Water quality remote sensing inversion method based on space spectrum fusion and ensemble learning

Abstract

The application discloses a water quality remote sensing inversion method based on space spectrum fusion and integrated learning, and has the advantages that the method can be applied to medium and small lakes and can be used for long-time sequence change dynamic monitoring of the water quality of the lakes while having high water quality inversion precision, and meets the needs of current actual application.

Description

Technical Field

[0001] This invention relates to water quality remote sensing inversion methods, and more particularly to a water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning. Background Technology

[0002] Water quality monitoring is a crucial reference for water quality status and water body management. Water quality is closely related to people's daily lives, making it essential to study how to conduct large-scale, rapid, and accurate water quality monitoring. Remote sensing technology has become an important tool for monitoring water quality in lakes and other bodies of water. The basic idea of ​​water quality remote sensing inversion is to first establish a regression model using water quality monitoring data and corresponding remote sensing satellite imagery. This model is then appropriately validated. Remote sensing satellite imagery can provide high-precision interpolation, more effectively generating spatially defined water quality maps, thus requiring less time and manpower. Water quality inversion methods based on multispectral remote sensing imagery have been widely applied in the field of water quality monitoring.

[0003] Existing water quality retrieval methods based on multispectral remote sensing imagery rely on regression models built from water quality monitoring data and corresponding remote sensing satellite imagery. Currently, the spectral band values ​​or band combinations of pixels corresponding to sampling points in multispectral remote sensing satellite imagery (i.e., remote sensing satellite imagery data) are used as independent variables, and the water quality parameter concentration values ​​of the sampling points (i.e., water quality monitoring data) are used as dependent variables, serving as the input and output of the regression model, respectively, to train the model. However, some remote sensing satellite imagery data obtained using remote sensing technology has high spatial resolution but low spectral resolution, resulting in low water quality retrieval accuracy; others have high spectral resolution but low spatial resolution, limiting their application in small and medium-sized lakes.

[0004] In addition, the concentration of water quality parameters in surface water is low. Current water quality inversion methods based on multispectral remote sensing images only use regression models with a single machine learning engine, resulting in low information mining ability and generalization. This has significant limitations in the dynamic monitoring of long-term water quality changes in lakes and cannot meet the needs of current practical applications. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning that has high water quality inversion accuracy, is applicable to small and medium-sized lakes, and can be used for dynamic monitoring of long-term water quality changes in lakes, thus meeting the needs of current practical applications.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: a water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning. This method constructs a spatial-spectral fusion model with multi-scale fusion and deep residual learning capabilities, and builds an ensemble learning regression model based on a Stacking strategy, comprising multiple different base models and a meta-model. The spatial-spectral fusion model is used to fuse features from Sentinel-2 and Gaofen-2 satellite images of the target water area with low spatial resolution, resulting in a high spatial resolution Sentinel-2 satellite image. Spectral band data of the water body in the high spatial resolution Sentinel-2 satellite image is extracted to obtain spectral band data. Each base model of the ensemble learning regression model is used to predict the spectral band data, resulting in multiple initial prediction results. Finally, the meta-model of the ensemble learning regression model is used to predict the multiple initial prediction results to obtain water quality inversion data.

[0007] The spatial-spectral fusion model includes two feature extraction units, one multi-scale fusion unit, and one stitching module, referred to as the first stitching module. The two feature extraction units are referred to as the first feature extraction unit and the second feature extraction unit, respectively. The first feature extraction unit extracts features from the Sentinel-2 satellite image of the target water area, and outputs the Sentinel-2 satellite image features to the multi-scale fusion unit. The second feature extraction unit extracts features from the Gaofen-2 satellite image of the target water area, and outputs the Gaofen-2 satellite image features to the multi-scale fusion unit. The multi-scale fusion unit performs multi-scale fusion of the Sentinel-2 satellite image features and the Gaofen-2 satellite image features output thereto, and outputs the fused high spatial resolution satellite image to the stitching module. The first stitching module stitches the fused high spatial resolution satellite image and the Sentinel-2 satellite image of the target water area to obtain a high spatial resolution Sentinel-2 satellite image.

[0008] Each feature extraction unit includes N feature extraction modules, where N is an integer greater than or equal to 4; each feature extraction module includes 4 first convolutional modules, each first convolutional module including a convolutional layer and a PReLU activation function, referred to as the first convolutional layer and the first PReLU activation function, both having input and output terminals, with the output of the first convolutional layer connected to the input of the first PReLU activation function; in each first convolutional module of each feature extraction module... The first convolutional layer is used to extract local features from the image features input to it, and the obtained local image features are output to the input of the first PReLU activation function. The first PReLU activation function is used to perform a nonlinear transformation on the image features input to it, and the corresponding image features are output through its output. The input of the first convolutional layer is the input of the first convolutional module, and the output of the first PReLU activation function is the output of the first convolutional module. In each feature extraction module, the output of the p-th first convolutional module is connected to the p-th... The inputs of the +1 first convolutional modules are connected, and the inputs of the first and fourth first convolutional modules are connected, p = 1, 2, 3. The input of the first convolutional module serves as the input of the feature extraction module, and the output of the fourth convolutional module serves as the output of the feature extraction module. Among the N feature extraction modules, the output of the i-th feature extraction module is connected to the input of the (i+1)-th feature extraction module, i = 1, 2, 3, ..., N-1. The input of the first feature extraction module is the input of the feature extraction unit, and the output of the k-th feature extraction module is the input of the feature extraction unit. The k-th output terminal of the feature extraction unit, k = 1, 2, 3, ..., N; the input terminal of the first feature extraction unit is used to access the Sentinel-2 satellite image of the target water area, and the k-th output terminal of the first feature extraction unit is used to output the Sentinel-2 satellite image features extracted by the k-th feature extraction module within it to the multi-scale fusion unit; the input terminal of the second feature extraction unit is used to access the Gaofen-2 satellite image of the target water area, and the k-th output terminal of the second feature extraction unit is used to output the Gaofen-2 satellite image features extracted by the k-th feature extraction module within it to the multi-scale fusion unit.

[0009] The multi-scale fusion unit comprises N multi-scale fusion modules and N-1 shallow fusion modules. Each multi-scale fusion module includes one computation module and four second convolution modules. Each computation module has two input terminals and one output terminal, referred to as the first input terminal and the second input terminal, respectively. The first input terminal of the computation module is used to input features from the Sentinel-2 satellite image, and the second input terminal is used to input features from the Gaofen-2 satellite image. The computation module performs a difference operation on the input Sentinel-2 satellite image features and Gaofen-2 satellite image features, and outputs the difference image features through its output terminal. Each second convolution module includes a convolutional layer and a PReLU activation function, referred to as the second convolutional layer and the PReLU activation function, respectively. Both the second convolutional layer and the second PReLU activation function have input terminals and output terminals, and the output terminal of the second convolutional layer is connected to the input terminal of the second PReLU activation function. In each second convolution module of each multi-scale fusion module, the second convolutional layer is used to connect its input terminal to the PReLU activation function. The input image features are used to extract local features, and the resulting local image features are output to the input of the second PReLU activation function. The second PReLU activation function is used to perform a nonlinear transformation on the input image features to obtain the corresponding image features, which are output through its output. The input of the second convolutional layer is the input of the second convolutional module, and the output of the second PReLU activation function is the output of the second convolutional module. In each multi-scale fusion module, the first input of the computation module is the first input of the multi-scale fusion module, the second input of the computation module is the second input of the multi-scale fusion module, and the output of the computation module is connected to the inputs of four second convolutional modules. The output of the first second convolutional module is connected to the inputs of the second, third, and fourth second convolutional modules, the output of the third second convolutional module is connected to the input of the fourth second convolutional module, and the output of the fourth second convolutional module is the output of the multi-scale fusion module.Each shallow fusion module consists of one stitching module and one convolutional layer. The stitching module is referred to as the second stitching module, and the convolutional layer as the third convolutional layer. The second stitching module has two inputs and one output, and the third convolutional layer has both inputs and an output. The output of the second stitching module is connected to the input of the third convolutional layer. The two inputs of the second stitching module are the two inputs of the shallow fusion module, and the output of the third convolutional layer is the output of the shallow fusion module. The outputs of the first and second multi-scale fusion modules are connected one-to-one with the two inputs of the first shallow fusion module. The j-th shallow... The output terminals of the layer fusion module and the (j+2)th multi-scale fusion module are connected one-to-one with the two input terminals of the (j+1)th shallow fusion module, where j = 1, 2, ..., N-2. The first input terminal of the kth multi-scale fusion module is connected to the kth output terminal of the first feature extraction unit, and the second input terminal of the kth multi-scale fusion module is connected to the kth output terminal of the second feature extraction unit. The first stitching module has a first input terminal, a second input terminal, and an output terminal. The output terminal of the (N-1)th shallow fusion module is connected to the first input terminal of the first stitching module. The second input terminal of the first stitching module is used to access the Sentinel-2 satellite image of the target water area.

[0010] The aforementioned spatial-spectral fusion model fuses features from Sentinel-2 and Gaofen-2 satellite images of the target water area after training. Before predicting the target water area, the aforementioned ensemble learning regression model also needs to be trained. The specific training processes for the spatial-spectral fusion model and the aforementioned ensemble learning regression model are as follows:

[0011] Step 1: Select a specific body of water as a reference body of water. Obtain N original Gaofen-2 satellite images of the reference body of water from the Natural Resources Satellite Remote Sensing Cloud Service Platform, and obtain N original Sentinel-2 satellite images of the reference body of water from the European Space Agency's Copernicus Open Data Platform. N is an integer greater than or equal to 6. Perform center-based cropping on the N original Gaofen-2 satellite images and N original Sentinel-2 satellite images respectively, ensuring that the size of the cropped N original Gaofen-2 satellite images and N original Sentinel-2 satellite images are integer multiples of 256 pixels. Perform image registration to obtain N pre-processed Gaofen-2 satellite images and N pre-processed Sentinel-2 satellite images.

[0012] Step 2: Based on the spectral characteristics of the Gaofen-2 and Sentinel-2 satellite images, each preprocessed Gaofen-2 satellite image includes four bands: B1, B2, B3, and B4; each preprocessed Sentinel-2 satellite image includes eight bands: B2, B3, B4, B5, B6, B7, B8, and B8A. Using the spectral response function of the Gaofen-2 satellite images, divide the N preprocessed Gaofen-2 satellite images into bands B1, B2, B3, and B4. The images were divided into four groups, resulting in four sets of preprocessed Gaofen-2 satellite images. The first group includes N preprocessed Gaofen-2 satellite images in the B1 band, the second group includes N preprocessed Gaofen-2 satellite images in the B2 band, the third group includes N preprocessed Gaofen-2 satellite images in the B3 band, and the fourth group includes N preprocessed Gaofen-2 satellite images in the B4 band. The four groups of preprocessed Gaofen-2 satellite images total a total of 4N single-band Gaofen-2 images. N preprocessed Sentinel-2 satellite images were analyzed using the spectral response function of the Sentinel-2 satellite. The eight bands (B2, B3, B4, B5, B6, B7, B8, and B8A) were divided into four groups, resulting in four groups of preprocessed Sentinel-2 satellite images. The first group includes N preprocessed Sentinel-2 satellite images with the B2 band; the second group includes N preprocessed Sentinel-2 satellite images with the B3 band; and the third group includes N preprocessed Sentinel-2 satellite images with the B4 band. The fourth group of preprocessed Sentinel-2 satellite images includes N preprocessed Sentinel-2 satellite images with B5, B6, B7, B8, and B8A bands. The four groups of preprocessed Sentinel-2 satellite images total 3N single-band Sentinel-2 satellite images and N composite-band Sentinel-2 satellite images. At this point, we obtain 3N single-band Gaofen-2 satellite images and N composite-band Sentinel-2 satellite images, for a total of 4N Sentinel-2 satellite images.

[0013] Step 3: Construct a sample set using 4N Gaofen-2 satellite images and 4N Sentinel-2 satellite images to obtain a sample set containing 4N Gaofen-2 satellite images and 4N Sentinel-2 satellite images;

[0014] Step 4: Crop the 4N Gaofen-2 satellite images and 4N Sentinel-2 satellite images in the sample set into multiple small image blocks of 256 pixels × 256 pixels with a step size of 128 pixels, to obtain multiple cropped Gaofen-2 satellite image blocks and multiple Sentinel-2 satellite image blocks.

[0015] Step 5: Construct a dataset using multiple cropped Gaofen-2 satellite images and multiple Sentinel-2 satellite images to obtain a dataset containing multiple Gaofen-2 satellite images and multiple Sentinel-2 satellite images. This dataset is the spatial-spectral fusion dataset.

[0016] Step 6: Divide the spatial-spectral fusion dataset into a training set and a test set according to the proportions. The training set contains 70% of the total dataset. If 70% of the total dataset is not an integer, round it up. The remaining part is used as the test set.

[0017] Step 7: Input the training set into the spatial-spectral fusion model, and use the Adam optimization algorithm to train the spatial-spectral fusion model. Use the test set to verify the accuracy of the spatial-spectral fusion model until the accuracy meets the requirements, and obtain the trained spatial-spectral fusion model.

[0018] Step 8: When it is necessary to invert the water quality parameters of the target water area, acquire the water quality data, geographical location data and Sentinel-2 satellite image data of the target area;

[0019] Step 9: Process the Sentinel-2 satellite image data from Step 8 using the same method as Steps 1 to 5 to obtain multiple small Sentinel-2 satellite images. Then, stitch the multiple small Sentinel-2 satellite images together and input them into the trained spatial-spectral fusion model. Use the trained spatial-spectral fusion model to generate high spatial resolution Sentinel-2 satellite images. Extract the spectral band data of the generated high spatial resolution Sentinel-2 satellite images. Combine the extracted spectral band data with the water quality data and geographic location data from Step S8 to construct a water quality inversion dataset.

[0020] Step 10: Use the water quality inversion dataset obtained in Step 9 to train the ensemble learning regression model to obtain the trained ensemble learning regression model.

[0021] When performing water quality remote sensing inversion on a target water area, after acquiring Gaofen-2 satellite images of the target water area from the Natural Resources Satellite Remote Sensing Cloud Service Platform and Sentinel-2 satellite images of the target water area from the European Space Agency's Copernicus Open Data Platform, the Gaofen-2 and Sentinel-2 satellite images of the target water area are first processed into multiple small Gaofen-2 and Sentinel-2 satellite images according to steps 1 to 5. Then, the multiple small Gaofen-2 and Sentinel-2 satellite images are matched one-to-one, and the corresponding small Gaofen-2 and Sentinel-2 satellite images are combined to obtain multiple combinations. The small Gaofen-2 and Sentinel-2 satellite images of each combination are input into the trained spatial-spectral fusion model for prediction to obtain the corresponding high spatial resolution Sentinel-2 satellite image. Each combination yields a high spatial resolution Sentinel-2 satellite image. Multiple combinations consisting of multiple small Gaofen-2 satellite images and multiple small Sentinel-2 satellite images yield multiple high spatial resolution Sentinel-2 satellite images. The spectral band data of each high spatial resolution Sentinel-2 satellite image are extracted and input into the trained ensemble learning regression model. The ensemble learning regression model outputs the corresponding water quality inversion data.

[0022] Compared with existing technologies, the advantages of this invention lie in constructing a spatial-spectral fusion model with multi-scale fusion and deep residual learning capabilities, and building an ensemble learning regression model based on a stacking strategy that includes multiple distinct base models and a meta-model. The spatial-spectral fusion model is used to fuse features from Sentinel-2 and Gaofen-2 satellite images of the target water area at low spatial resolution to obtain a high spatial resolution Sentinel-2 satellite image. Spectral band data of the water body in the high spatial resolution Sentinel-2 satellite image is extracted to obtain spectral band data. Each base model of the ensemble learning regression model is used to predict the spectral band data, resulting in multiple initial predictions. Finally, the meta-model of the ensemble learning regression model is used to make predictions. The model predicts multiple initial prediction results to obtain water quality inversion data. Since the spatial-spectral fusion model integrates multi-scale fusion and deep residual learning functions, it has high information mining ability and generalization ability. The spatial-spectral fusion model can fuse the features of Sentinel-2 satellite images and Gaofen-2 satellite images of target water areas with low spatial resolution to obtain high spatial resolution Sentinel-2 satellite images. Then, based on the high spatial resolution Sentinel-2 satellite images, spectral band data is obtained as input to the spatial-spectral fusion model for water quality inversion. Thus, this invention has high water quality inversion accuracy and is also applicable to small and medium-sized lakes. It can also be used for dynamic monitoring of long-term water quality changes in lakes, meeting the needs of current practical applications. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the structure of the ensemble learning regression model in the water quality remote sensing inversion method based on spatial spectrum fusion and ensemble learning of the present invention.

[0024] Figure 2 This is a schematic diagram of the spatial spectrum fusion model in the water quality remote sensing inversion method based on spatial spectrum fusion and ensemble learning of the present invention.

[0025] Figure 3 This is a schematic diagram of the feature extraction module of the spatial spectrum fusion model in the water quality remote sensing inversion method based on spatial spectrum fusion and ensemble learning of the present invention.

[0026] Figure 4 This is a schematic diagram of the structure of the first convolutional module of the spatial spectrum fusion model in the water quality remote sensing inversion method based on spatial spectrum fusion and ensemble learning of the present invention.

[0027] Figure 5 This is a schematic diagram of the multi-scale fusion module of the spatial-spectral fusion model in the water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning of the present invention.

[0028] Figure 6 This is a schematic diagram of the structure of the second convolutional module of the spatial spectrum fusion model in the water quality remote sensing inversion method based on spatial spectrum fusion and ensemble learning of the present invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0030] Example 1: As Figure 1 As shown, a water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning is proposed. This method constructs a spatial-spectral fusion model with multi-scale fusion and deep residual learning capabilities. An ensemble learning regression model, comprising multiple distinct base models and a meta-model, is built based on a stacking strategy. The spatial-spectral fusion model fuses features from Sentinel-2 and Gaofen-2 satellite images of the target water area at low spatial resolution to obtain a high spatial resolution Sentinel-2 satellite image. Spectral band data of the water body in the high spatial resolution Sentinel-2 satellite image is extracted to obtain spectral band data. Each base model of the ensemble learning regression model is used to predict the spectral band data, resulting in multiple initial predictions. Finally, the meta-model of the ensemble learning regression model is used to predict these initial predictions, yielding the water quality inversion data.

[0031] In this embodiment, because the spatial-spectral fusion model integrates multi-scale fusion and deep residual learning functions, it has high information mining capabilities and generalization. The spatial-spectral fusion model can fuse the features of Sentinel-2 satellite images and Gaofen-2 satellite images of target water areas with low spatial resolution to obtain high spatial resolution Sentinel-2 satellite images. Then, based on the high spatial resolution Sentinel-2 satellite images, spectral band data is obtained as input to the spatial-spectral fusion model for water quality inversion. Thus, this invention has high water quality inversion accuracy and is also applicable to small and medium-sized lakes. It is also applicable to the dynamic monitoring of long-term water quality changes in lakes, meeting the current practical application needs.

[0032] Example 2: This example is basically the same as Example 1, except that: Figure 2 As shown, in this embodiment, the spatial-spectral fusion model includes two feature extraction units, one multi-scale fusion unit, and one stitching module, referred to as the first stitching module; the two feature extraction units are referred to as the first feature extraction unit and the second feature extraction unit, respectively; the first feature extraction unit is used to extract features from the Sentinel-2 satellite image of the target water area, and outputs the Sentinel-2 satellite image features to the multi-scale fusion unit; the second feature extraction unit is used to extract features from the Gaofen-2 satellite image of the target water area, and outputs the Gaofen-2 satellite image features to the multi-scale fusion unit; the multi-scale fusion unit is used to perform multi-scale fusion of the Sentinel-2 satellite image features and the Gaofen-2 satellite image features output thereto, and outputs the fused high spatial resolution satellite image to the stitching module; the first stitching module is used to stitch the fused high spatial resolution satellite image and the Sentinel-2 satellite image of the target water area, and obtain a high spatial resolution Sentinel-2 satellite image.

[0033] like Figure 3 and Figure 4As shown in this embodiment, each feature extraction unit includes N feature extraction modules, where N is an integer greater than or equal to 4. Each feature extraction module includes four first convolutional modules, each of which includes a convolutional layer and a PReLU activation function. The convolutional layer is referred to as the first convolutional layer, and the PReLU activation function is referred to as the first PReLU activation function. Both the first convolutional layer and the first PReLU activation function have input and output terminals, and the output terminal of the first convolutional layer is connected to the input terminal of the first PReLU activation function. In each first convolutional module of each feature extraction module, the first convolutional layer is used to perform local feature extraction on the image features input to its input terminal, and the obtained local image features are output to the input terminal of the first PReLU activation function through its output terminal. The first PReLU activation function is used to perform nonlinear transformation on the image features input to its input terminal, and the corresponding image features are output through its output terminal. The input terminal of the first convolutional layer is the input terminal of the first convolutional module, and the output terminal of the first PReLU activation function is the output terminal of the first convolutional module. In each feature extraction module, the output terminal of the p-th first convolutional module is connected to the p+th first convolutional module. The input of the first convolutional module is connected, and the input of the first convolutional module is connected to the output of the fourth convolutional module, p = 1, 2, 3. The input of the first convolutional module serves as the input of the feature extraction module, and the output of the fourth convolutional module serves as the output of the feature extraction module. In N feature extraction modules, the output of the i-th feature extraction module is connected to the input of the (i+1)-th feature extraction module, i = 1, 2, 3, ..., N-1. The input of the first feature extraction module serves as the input of the feature extraction unit, and the input of the k-th feature extraction module... The output terminal is the kth output terminal of the feature extraction unit, k = 1, 2, 3, ..., N; the input terminal of the first feature extraction unit is used to access the Sentinel-2 satellite image of the target water area, and the kth output terminal of the first feature extraction unit is used to output the Sentinel-2 satellite image features extracted by the kth feature extraction module within it to the multi-scale fusion unit; the input terminal of the second feature extraction unit is used to access the Gaofen-2 satellite image of the target water area, and the kth output terminal of the second feature extraction unit is used to output the Gaofen-2 satellite image features extracted by the kth feature extraction module within it to the multi-scale fusion unit.

[0034] like Figure 5 and Figure 6As shown, in this embodiment, the multi-scale fusion unit includes N multi-scale fusion modules and N-1 shallow fusion modules. Each multi-scale fusion module includes one computation module and four second convolution modules. Each computation module has two input terminals and one output terminal. The two input terminals are referred to as the first input terminal and the second input terminal, respectively. The first input terminal of the computation module is used to input features from the Sentinel-2 satellite image, and the second input terminal is used to input features from the Gaofen-2 satellite image. The computation module performs a difference operation on the input Sentinel-2 satellite image features and Gaofen-2 satellite image features, and outputs the difference image features through its output terminal. Each second convolution module includes one convolutional layer and one PReLU activation function. This convolutional layer is referred to as the first convolutional layer. The second convolutional layer uses the PReLU activation function as its second PReLU activation function. Both the second convolutional layer and the second PReLU activation function have input and output terminals. The output of the second convolutional layer is connected to the input of the second PReLU activation function. In each second convolutional module of each multi-scale fusion module, the second convolutional layer extracts local features from the image features input to its input terminal. The resulting local image features are output to the input of the second PReLU activation function. The second PReLU activation function performs a nonlinear transformation on the image features input to its input terminal, obtaining corresponding image features that are output. The input of the second convolutional layer is the input of the second convolutional module. The output of the calculation module is the output of the second convolutional module; in each multi-scale fusion module, the first input of the calculation module is the first input of the multi-scale fusion module, the second input of the calculation module is the second input of the multi-scale fusion module, and the output of the calculation module is connected to the inputs of four second convolutional modules respectively. The output of the first second convolutional module is connected to the inputs of the second, third, and fourth second convolutional modules respectively, the output of the third second convolutional module is connected to the input of the fourth second convolutional module, and the output of the fourth second convolutional module is the output of the multi-scale fusion module; each shallow fusion module consists of one stitching module and one convolutional layer, which are stitched together... The convolutional module is called the second splicing module, and the convolutional layer is called the third convolutional layer. The second splicing module has two inputs and one output, and the third convolutional layer has inputs and an output. The output of the second splicing module is connected to the input of the third convolutional layer. The two inputs of the second splicing module are the two inputs of the shallow fusion module, and the output of the third convolutional layer is the output of the shallow fusion module. The outputs of the first and second multi-scale fusion modules are connected one-to-one with the two inputs of the first shallow fusion module. The outputs of the j-th and (j+2)-th multi-scale fusion modules are connected one-to-one with the two inputs of the (j+1)-th shallow fusion module, where j = 1, 2, ..., N-2.The first input of the k-th multi-scale fusion module is connected to the k-th output of the first feature extraction unit, and the second input of the k-th multi-scale fusion module is connected to the k-th output of the second feature extraction unit. The first stitching module has a first input, a second input, and an output. The output of the (N-1)-th shallow fusion module is connected to the first input of the first stitching module. The second input of the first stitching module is used to access the Sentinel-2 satellite image of the target water area.

[0035] In this embodiment, when performing water quality remote sensing inversion on the target water area, Gaofen-2 satellite images of the target water area are obtained from the Natural Resources Satellite Remote Sensing Cloud Service Platform, and Sentinel-2 satellite images of the target water area are obtained from the European Space Agency's Copernicus Open Data Platform. First, the Gaofen-2 and Sentinel-2 satellite images of the target water area are processed into multiple small Gaofen-2 satellite images and multiple small Sentinel-2 satellite images, respectively. Then, the multiple small Gaofen-2 satellite images and multiple small Sentinel-2 satellite images are matched one-to-one, and a corresponding small Gaofen-2 satellite image and a small... Pieces of Sentinel-2 satellite images are combined to form multiple combinations. Small pieces of Gaofen-2 satellite images from each combination are input to the input terminals of the first feature extraction unit and the first stitching module. Similarly, small pieces of Sentinel-2 satellite images from each combination are input to the input terminal of the second feature extraction unit. At this point, the first feature extraction unit performs image feature extraction on the small pieces of Sentinel-2 satellite images of the target water area, and the second feature extraction unit performs feature extraction on the small pieces of Gaofen-2 satellite images of the target water area. Specifically, the first feature extraction unit performs image feature extraction on the small pieces of Sentinel-2 satellite images of the target water area. The specific extraction process is as follows: First, the first feature extraction module of the first feature extraction unit extracts features from the Sentinel-2 satellite image of the target water area, and outputs the corresponding Sentinel-2 satellite image features through its output terminal. Then, in the second to nth feature extraction modules of the first feature extraction unit, when a corresponding Sentinel-2 satellite image feature is input to the input terminal of a feature extraction module, that feature extraction module extracts features from the input Sentinel-2 satellite image feature, and outputs the corresponding Sentinel-2 satellite image features through its output terminal. The second feature extraction unit then extracts features from the target water area. The specific process of image feature extraction for small blocks of Gaofen-2 satellite images is as follows: First, the first feature extraction module of the second feature extraction unit extracts features from the Gaofen-2 satellite image of the target water area and outputs the corresponding Gaofen-2 satellite image features through its output terminal. In the second to nth feature extraction modules of the first feature extraction unit, when a Gaofen-2 satellite image feature is input to the input terminal of a certain feature extraction module, the feature extraction module extracts features from the Gaofen-2 satellite image feature input therein and outputs the corresponding Gaofen-2 satellite image features through its output terminal.When both the first and second input terminals of a multi-scale fusion module in a multi-scale fusion unit are connected to Sentinel-2 satellite image features and Gaofen-2 satellite image features, the multi-scale fusion module performs feature fusion on the connected Sentinel-2 and Gaofen-2 satellite image features, and outputs the corresponding fused Gaofen-2 satellite image features through its output terminal. Similarly, when both the first and second input terminals of a shallow fusion module are connected to fused Gaofen-2 satellite image features, the shallow fusion module performs feature fusion on the connected two fused Gaofen-2 satellite image features, and outputs the corresponding fused Gaofen-2 satellite image features through its output terminal. In multi-scale fusion... When the unit completes the multi-scale fusion of the Sentinel-2 satellite image features and Gaofen-2 satellite image features at the input location, its output at this time is the fused Gaofen-2 satellite image features, which is the fused high spatial resolution satellite image. At this time, the first stitching module stitches the fused high spatial resolution satellite image and the Sentinel-2 satellite image of the target water area to obtain the high spatial resolution Sentinel-2 satellite image of each combination. Then, all the combined high spatial resolution Sentinel-2 satellite images are stitched together to obtain the high spatial resolution Sentinel-2 satellite image of the target water area. This high spatial resolution Sentinel-2 satellite image of the target water area can be used for subsequent spectral band data extraction.

[0036] Example 3: This example is basically the same as Example 1, except that: Figure 2 As shown, in this embodiment, the spatial-spectral fusion model fuses the features of Sentinel-2 and Gaofen-2 satellite images of the target water area after training. Before predicting the target water area, the ensemble learning regression model also needs to be trained. The specific training process of the spatial-spectral fusion model and the ensemble learning regression model is as follows:

[0037] Step 1: Select a specific body of water as a reference body of water. Obtain N original Gaofen-2 satellite images of the reference body of water from the Natural Resources Satellite Remote Sensing Cloud Service Platform, and obtain N original Sentinel-2 satellite images of the reference body of water from the European Space Agency's Copernicus Open Data Platform. N is an integer greater than or equal to 6. Perform center-based cropping on the N original Gaofen-2 satellite images and N original Sentinel-2 satellite images respectively, ensuring that the size of the cropped N original Gaofen-2 satellite images and N original Sentinel-2 satellite images are integer multiples of 256 pixels. Perform image registration to obtain N pre-processed Gaofen-2 satellite images and N pre-processed Sentinel-2 satellite images.

[0038] Step 2: Based on the spectral characteristics of the Gaofen-2 and Sentinel-2 satellite images, each preprocessed Gaofen-2 satellite image includes four bands: B1, B2, B3, and B4; each preprocessed Sentinel-2 satellite image includes eight bands: B2, B3, B4, B5, B6, B7, B8, and B8A. Using the spectral response function of the Gaofen-2 satellite images, divide the N preprocessed Gaofen-2 satellite images into bands B1, B2, B3, and B4. The images were divided into four groups, resulting in four sets of preprocessed Gaofen-2 satellite images. The first group includes N preprocessed Gaofen-2 satellite images in the B1 band, the second group includes N preprocessed Gaofen-2 satellite images in the B2 band, the third group includes N preprocessed Gaofen-2 satellite images in the B3 band, and the fourth group includes N preprocessed Gaofen-2 satellite images in the B4 band. The four groups of preprocessed Gaofen-2 satellite images total a total of 4N single-band Gaofen-2 images. N preprocessed Sentinel-2 satellite images were analyzed using the spectral response function of the Sentinel-2 satellite. The eight bands (B2, B3, B4, B5, B6, B7, B8, and B8A) were divided into four groups, resulting in four groups of preprocessed Sentinel-2 satellite images. The first group includes N preprocessed Sentinel-2 satellite images with the B2 band; the second group includes N preprocessed Sentinel-2 satellite images with the B3 band; and the third group includes N preprocessed Sentinel-2 satellite images with the B4 band. The fourth group of preprocessed Sentinel-2 satellite images includes N preprocessed Sentinel-2 satellite images with B5, B6, B7, B8, and B8A bands. The four groups of preprocessed Sentinel-2 satellite images total 3N single-band Sentinel-2 satellite images and N composite-band Sentinel-2 satellite images. At this point, we obtain 3N single-band Gaofen-2 satellite images and N composite-band Sentinel-2 satellite images, for a total of 4N Sentinel-2 satellite images.

[0039] Step 3: Construct a sample set using 4N Gaofen-2 satellite images and 4N Sentinel-2 satellite images to obtain a sample set containing 4N Gaofen-2 satellite images and 4N Sentinel-2 satellite images;

[0040] Step 4: Crop the 4N Gaofen-2 satellite images and 4N Sentinel-2 satellite images in the sample set into multiple small image blocks of 256 pixels × 256 pixels with a step size of 128 pixels, to obtain multiple cropped Gaofen-2 satellite image blocks and multiple Sentinel-2 satellite image blocks.

[0041] Step 5: Construct a dataset using multiple cropped Gaofen-2 satellite images and multiple Sentinel-2 satellite images to obtain a dataset containing multiple Gaofen-2 satellite images and multiple Sentinel-2 satellite images. This dataset is the spatial-spectral fusion dataset.

[0042] Step 6: Divide the spatial-spectral fusion dataset into a training set and a test set according to the proportions. The training set contains 70% of the total dataset. If 70% of the total dataset is not an integer, round it up. The remaining part is used as the test set.

[0043] Step 7: Input the training set into the spatial-spectral fusion model, and use the Adam optimization algorithm to train the spatial-spectral fusion model. Use the test set to verify the accuracy of the spatial-spectral fusion model until the accuracy meets the requirements, and obtain the trained spatial-spectral fusion model.

[0044] Step 8: When it is necessary to invert the water quality parameters of the target water area, acquire the water quality data, geographical location data and Sentinel-2 satellite image data of the target area;

[0045] Step 9: Process the Sentinel-2 satellite image data from Step 8 using the same method as Steps 1 to 5 to obtain multiple small Sentinel-2 satellite images. Then, stitch the multiple small Sentinel-2 satellite images together and input them into the trained spatial-spectral fusion model. Use the trained spatial-spectral fusion model to generate high spatial resolution Sentinel-2 satellite images. Extract the spectral band data of the generated high spatial resolution Sentinel-2 satellite images. Combine the extracted spectral band data with the water quality data and geographic location data from Step S8 to construct a water quality inversion dataset.

[0046] Step 10: Use the water quality inversion dataset obtained in Step 9 to train the ensemble learning regression model to obtain the trained ensemble learning regression model.

[0047] In this embodiment, when performing water quality remote sensing inversion on the target water area, after obtaining Gaofen-2 satellite images of the target water area from the Natural Resources Satellite Remote Sensing Cloud Service Platform and Sentinel-2 satellite images of the target water area from the European Space Agency's Copernicus Open Data Platform, the Gaofen-2 and Sentinel-2 satellite images of the target water area are first processed into multiple small Gaofen-2 and Sentinel-2 satellite images according to steps 1 to 5. Then, the multiple small Gaofen-2 and Sentinel-2 satellite images are matched one-to-one, and the corresponding small Gaofen-2 and Sentinel-2 satellite images form a combination, resulting in multiple combinations. The small Gaofen-2 and Sentinel-2 satellite images of each combination are input into the trained spatial-spectral fusion model for prediction to obtain the corresponding high spatial resolution Sentinel-2 satellite image. Each combination yields a high spatial resolution Sentinel-2 satellite image. Multiple combinations consisting of multiple small Gaofen-2 satellite images and multiple small Sentinel-2 satellite images yield multiple high spatial resolution Sentinel-2 satellite images. The spectral band data of each high spatial resolution Sentinel-2 satellite image are extracted and input into the trained ensemble learning regression model. The ensemble learning regression model outputs the corresponding water quality inversion data.

[0048] In this embodiment, when training the spatial-spectral fusion model, multi-band data from Sentinel-2 and Gaofen-2 satellite images are fused, and centered cropping, image registration, and grouping are performed. The images are then cropped into small blocks of 256×256 pixels, thereby constructing a rich and diverse spatial-spectral fusion dataset. The Adam optimization algorithm is then used to train the spatial-spectral fusion model, improving image resolution, enhancing the generalization ability of the spatial-spectral fusion model, and improving training efficiency, ensuring that the spatial-spectral fusion model can accurately invert the actual situation of the target water area.

Claims

1. A water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning, characterized in that... By constructing a spatial-spectral fusion model with multi-scale fusion and deep residual learning capabilities, and building an ensemble learning regression model based on a stacking strategy that includes multiple different base models and a meta-model, the spatial-spectral fusion model is used to fuse features of Sentinel-2 and Gaofen-2 satellite images of the target water area with low spatial resolution to obtain high spatial resolution Sentinel-2 satellite images. Spectral band data of the water body in the high spatial resolution Sentinel-2 satellite images are extracted to obtain spectral band data. Each base model of the ensemble learning regression model is used to predict the spectral band data to obtain multiple initial prediction results. Then, the meta-model of the ensemble learning regression model is used to predict the multiple initial prediction results to obtain water quality inversion data. The spatial-spectral fusion model includes two feature extraction units, one multi-scale fusion unit, and one stitching module, referred to as the first stitching module. The two feature extraction units are referred to as the first feature extraction unit and the second feature extraction unit, respectively. The first feature extraction unit extracts features from the Sentinel-2 satellite image of the target water area, and outputs the Sentinel-2 satellite image features to the multi-scale fusion unit. The second feature extraction unit extracts features from the Gaofen-2 satellite image of the target water area, and outputs the Gaofen-2 satellite image features to the multi-scale fusion unit. The multi-scale fusion unit performs multi-scale fusion of the Sentinel-2 satellite image features and the Gaofen-2 satellite image features output thereto, and outputs the fused high spatial resolution satellite image to the stitching module. The first stitching module stitches the fused high spatial resolution satellite image and the Sentinel-2 satellite image of the target water area to obtain a high spatial resolution Sentinel-2 satellite image.

2. The water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning according to claim 1, characterized in that... Each feature extraction unit includes N feature extraction modules, where N is an integer greater than or equal to 4. Each feature extraction module includes four first convolutional modules. Each first convolutional module includes a convolutional layer and a PReLU activation function. The convolutional layer is called the first convolutional layer, and the PReLU activation function is called the first PReLU activation function. Both the first convolutional layer and the first PReLU activation function have input and output terminals. The output terminal of the first convolutional layer is connected to the input terminal of the first PReLU activation function. In each first convolutional module of each feature extraction module, the first convolutional layer is used to extract local features from the image features input to its input terminal. The obtained local image features are output to the input terminal of the first PReLU activation function through its output terminal. The first PReLU activation function is used to perform a nonlinear transformation on the image features input to its input terminal. The obtained corresponding image features are output through its output terminal. The input terminal of the first convolutional layer is the input terminal of the first convolutional module, and the output terminal of the first PReLU activation function is the output terminal of the first convolutional module. In each feature extraction module, the output of the p-th first convolutional module is connected to the input of the (p+1)-th first convolutional module, and the input of the 1st first convolutional module is connected to the output of the 4th first convolutional module. p=1,2,3. The input of the 1st first convolutional module serves as the input of the feature extraction module, and the output of the 4th first convolutional module serves as the output of the feature extraction module. In the N feature extraction modules, the output of the i-th feature extraction module is connected to the input of the (i+1)-th feature extraction module, i=1,2,3,…,N-1; the input of the 1-th feature extraction module is the input of the feature extraction unit, and the output of the k-th feature extraction module is the k-th output of the feature extraction unit, k=1,2,3,…,N; The input terminal of the first feature extraction unit is used to access the Sentinel-2 satellite image of the target water area, and the kth output terminal of the first feature extraction unit is used to output the Sentinel-2 satellite image features extracted by the kth feature extraction module within it to the multi-scale fusion unit; the input terminal of the second feature extraction unit is used to access the Gaofen-2 satellite image of the target water area, and the kth output terminal of the second feature extraction unit is used to output the Gaofen-2 satellite image features extracted by the kth feature extraction module within it to the multi-scale fusion unit.

3. The water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning according to claim 1, characterized in that... The multi-scale fusion unit comprises N multi-scale fusion modules and N-1 shallow fusion modules. Each multi-scale fusion module includes one computation module and four second convolution modules. Each computation module has two input terminals and one output terminal, referred to as the first input terminal and the second input terminal, respectively. The first input terminal of the computation module is used to input features from the Sentinel-2 satellite image, and the second input terminal is used to input features from the Gaofen-2 satellite image. The computation module performs a difference operation on the input Sentinel-2 satellite image features and Gaofen-2 satellite image features, and outputs the difference image features through its output terminal. Each second convolution module includes a convolutional layer and a PReLU activation function, referred to as the second convolutional layer and the PReLU activation function, respectively. Both the second convolutional layer and the second PReLU activation function have input terminals and output terminals, and the output terminal of the second convolutional layer is connected to the input terminal of the second PReLU activation function. In each second convolutional module of each multi-scale fusion module, the second convolutional layer is used to extract local features from the image features input to its input end, and outputs the obtained local image features to the input end of the second PReLU activation function through its output end. The second PReLU activation function is used to perform nonlinear transformation on the image features input to its input end, and outputs the corresponding image features through its output end. The input end of the second convolutional layer is the input end of the second convolutional module, and the output end of the second PReLU activation function is the output end of the second convolutional module. In each multi-scale fusion module, the first input terminal of the computation module is the first input terminal of the multi-scale fusion module, the second input terminal of the computation module is the second input terminal of the multi-scale fusion module, the output terminal of the computation module is connected to the input terminals of four second convolution modules respectively, the output terminal of the first second convolution module is connected to the input terminals of the second, third and fourth second convolution modules respectively, the output terminal of the third second convolution module is connected to the input terminal of the fourth second convolution module, and the output terminal of the fourth second convolution module is the output terminal of the multi-scale fusion module; Each shallow fusion module consists of one stitching module and one convolutional layer. The stitching module is referred to as the second stitching module, and the convolutional layer as the third convolutional layer. The second stitching module has two inputs and one output, and the third convolutional layer has both inputs and an output. The output of the second stitching module is connected to the input of the third convolutional layer. The two inputs of the second stitching module are the two inputs of the shallow fusion module, and the output of the third convolutional layer is the output of the shallow fusion module. The outputs of the first and second multi-scale fusion modules are connected one-to-one with the two inputs of the first shallow fusion module. The j-th shallow... The output of the layer fusion module and the output of the (j+2)th multi-scale fusion module are connected one-to-one with the two inputs of the (j+1)th shallow fusion module, j=1,2,…,N-2; the first input of the kth multi-scale fusion module is connected to the kth output of the first feature extraction unit, and the second input of the kth multi-scale fusion module is connected to the kth output of the second feature extraction unit. The first stitching module has a first input, a second input, and an output. The output of the (N-1)th shallow fusion module is connected to the first input of the first stitching module. The second input of the first stitching module is used to access the Sentinel-2 satellite image of the target water area.

4. A water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning according to any one of claims 1 to 3, characterized in that... The aforementioned spatial-spectral fusion model fuses features from Sentinel-2 and Gaofen-2 satellite images of the target water area after training. Before predicting the target water area, the aforementioned ensemble learning regression model also needs to be trained. The specific training processes for the spatial-spectral fusion model and the aforementioned ensemble learning regression model are as follows: Step 1: Select a specific body of water as a reference body of water. Obtain N original Gaofen-2 satellite images of the reference body of water from the Natural Resources Satellite Remote Sensing Cloud Service Platform, and obtain N original Sentinel-2 satellite images of the reference body of water from the European Space Agency's Copernicus Open Data Platform. N is an integer greater than or equal to 6. Perform center-based cropping on the N original Gaofen-2 satellite images and N original Sentinel-2 satellite images respectively, ensuring that the size of the cropped N original Gaofen-2 satellite images and N original Sentinel-2 satellite images are integer multiples of 256 pixels. Perform image registration to obtain N pre-processed Gaofen-2 satellite images and N pre-processed Sentinel-2 satellite images. Step 2: Based on the spectral characteristics of the Gaofen-2 and Sentinel-2 satellite images, each preprocessed Gaofen-2 satellite image includes four bands: B1, B2, B3, and B4; each preprocessed Sentinel-2 satellite image includes eight bands: B2, B3, B4, B5, B6, B7, B8, and B8A. Using the spectral response function of the Gaofen-2 satellite images, the N preprocessed Gaofen-2 satellite images are divided into four bands according to the four bands B1, B2, B3, and B4. The process yielded four sets of preprocessed Gaofen-2 satellite images. The first set included N preprocessed Gaofen-2 satellite images with the B1 band, the second set included N preprocessed Gaofen-2 satellite images with the B2 band, the third set included N preprocessed Gaofen-2 satellite images with the B3 band, and the fourth set included N preprocessed Gaofen-2 satellite images with the B4 band. The four sets of preprocessed Gaofen-2 satellite images totaled 4N single-band Gaofen-2 satellite images. Using the spectral response function of Sentinel-2 satellite images, the eight bands (B2, B3, B4, B5, B6, B7, B8, and B8A) of N preprocessed Sentinel-2 satellite images are divided into four groups, resulting in four groups of preprocessed Sentinel-2 satellite images. The first group includes N preprocessed Sentinel-2 satellite images with the B2 band, the second group includes N preprocessed Sentinel-2 satellite images with the B3 band, and the third group includes N preprocessed Sentinel-2 satellite images with the B4 band. The fourth group of Sentinel-2 satellite images includes N preprocessed Sentinel-2 satellite images with B5, B6, B7, B8, and B8A bands. The four groups of preprocessed Sentinel-2 satellite images total 3N single-band Sentinel-2 satellite images and N composite-band Sentinel-2 satellite images. At this point, we obtain 3N single-band Gaofen-2 satellite images and N composite-band Sentinel-2 satellite images, for a total of 4N Sentinel-2 satellite images. Step 3: Construct a sample set using 4N Gaofen-2 satellite images and 4N Sentinel-2 satellite images to obtain a sample set containing 4N Gaofen-2 satellite images and 4N Sentinel-2 satellite images; Step 4: Crop the 4N Gaofen-2 satellite images and 4N Sentinel-2 satellite images in the sample set into multiple small image blocks of 256 pixels × 256 pixels with a step size of 128 pixels, to obtain multiple cropped Gaofen-2 satellite image blocks and multiple Sentinel-2 satellite image blocks. Step 5: Construct a dataset using multiple cropped Gaofen-2 satellite images and multiple Sentinel-2 satellite images to obtain a dataset containing multiple Gaofen-2 satellite images and multiple Sentinel-2 satellite images. This dataset is the spatial-spectral fusion dataset. Step 6: Divide the spatial-spectral fusion dataset into a training set and a test set according to the proportion. The training set contains 70% of the total dataset. If 70% of the total dataset is not an integer, round it up. The remaining part is used as the test set. Step 7: Input the training set into the spatial-spectral fusion model, and use the Adam optimization algorithm to train the spatial-spectral fusion model. Use the test set to verify the accuracy of the spatial-spectral fusion model until the accuracy meets the requirements, and obtain the trained spatial-spectral fusion model. Step 8: When it is necessary to invert the water quality parameters of the target water area, acquire the water quality data, geographical location data and Sentinel-2 satellite image data of the target area; Step 9: Process the Sentinel-2 satellite image data from Step 8 using the same method as Steps 1 to 5 to obtain multiple small Sentinel-2 satellite images. Then, stitch the multiple small Sentinel-2 satellite images together and input them into the trained spatial-spectral fusion model. Use the trained spatial-spectral fusion model to generate high spatial resolution Sentinel-2 satellite images. Extract the spectral band data of the generated high spatial resolution Sentinel-2 satellite images. Combine the extracted spectral band data with the water quality data and geographic location data from Step S8 to construct a water quality inversion dataset. Step 10: Use the water quality inversion dataset obtained in Step 9 to train the ensemble learning regression model to obtain the trained ensemble learning regression model.

5. The water quality remote sensing inversion method based on spatial-spectral fusion and ensemble learning according to claim 4, characterized in that... When performing water quality remote sensing inversion on a target water area, after acquiring Gaofen-2 satellite images of the target water area from the Natural Resources Satellite Remote Sensing Cloud Service Platform and Sentinel-2 satellite images of the target water area from the European Space Agency's Copernicus Open Data Platform, the Gaofen-2 and Sentinel-2 satellite images of the target water area are first processed into multiple small Gaofen-2 and Sentinel-2 satellite images according to steps 1 to 5. Then, the multiple small Gaofen-2 and Sentinel-2 satellite images are matched one-to-one, and the corresponding small Gaofen-2 and Sentinel-2 satellite images constitute a... One combination yields multiple combinations. Small patches of Gaofen-2 and Sentinel-2 satellite images from each combination are input into a trained spatial-spectral fusion model for prediction, resulting in corresponding high spatial resolution Sentinel-2 satellite images. Each combination yields one high spatial resolution Sentinel-2 satellite image. Multiple combinations consisting of multiple small patches of Gaofen-2 and Sentinel-2 satellite images yield multiple high spatial resolution Sentinel-2 satellite images. The spectral band data of each high spatial resolution Sentinel-2 satellite image are extracted and input into a trained ensemble learning regression model. The ensemble learning regression model outputs the corresponding water quality inversion data.

Images