll concentration prediction; the images generated by this method can have higher spatial resolution and temporal resolution at the same time, which can make up for the limitations of a single sensor. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is an overall flow chart of a high-precision sea surface chlorophyll concentration inversion method using a cross-attention mechanism and spatial-spectral fusion in an embodiment of the present invention.

[0033] Figure 2 This is a partial flow chart of spatial-spectrum fusion according to an embodiment of the present invention.

[0034] Figure 3 This is a flow chart of the inversion fitting part of an embodiment of the present invention.

[0035] Figure 4 Schematic diagram of chlorophyll concentration after adopting the present invention and the comparative method in this embodiment; wherein, Figure (a) shows the 10-meter resolution chlorophyll concentration inversion result based on NN after adopting the present invention; Figure (b) shows the 300-meter resolution chlorophyll concentration inversion result based on NN in the WFR dataset; Figure (c) shows the 10-meter resolution chlorophyll concentration inversion result based on OC4M after adopting the present invention; Figure (d) shows the 300-meter resolution chlorophyll concentration inversion result based on OC4ME in the WFR dataset. DETAILED DESCRIPTION

[0036] The present invention will be described in detail below based on the accompanying drawings and preferred embodiments, and the purpose and effects of the present invention will become more clear. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0037] Technical term explanation:

[0038] 6S model: The 6S (Second Simulation of the Satellite Sigal in the Solar Spectrum) model is a widely used atmospheric radiation transfer model, which is used to simulate the propagation process of solar radiation in the atmosphere, especially in satellite remote sensing applications.

[0039] like Figure 1 As shown, the high-precision sea surface chlorophyll concentration inversion method using the cross-attention mechanism and spatial spectrum fusion of the present invention includes the following steps:

[0040] Step 1: To obtain the spatial resolution SR of the predicted location p2 at the predicted time t3 H , time resolution is TR H The chlorophyll concentration data is constructed with a spatial resolution of SR H , time resolution is TR L The multi-band remote sensing image dataset is X1, and the spatial resolution is SR L , time resolution is TR H The multi-band remote sensing image dataset is X2; among them, the spatial resolution SR H >SR L , time resolution TR L <TR H .

[0041] Select from X1 a multi-band remote sensing image X11 of date t1 of location p1 that is adjacent to or contains the location p2 to be predicted and is used for training the spatial-spectral fusion model;

[0042] Select from X1 a multi-band remote sensing image X10 of a location p1 that is adjacent to or contains the location p2 to be predicted and at date t0 that is close to t1 for training of the spatial-spectral fusion model;

[0043] Select from X2 a multi-band remote sensing image X21 of date t1 of location p1 that is adjacent to or contains the location p2 to be predicted and is used for training the spatial-spectral fusion model;

[0044] Select from X2 a multi-band remote sensing image X20 of a location p1 that is adjacent to or contains the location p2 to be predicted and at a date t0 that is adjacent to t1 for training of a spatial-spectral fusion model;

[0045] Select a multi-band remote sensing image X23 located at the location p2 to be predicted at the time t3 to be predicted from X2;

[0046] Select from X2 a multi-band remote sensing image X22 located at the location p2 to be predicted at t2 close to the time t3 to be predicted;

[0047] A multi-band remote sensing image X12 located at the location p2 to be predicted at t2 close to the time t3 to be predicted is selected from X1.

[0048] The bands selected by the multi-band remote sensing images X11, X10, X21, X20 and X12, X22, and X23 here need to select blue, green, red, and near-infrared bands, and the multi-band remote sensing images X11, X10, X12 and the multi-band remote sensing images X21, X2 0、 The wavelength ranges of X22 and X23 in corresponding bands are as close as possible.

[0049] The locations p1 and p2 are in an inclusion relationship, and preferably the location p2 to be predicted is included in the known location p1.

[0050] Because the chlorophyll concentration varies greatly at different times, t0, t1, t2, and t3 are as close as possible; at the same time, the time interval between t0 and t1 is as close as possible to the time interval between t2 and t3.

[0051] In this embodiment, the selected reference date t0 is December 3, the adjacent prediction date t1 is December 8, the designated sea area p1 is the South China Sea with a latitude from 20.807°N to 21.363° and a longitude from 109.241°E to 109.651°E, the multi-band remote sensing images X11 and X10 are multi-band Sentinel-2L1C image pairs with a spatial resolution of 10 meters of the designated sea area on December 8 and December 3, respectively, and the multi-band remote sensing images X21 and X20 are multi-band image pairs with a resolution of 300 meters from Sentinel-3 of the designated sea area from OLCI WFR on December 8 and December 3, respectively. Among them, the corresponding file processing of Sentinel-2 is carried out on the GEE platform, and the Sentinel-3 data files are: S3A_OL_2_WFR____20201203T025356_20201203T025656_20201204T104820_0179_065_360_2520_MAR_O_NT_002.SEN3 and S3A_OL_2_WFR____20201208T022401_20201208T022701_20201209T130731_0179_066_046_2520_MAR_O_NT_002.SEN3.

[0052] In this embodiment, the bands selected for the Sentinel-2 sea surface reflectance data image are: B2, B3, B4, B5, B6, and B8; the bands selected for the Sentinel-3 surface reflectance data in the WFR dataset are: Oa04, Oa06, Oa08, Oa11, Oa12, and Oa17.

[0053] In this embodiment, in order to test the model output, t2=t0, t3=t1, p2 is a part of the area contained in p1, X12 and X2 2、 X23 can be respectively from X10 and X2 0、 X21 remote sensing image is captured.

[0054] Step 2: Take the sea surface reflectance data images corresponding to the multi-band remote sensing images X11, X10, X21 and X20 at t1 and t0, and make a training data set for the spatial-spectral fusion model; build a spatial-spectral fusion model based on the cross-attention mechanism and complete the training of the spatial-spectral fusion model. Figure 2 As shown, step 2 specifically includes the following sub-steps:

[0055] S2.1: For the sea surface reflectance data images corresponding to the multi-band remote sensing images X11, X10, X21 and X20 at t1 and t0, perform spatial segmentation and land masking on location p1 to remove all land areas; then perform interpolation filling of missing values ​​and outliers; then upsample the filled X21 and X20 to the same spatial resolution as the multi-band remote sensing images X11 and X10;

[0056] S2.2: Cut the four sea surface reflectance data images with the same spatial resolution obtained after processing in step S2.1 into three-dimensional image blocks of the same size and quantity, use the sea surface reflectance data images corresponding to X10, X20 and X21 as the input of the spatial-spectral fusion model, and use the sea surface reflectance data image corresponding to X11 as the output of the spatial-spectral fusion model to produce a training data set;

[0057] S2.3: Construct a spatial-spectral fusion model based on a convolutional neural network, implement a cross-attention mechanism, and use the training data set of S2.2 to train the spatial-spectral fusion model to obtain a trained spatial-spectral fusion model; the spatial-spectral fusion model at this time can use the convolution method to simultaneously examine the temporal changes in the sea surface reflectivity of a specified area on adjacent dates and the channel differences of remote sensing images from different sensors in a specified area and time;

[0058] Based on the image fusion model of convolutional neural network, the following method is used to examine the phase differences of the same area at different times:

[0059] F SPM =(W a *F H +b a )⊙σ(W b *F L +b b )

[0060] At the same time, the following methods are used to investigate the differences between different sensor channels:

[0061]

[0062] Among them, F SPM 、F SRM They represent the feature maps output by the SPM (Spatial Perception Module) spatial perception module and the SRM (Spectral Reconstruction Module) spectral perception module, respectively. H Represents high spatial resolution image (SR H ) corresponds to the feature map, F L Represents low spatial resolution images (SR L ) corresponds to the feature map, represents the feature map obtained by the SRM module for the i-th band, f concat is a cascade operation, ⊙ is element-by-element multiplication, σ represents the Sigmoid function, W represents the weight parameter, and b represents the bias parameter.

[0063] In this embodiment, due to the multi-band Sentinel-2L1C image selected, it is necessary to invert the multi-band Sentinel-2L1C image into a sea surface reflectivity data image through the atmospheric radiation transmission model 6S. The multi-band image of Sentinel-3 with a resolution of 300 meters of OLCI WFR is itself a sea surface reflectivity data image, and the land mask has been completed. Therefore, the multi-band image of Sentinel-3 with a resolution of 300 meters of OLCI WFR is directly spatially cut, and missing values ​​and outliers are interpolated and filled. The multi-band image of Sentinel-3 with a resolution of 300 meters is upsampled to obtain a sea surface reflectivity data image with a resolution of 10 meters. In this embodiment, the processing of Sentinel-2 images is performed on the GEE platform.

[0064] Step 3: Take the sea surface reflectance data images corresponding to the multi-band remote sensing images X12, X22 and X23, make a prediction data set, input the spatial spectrum fusion model trained in step 2, and predict the spatial resolution of time t3 to be SR H , time resolution is TR H Multi-band remote sensing image X3. Figure 2 As shown, step three specifically includes the following sub-steps:

[0065] S3.1: Perform spatial cutting and land masking on the sea surface reflectance data images corresponding to X12, X22 and X23 for the predicted location p2 to remove all land areas; then perform interpolation filling of missing values ​​and outliers; then upsample the filled sea surface reflectance data images corresponding to X22 and X23 to the same spatial resolution as the multi-band remote sensing image X12;

[0066] S3.2: Cut the three sea surface reflectance data images with the same spatial resolution obtained after processing in S3.1 into three-dimensional image blocks of the same size as the training data set for the spatial-spectral fusion model prepared in step 2, and use the sea surface reflectance data images corresponding to X12, X22 and X23 as the input of the spatial-spectral fusion model trained in step 2 to prepare a prediction data set;

[0067] S3.3: Input the prediction data set prepared in S3.2 into the spatial-spectral fusion model trained in step 2 to obtain the multi-band sea surface reflectance data image X3 at the prediction time t3. At this time, X3 has the same high spatial resolution SR as the X1 data set. H, and because the corresponding X3 can be predicted as long as X23 is available, X3 at this time also has the same high temporal resolution TR as X2 H .

[0068] Step 4: Select several sea surface reflectance data images from the multi-band remote sensing image dataset X1, and the known chlorophyll concentrations in the same time and area as the selected sea surface reflectance data images, perform longitude and latitude matching, and obtain a training data set for the chlorophyll concentration inversion model;

[0069] The selected sea surface reflectance data images from the multi-band remote sensing image dataset X1 should have the same band selection as the previous X10, X11 and X12, and the time and space selection should also be as close as possible.

[0070] When the known chlorophyll concentration in the same time and area is a grid image, and the spatial resolution of the grid image is different from the spatial resolution of the multi-band remote sensing image dataset X1, it is necessary to downsample the image with the higher spatial resolution to ensure that the longitude and latitude of the two are matched at the same spatial resolution.

[0071] In this embodiment, the selected sea surface reflectance data images from the multi-band remote sensing image dataset X1 are X10 and X11, and the known chlorophyll concentration of the same time and region is from the WFR dataset of Sentinel-3. Since the known chlorophyll concentration comes from Sentinel-3, which has a spatial resolution of 300m, and the spatial resolution of the Sentinel-2 sea surface reflectance data is 10m, the 10-meter spatial resolution of the Sentinel-2 sea surface reflectance data is first downsampled to a resolution of 300m, and then matched with the chlorophyll concentration inverted based on the OC4ME and NN methods in the WFR according to time, longitude and latitude, to obtain two inversion fitting data sets.

[0072] Step 5: Construct a chlorophyll concentration inversion model based on random forest fitting, and use the training data set of the chlorophyll concentration inversion model in step 4 to train the chlorophyll concentration inversion model to obtain a trained chlorophyll concentration inversion model.

[0073] In this embodiment, two inversion fitting data sets are used to train the chlorophyll concentration inversion model, respectively, to obtain two chlorophyll concentration inversion models.

[0074] Step 6: The spatial resolution of time t3 obtained in step 3 is SR H , time resolution is TR H The multi-band remote sensing image X3 is input into the trained chlorophyll concentration inversion model obtained in step 5. The chlorophyll concentration inversion model outputs the spatial resolution of the specified area p2 at time t3 as SRH , time resolution is TR H chlorophyll concentration.

[0075] The inversion fitting process of step 4, step 5 and step 6 is as follows Figure 3 shown.

[0076] The results obtained in this example are as follows Figure 4 As shown, (a) represents the chlorophyll concentration inverted with a spatial resolution of 10 m based on NN after adopting the method of the present invention, (b) represents the chlorophyll concentration inverted with a spatial resolution of 300 m based on NN in the WFR dataset, (c) represents the chlorophyll concentration inverted with a spatial resolution of 10 m based on OC4M after adopting the method of the present invention, and (d) represents the chlorophyll concentration inverted with a spatial resolution of 300 m based on OC4ME in the WFR dataset.

[0077] By comparing (a) with (b), and (c) with (d), it can be seen that the chlorophyll concentrations are highly consistent, and the image texture details of the chlorophyll concentration inverted by the method of the present invention are richer. By comparing (c) with (d), the data missing areas in Figure (d) are supplemented in Figure (c).

[0078] In the method of the present invention, as long as there is available sea surface reflectance data in a set area with high temporal resolution remote sensing image data X2 of a date close to the prediction date, the sea surface chlorophyll concentration of the set area on the prediction date with high spatial resolution can be inverted. Therefore, the sea surface chlorophyll concentration obtained by the method of the present invention has both a spatial resolution consistent with the X1 data set and a temporal resolution consistent with the X2 data set, thereby generating sea surface chlorophyll concentration grid data with both high spatial resolution and high temporal resolution.

[0079] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the invention.

Claims

1. A high-precision sea surface chlorophyll concentration inversion method using a cross-attention mechanism and spatial-spectral fusion, characterized in that: The steps include: Step 1: To obtain the spatial resolution SR of the predicted location p2 at the predicted time t3 H , time resolution is TR H The chlorophyll concentration data is constructed with a spatial resolution of SR H , time resolution is TR L The multi-band remote sensing image dataset is X1, and the spatial resolution is SR L , time resolution is TR H The multi-band remote sensing image dataset is X2; among them, the spatial resolution SR H >SR L , time resolution TR L <TR H ; Select from X1 a multi-band remote sensing image X11 of date t1 of location p1 that is adjacent to or contains the location p2 to be predicted and is used for training the spatial-spectral fusion model; Select from X1 a multi-band remote sensing image X10 of a location p1 that is adjacent to or contains the location p2 to be predicted and at date t0 that is close to t1 for training of the spatial-spectral fusion model; Select from X2 a multi-band remote sensing image X21 of date t1 of location p1 that is adjacent to or contains the location p2 to be predicted and is used for training the spatial-spectral fusion model; Select from X2 a multi-band remote sensing image X20 of a location p1 that is adjacent to or contains the location p2 to be predicted and at a date t0 that is adjacent to t1 for training of a spatial-spectral fusion model; Select a multi-band remote sensing image X23 located at the location p2 to be predicted at the time t3 to be predicted from X2; Select from X2 a multi-band remote sensing image X22 located at the location p2 to be predicted at t2 close to the time t3 to be predicted; Select from X1 a multi-band remote sensing image X12 located at the location p2 to be predicted at t2 close to the time t3 to be predicted; Step 2: Take the sea surface reflectance data images corresponding to the multi-band remote sensing images X11, X10, X21 and X20 to create a training data set for the spatial-spectral fusion model; build a spatial-spectral fusion model based on the cross-attention mechanism and complete the training of the spatial-spectral fusion model; Step 3: Take the sea surface reflectance data images corresponding to the multi-band remote sensing images X12, X22 and X23, make a prediction data set, input the spatial spectrum fusion model trained in step 2, and predict the spatial resolution of time t3 to be SR H , time resolution is TR H Multi-band remote sensing image X3; Step 4: Select several sea surface reflectance data images from the multi-band remote sensing image dataset X1, and the known chlorophyll concentrations in the same time and area as the selected sea surface reflectance data images, perform longitude and latitude matching, and obtain a training data set for the chlorophyll concentration inversion model; Step 5: Construct a chlorophyll concentration inversion model based on random forest fitting, and use the training data set of the chlorophyll concentration inversion model in step 4 to train the chlorophyll concentration inversion model to obtain a trained chlorophyll concentration inversion model; Step 6: Set the spatial resolution of time t3 obtained in step 3 to SR H , time resolution is TR H The multi-band remote sensing image X3 is input into the trained chlorophyll concentration inversion model obtained in step 5. The chlorophyll concentration inversion model outputs the spatial resolution of the specified area p2 at time t3 as SR H , time resolution is TR H chlorophyll concentration.

2. The high-precision sea surface chlorophyll concentration inversion method using cross-attention mechanism and spatial-spectral fusion according to claim 1 is characterized in that: The bands selected by the multi-band remote sensing images X11, X10, X21, X20 and X12, X22, X23 need to select blue, green, red, and near-infrared bands, and the multi-band remote sensing images X11, X10, X12 and the multi-band remote sensing images X21, X2 0、 The wavelength ranges of X22 and X23 in corresponding bands are as close as possible.

3. The high-precision sea surface chlorophyll concentration inversion method using cross-attention mechanism and spatial-spectral fusion according to claim 1 is characterized in that: The location p2 to be predicted is included in the location p1.

4. The high-precision sea surface chlorophyll concentration inversion method using cross-attention mechanism and spatial-spectral fusion according to claim 1 is characterized in that: t0, t1, t2, and t3 are close; at the same time, the time interval between t0 and t1 is the same as the time interval between t2 and t3.

5. The high-precision sea surface chlorophyll concentration inversion method using cross-attention mechanism and spatial-spectral fusion according to claim 1 is characterized in that: The step 2 includes the following sub-steps: S2.1: For the sea surface reflectance data images corresponding to the multi-band remote sensing images X11, X10, X21 and X20, perform spatial segmentation and land masking on location p1 to remove all land areas; then perform interpolation filling of missing values ​​and outliers; then upsample the filled X21 and X20 to the same spatial resolution as the multi-band remote sensing images X11 and X10; S2.2: Cut the four sea surface reflectance data images with the same spatial resolution obtained after processing in step S2.1 into three-dimensional image blocks of the same size and quantity, use the sea surface reflectance data images corresponding to X10, X20 and X21 as the input of the spatial-spectral fusion model, and use the sea surface reflectance data image corresponding to X11 as the output of the spatial-spectral fusion model to produce a training data set; S2.3: Construct a spatial-spectral fusion model based on convolutional neural network, implement the cross-attention mechanism, use the training data set of S2.2 to train the spatial-spectral fusion model, and obtain the trained spatial-spectral fusion model.

6. The high-precision sea surface chlorophyll concentration inversion method using cross-attention mechanism and spatial-spectral fusion according to claim 5 is characterized in that: The step three includes the following sub-steps: S3.1: Perform spatial cutting and land masking on the sea surface reflectance data images corresponding to X12, X22 and X23 for the predicted location p2 to remove all land areas; then perform interpolation filling of missing values ​​and outliers; then upsample the filled sea surface reflectance data images corresponding to X22 and X23 to the same spatial resolution as the multi-band remote sensing image X12; S3.2: Cut the three sea surface reflectance data images with the same spatial resolution obtained after processing in S3.1 into three-dimensional image blocks of the same size as the training data set for the spatial-spectral fusion model prepared in step 2, and use the sea surface reflectance data images corresponding to X12, X22 and X23 as the input of the spatial-spectral fusion model trained in step 2 to prepare a prediction data set; S3.3: Input the prediction data set prepared in S3.2 into the spatial-spectral fusion model trained in step 2, and obtain the spatial resolution of the prediction time t3 as SR H , time resolution is TR H Multi-band remote sensing image X3.

7. The high-precision sea surface chlorophyll concentration inversion method using cross-attention mechanism and spatial-spectral fusion according to claim 5 is characterized in that: When performing longitude and latitude matching in step 4, when the known chlorophyll concentration in the same time and area is a grid image, and the spatial resolution of the grid image is different from the spatial resolution of the multi-band remote sensing image dataset X1, it is necessary to downsample the image with the higher spatial resolution of the two images to ensure that the longitude and latitude matching of the two is performed at the same spatial resolution.

Images