Cnn super-resolution land cover mapping method and system fusing geoinformation

Abstract

The application relates to a CNN super-resolution land cover mapping method and system fusing geographical information, and the method comprises the following steps: S1, remote sensing image data and geographical information data with geographical position information are acquired, an improved multi-scale U-Net model is input, and remote sensing image features and geographical information features are output; S2, the remote sensing image features and the geographical information features are input into an attention mechanism position weight calculation module, and a geographical information feature map with position weight is constructed; S3, the geographical information feature map with position weight is superimposed with the remote sensing image features through a feature fusion module, and an abstract feature map fusing the geographical information and the image features is generated; and S4, a continuous upsampling operation is performed on the abstract feature map, a feature map with higher spatial resolution under a specific scale is acquired, and a super-resolution mapping result of each sub-pixel is acquired by using an activation function. The method provided by the application can acquire high-resolution and high-precision land cover mapping.

Description

Technical Field

[0001] This invention relates to the field of geospatial information technology, specifically to a CNN super-resolution land cover mapping method and system that integrates geospatial information. Background Technology

[0002] Due to the complexity of ground feature distribution, the influence of various environments during electromagnetic radiation transmission, and the physical characteristics of the sensors themselves, pixels in remote sensing images are rarely composed of a single ground feature. Pixels containing multiple ground feature types are called mixed pixels (Hu Maogui and Wang Jinfeng 2010, Ren Wu and Ge Yong 2011). The existence of mixed pixels increases the uncertainty of traditional hard classification results, seriously affecting the accuracy of remote sensing image information extraction. To address this issue, the concept of mixed pixel decomposition was proposed. It refers to determining the component information of different categories within each pixel by constructing the transformation relationship between the spectral values ​​of the pixel and the pure category pixels. Although unmixing mixed pixels can reduce the uncertainty of categories within a pixel, the uncertainty of the location of ground features within the pixel still exists. To identify the distribution of different ground features in mixed pixels, Atkinson first proposed the concept of super-resolution mapping (also known as sub-pixel mapping). It determines the location of sub-pixel categories based on the spatial structure of the pixel, under the constraint of the proportion of category components. This method can accurately determine the land cover category and distribution in mixed pixels, and has shown good application in land cover renewal, water boundary extraction, fine land cover extraction and change detection.

[0003] In recent years, deep learning models in the field of computer vision have developed rapidly and demonstrated outstanding performance in remote sensing image information extraction. Deep learning possesses powerful learning and representation capabilities, enabling effective image feature segmentation and classification. However, deep learning is a typical data-driven method, suffering from problems such as poor interpretability. Given these shortcomings in remote sensing information extraction, some research has begun to consider integrating geoscientific information into deep learning models. Geoscientific information describes and understands geological patterns and processes, possessing strong logical reasoning capabilities. Integrating geoscientific information with deep learning models can effectively improve the accuracy and interpretability of information extraction.

[0004] Currently, the main methods for fusing geoscientific information include input overlay and feature overlay. However, these two methods ignore the varying importance of geoscientific information at different locations, potentially leading to wasted features and classification errors resulting from feature combinations. Therefore, how to effectively overlay geoscientific information with image features across different geographical locations to improve mapping accuracy has become an urgent problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a CNN super-resolution land cover mapping method and system that integrates geoscientific information.

[0006] The technical solution of this invention is: a CNN super-resolution land cover mapping method that integrates geoscientific information, comprising:

[0007] Step S1: Acquire remote sensing image data and geoscientific information data with geographic location information, input the remote sensing image data and the geoscientific information data into the improved multi-scale U-Net model, and extract the corresponding remote sensing image features and geoscientific information features respectively;

[0008] Step S2: Input the remote sensing image features and geoscientific information features into the attention mechanism location weight calculation module to calculate the weights of the geoscientific information features and construct a geoscientific information feature map with location weights;

[0009] Step S3: The geoscientific information feature map with location weights is superimposed with the remote sensing image features through the feature fusion module to generate an abstract feature map that fuses geoscientific information and image features;

[0010] Step S4: Perform continuous upsampling operations on the abstract feature map to obtain a feature map with higher spatial resolution at a specific scale, and then use the activation function to obtain the classification results of each sub-pixel, i.e., the super-resolution mapping results.

[0011] Compared with the prior art, the present invention has the following advantages:

[0012] 1. This invention discloses a CNN super-resolution land cover mapping method that integrates geoscientific information. It fully leverages the characteristics of deep learning models and geoscientific information, incorporating geoscientific information into deep learning models. Compared with a single deep learning model, the mapping results using the method of this invention have higher accuracy and a certain degree of interpretability.

[0013] 2. This invention addresses the shortcomings of current methods for fusing geoscientific information by designing a dual-branch network structure model based on a multi-scale U-Net. By introducing dilated convolution into the U-Net model, multi-scale feature extraction from remote sensing imagery and geoscientific information is achieved. Furthermore, a spatial attention mechanism is incorporated into the model to enable efficient fusion of remote sensing imagery and geoscientific information at specific locations, solving the problem of neglecting the location of geoscientific information in existing methods and providing a new approach for fusing geoscientific information.

[0014] 3. The method of this invention can obtain high-resolution, high-precision land cover products, which can be applied to various types of geoscientific information to participate in super-resolution land cover mapping. It has strong scalability and helps to promote the development and application of remote sensing information extraction technology. Attached Figure Description

[0015] Figure 1 This is a flowchart of a CNN super-resolution land cover mapping method that integrates geoscientific information, as described in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of the structure of the improved multi-scale U-Net model in an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram of the attention mechanism position weight calculation module in an embodiment of the present invention;

[0018] Figure 4 This is a schematic diagram of the feature fusion module and the super-resolution classification module in an embodiment of the present invention;

[0019] Figure 5 This is a schematic diagram of the process for super-resolution land cover mapping that integrates remote sensing imagery and the Normalized Difference Vegetation Index (NDVI) in an embodiment of the present invention.

[0020] Figure 6 This is a structural block diagram of a CNN super-resolution land cover mapping system that integrates geoscientific information, as described in an embodiment of the present invention. Detailed Implementation

[0021] This invention provides a CNN super-resolution land cover mapping method that integrates geoscientific information, which can obtain high-resolution, high-precision land cover maps.

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below through specific implementations and in conjunction with the accompanying drawings.

[0023] Example 1

[0024] like Figure 1 As shown in the figure, an embodiment of the present invention provides a CNN super-resolution land cover mapping method that integrates geoscientific information, comprising the following steps:

[0025] Step S1: Acquire remote sensing image data and geoscientific information data with geographic location information. Input the remote sensing image data and geoscientific information data into the improved multi-scale U-Net model and extract the corresponding remote sensing image features and geoscientific information features respectively.

[0026] Step S2: Input the remote sensing image features and geoscientific information features into the attention mechanism location weight calculation module to calculate the weights of the geoscientific information features and construct a geoscientific information feature map with location weights;

[0027] Step S3: Overlay the geoscientific information feature map with location weights and the remote sensing image features through the feature fusion module to generate an abstract feature map that integrates geoscientific information and image features;

[0028] Step S4: Perform continuous upsampling operations on the abstract feature map to obtain a feature map with higher spatial resolution at a specific scale, and then use the activation function to obtain the classification results of each sub-pixel, i.e., the super-resolution mapping results.

[0029] In one embodiment, step S1 above: acquiring remote sensing image data and geoscientific information data with geographic location information, inputting the remote sensing image data and geoscientific information data into an improved multi-scale U-Net model, and extracting the corresponding remote sensing image features and geoscientific information features respectively, specifically including:

[0030] Step S11: Construct an improved multi-scale U-Net model, including an encoder, a dilated convolution module, and a decoder. The encoder consists of a 3-layer structure: the first two layers consist of a convolutional layer, a ReLU layer, and a pooling layer; the third layer consists of a convolutional layer and a ReLU layer. The dilated convolution module consists of three dilated convolutions with different dilation coefficients. The decoder consists of a 2-layer structure: each layer consists of a deconvolutional layer, a convolutional layer, and a ReLU layer.

[0031] Step S12: Input the remote sensing image data and geoscientific information data into the encoder of the improved multi-scale U-Net model, and perform feature extraction through the encoder's 3-layer structure:

[0032] I RS_en_3 =F RS_en_3 (F RS_en_2 (F RS_en_1 (I rs )))

[0033] I GI_en_3 =F GI_en_3 (F GI_en_2 (F GI_en_1 (I gi )))

[0034] Among them, I RS_en_3 F represents the remote sensing image features extracted by the third layer structure of the encoder. RS_en_1 F represents the first layer structure for extracting image features using the encoder. RS_en_2 F represents the second layer structure of the encoder for extracting image features. RS_en_3 This represents the third layer structure after the encoder extracts image features; I rs This represents the input remote sensing image data;

[0035] I GI_en_3F represents the geoscientific information features extracted by the third layer structure of the encoder. GI_en_1 F represents the first layer structure of the encoder for extracting geoscientific information features. GI_en_2 F represents the second-layer structure of the encoder for extracting geoscientific information features. GI_en_3 I represents the third layer structure of the encoder for extracting geoscientific information features. gi This represents the input geoscientific information data;

[0036] Step S13: Place I RS_en_3 and I GI_en_3 Input the dilated convolution module, and use three layers of dilated convolution with dilation coefficients of 1, 2, and 3 to perform multi-scale feature extraction:

[0037] I RS_1 =Contact(F RS_MR1 (I RS_en_3 ),F RS_MR2 (I RS_en_3 ),F RS_MR3 (I RS_en_3 ))

[0038] I GI_1 =Contact(F GI_MR1 (I GI_en_3 ),F GI_MR2 (I GI_en_3 ),F GI_MR3 (I GI_en_3 ))

[0039] Among them, I RS_1 This represents the extracted multi-scale remote sensing image features, Contact represents the connection operation, and F RS_MR1 F RS_MR2 F RS_MR3 These represent dilated convolutional features in remote sensing images with dilation rates of 1, 2, and 3, respectively; I GI_1 F represents the extracted multi-scale geoscientific information features. GI_MR1 F GI_MR2 F GI_MR3 These represent geoscientific information feature dilated convolutions with dilation rates of 1, 2, and 3, respectively.

[0040] Step S14: Place I RS_1 and I GI_1 The input is processed by the decoder, undergoing upsampling and multi-level feature fusion. During this process, to fuse features from different levels, the skip connection operation of the U-Net model is preserved to stitch together deep and shallow features of the same resolution. The decoding process is as follows:

[0041] I RS_de_1 =ReLU(Conv(Trans(connect(I)) RS_en_2 ,IRS_1 ))))

[0042] I RS_de_2 =ReLU(Conv(Trans(connect(I)) RS_en_1 ,I RS_de_1 ))))

[0043] I GI_de_1 =ReLU(Conv(Trans(connect(I)) GI_en_2 ,I GI_1 ))))

[0044] I GI_de_2 =ReLU(Conv(Trans(connect(I)) GI_en_1 ,I GI_de_1 ))))

[0045] Among them, I RS_1 For the input multi-scale remote sensing image features, I RS_en_2 The remote sensing image features extracted from the second layer of the encoder structure, connect represents the skip connection operation, Trans represents the deconvolution operation, Conv represents the convolution operation, ReLU represents the ReLU activation function, and I RS_de_1 This represents the remote sensing image features extracted by the first layer of the decoder; I RS_en_1 I represents the remote sensing image features extracted from the first layer of the encoder structure. RS_de_2 This represents the remote sensing image features extracted by the second layer of the decoder;

[0046] I GI_1 For the input multi-scale geoscientific information features, I GI_en_2 For the geoscientific information features extracted from the second layer structure of the encoder, I GI_de_1 I represents the geoscientific information features extracted by the first layer of the decoder. GI_en_1 For the geoscientific information features extracted from the first layer structure of the encoder, I GI_de_2 This represents the geoscientific information features extracted by the second layer structure of the decoder;

[0047] Let the final output of the decoding part be represented as follows:

[0048] I RS =I RS_de_2

[0049] I GI =I GI_de_2

[0050] Among them, I RS Indicates remote sensing image features, I GI It represents the characteristics of geoscientific information.

[0051] like Figure 2 The diagram shows an improved multi-scale U-Net model that retains the classic encoding and decoding structure of the U-Net model. The encoder of this improved multi-scale U-Net model mainly consists of a three-layer structure, each layer comprising a convolutional layer, a ReLU layer, and a pooling layer. The pooling layer is removed in the third layer, so the decoding part only includes two downsampling operations. The middle part of the improved multi-scale U-Net model consists of three dilated convolutional modules used to extract multi-scale features from remote sensing images and geoscientific information at different scales. For the extracted information at different scales, a feature fusion operation is performed to input the data into the model's decoder. The decoder of the improved multi-scale U-Net model consists of a two-layer structure, each layer comprising a deconvolutional layer, a convolutional layer, and a ReLU layer. Two deconvolutions are used to upsample the feature maps, restoring them to the same size as the model's input data. For features at different levels, the classic skip-layer connection operation of the U-Net model is still retained, that is, feature maps of the same size as the decoder and encoder are concatenated. Using the features extracted from the second layer of the encoder and the dilated convolutional features extracted by the dilated convolution module as inputs to the next layer, the improved multi-scale U-Net model of this invention adopts a three-step hop connection, specifically: the feature map extracted from the second layer of the encoder and the dilated convolutional features extracted by the dilated convolution module are concatenated as inputs to the first deconvolution of the decoder; the features extracted from the first layer of the encoder and the features extracted from the first layer of the decoder are concatenated as inputs to the second deconvolution of the decoder; and the input data and the features extracted from the second layer of the decoder are concatenated as the output.

[0052] In one embodiment, step S2 above—inputting remote sensing image features and geoscientific information features into the attention mechanism location weight calculation module, calculating the weights of the geoscientific information features, and constructing a geoscientific information feature map with location weights—specifically includes:

[0053] Step S21: For I RS and I GI Further feature extraction is performed using two consecutive convolutions:

[0054] I RS' =Conv1(Conv 32 (I RS ))

[0055] I GI' =Conv1(Conv 32 (I GI ))

[0056] Among them, Conv 32 Conv1 indicates the first convolution operation with a kernel size of 32 and a convolution size of 3×3; Conv2 indicates the second convolution operation with a kernel size of 1 and a convolution size of 3×3.

[0057] Step S22: For I RS' and I GI' Perform location-by-location multiplication to determine the correlation between remote sensing image features and geoscientific information features at each location. SA :

[0058] I SA =Multiply(I RS' I GI' )

[0059] Step S23: Use the Sigmoid activation function to activate I. SA Mapped to a value between 0 and 1, the weight value SA represents the weight of the remote sensing image features and the geoscientific information features at each location:

[0060] SA = Sigmoid(I) SA )

[0061] Step S24: Transfer geoscientific information features I GI Multiplying with SA generates a geoscientific feature map W with location weights. GA :

[0062] W GA =Multiply(I GI SA)

[0063] Here, Multiply represents the position-by-position multiplication operation.

[0064] like Figure 3 The diagram shows the location weight calculation module of the attention mechanism. The input to this module is remote sensing image features and geoscientific information features. This module mainly consists of two branches, each containing two different convolutional layers. The first layer has 32 kernels, and the second layer has 1 kernel. The features extracted from the two branches are fused through a Multiply layer, and finally, a sigmoid layer is used to perform weight mapping. This module utilizes a spatial attention mechanism to calculate the weights of the geoscientific information features, forming a location weight map of the geoscientific information features.

[0065] In one embodiment, step S3 above, which involves overlaying the location-weighted geoscientific information feature map with remote sensing image features using a feature fusion module to generate an abstract feature map that fuses geoscientific information and image features, specifically includes:

[0066] Remote sensing image features I RS With W GA Input feature fusion module:

[0067] Fconc = Add(I RS W GA )

[0068] Among them, I RS This represents the features of remote sensing images. Add represents the overlay operation, and Fconc represents the abstract feature map that integrates geoscientific information and image features.

[0069] Feature fusion module, such as Figure 4 As shown, this module mainly includes an Add layer, which is used to fuse the geoscientific information feature map with location weights with remote sensing image features to form an abstract feature map of fused image and geoscientific information.

[0070] In one embodiment, step S4 above involves performing continuous upsampling operations on the abstract feature map to obtain a feature map with higher spatial resolution at a specific scale, and then using an activation function to obtain the classification results of each sub-pixel, i.e., the super-resolution mapping results, specifically including:

[0071] The pixel shuffle method is used to upsample the abstract feature map that integrates geoscientific information and image features to form a high-resolution feature map. The sigmoid function is then used to classify the high-resolution feature map, thereby obtaining the land cover mapping results for each land cover type at the sub-pixel scale.

[0072] H Fconc = pixelshuffle(Fconc, r)

[0073] F pixel_class =sigmoid(H Fconc )

[0074] Where pixelshuffle is the pixelshuffle upsampling operation, r represents the upsampling amplification factor, and H Fconc F represents the high-resolution abstract feature map after upsampling. pixel_class This represents the classification result for each sub-pixel of the high-resolution abstract feature map.

[0075] Super-resolution classification module such as Figure 4 As shown, this module uses the pixel shuffle method for upsampling to form a high-resolution feature map, and then uses the sigmoid function to classify the high-resolution feature map, thereby obtaining land cover mapping results for each land cover type at the sub-pixel scale.

[0076] Taking the fusion of geospatial data—Normalized Difference Vegetation Index (NDVI) and remote sensing imagery—for example, super-resolution land cover mapping is performed. Figure 5This document illustrates the workflow of fusing remote sensing imagery and Normalized Difference Vegetation Index (NDVI). Multi-temporal NDVI can reflect the growth status of vegetation in different seasons. NDVI can effectively distinguish vegetation from other types of land cover, providing good assistance for the extraction and classification of geoscientific information on vegetation types. Therefore, the super-resolution land cover mapping workflow of NDVI fusion mainly includes: First, acquiring low-resolution remote sensing imagery and MODIS NDVI data and preprocessing them; second, using an improved multi-scale U-Net model to extract remote sensing imagery features and NDVI geoscientific information features respectively; next, calculating the correlation between remote sensing imagery features and NDVI geoscientific information features through an attention mechanism, and multiplying them with the deep features of NDVI to form a location-weighted NDVI feature map; then, superimposing the deep features of the remote sensing imagery with the location-weighted NDVI feature map to form an abstract feature map that fuses NDVI features and remote sensing imagery features; finally, inputting the abstract feature map into the super-resolution classification module to achieve feature map upsampling and classification, obtaining high-resolution land cover mapping results.

[0077] This invention discloses a CNN-based super-resolution land cover mapping method that integrates geoscientific information. It fully leverages the characteristics of deep learning models and geoscientific information, incorporating geoscientific information into the deep learning model. Compared to a single deep learning model, the mapping results obtained using this method are more accurate and have a certain degree of interpretability. Addressing the shortcomings of current geoscientific information integration methods, this invention designs a dual-branch network structure model based on a multi-scale U-Net. By introducing dilated convolution into the U-Net model, multi-scale feature extraction from remote sensing imagery and geoscientific information is achieved. A spatial attention mechanism is introduced into the model, enabling efficient fusion of remote sensing imagery and geoscientific information at specific locations. This solves the problem of neglecting the location of geoscientific information in existing methods, providing a new approach for geoscientific information integration. Using this method, high-resolution, high-precision land cover products can be obtained. It can be applied to various types of geoscientific information for super-resolution land cover mapping, exhibiting strong scalability and contributing to the development and application of remote sensing information extraction technology.

[0078] Example 2

[0079] like Figure 6 As shown, this embodiment of the invention provides a CNN super-resolution land cover mapping system that integrates geoscientific information, including the following modules:

[0080] The multi-scale U-Net semantic feature extraction module 51 is used to acquire remote sensing image data and geoscientific information data with geographic location information. The remote sensing image data and geoscientific information data are input into the improved multi-scale U-Net model to extract the corresponding remote sensing image features and geoscientific information features respectively.

[0081] The attention mechanism position weight calculation module 52 is used to input remote sensing image features and geoscientific information features into the attention mechanism position weight calculation module, calculate the weight of geoscientific information features, and construct a geoscientific information feature map with position weights.

[0082] The feature fusion module 53 is used to overlay the geoscientific information feature map with location weights and the remote sensing image features through the feature fusion module to generate an abstract feature map that integrates geoscientific information and image features.

[0083] The super-resolution classification module 54 is used to perform continuous upsampling operations on the abstract feature map to obtain a feature map with higher spatial resolution at a specific scale, and then use the activation function to obtain the classification results of each sub-pixel, i.e., the super-resolution mapping results.

[0084] The above embodiments are provided merely for the purpose of describing the present invention and are not intended to limit the scope of the invention. The scope of the invention is defined by the appended claims. Various equivalent substitutions and modifications made without departing from the spirit and principles of the invention should be covered within the scope of the invention.

Claims

1. A CNN-based super-resolution land cover mapping method integrating geoscientific information, characterized in that, include: Step S1: Acquire remote sensing image data and geoscientific information data with geographic location information. Input the remote sensing image data and the geoscientific information data into the improved multi-scale U-Net model to extract the corresponding remote sensing image features and geoscientific information features, specifically including: Step S11: Construct the improved multi-scale U-Net model, including an encoder, a dilated convolution module, and a decoder. The encoder has a three-layer structure: the first two layers consist of a convolutional layer, a ReLU layer, and a pooling layer; the third layer consists of a convolutional layer and a ReLU layer. The dilated convolution module consists of three dilated convolutions with different dilation coefficients. The decoder has a two-layer structure: each layer consists of a deconvolutional layer, a convolutional layer, and a ReLU layer. Step S12: Input the remote sensing image data and the geoscientific information data into the encoder of the improved multi-scale U-Net model, and perform feature extraction through the three-layer structure of the encoder: Among them, I RS_en_3 F represents the remote sensing image features extracted by the third layer structure of the encoder. RS_en_1 F represents the first layer structure through which image features are extracted by the encoder. RS_en_2 F represents the second layer structure of the encoder for extracting image features. RS_en_3 This represents the third layer structure from which image features are extracted by the encoder; I rs This represents the input remote sensing image data; I GI_en_3 F represents the geoscientific information features extracted by the third layer structure of the encoder. GI_en_1 F represents the first layer structure of the encoder for extracting geoscientific information features. GI_en_2 F represents the second-layer structure of the encoder for extracting geoscientific information features. GI_en_3 I represents the third layer structure of the encoder for extracting geoscientific information features. gi This represents the input geoscientific information data; Step S13: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and The input to the dilated convolution module utilizes three layers of dilated convolution with dilation coefficients of 1, 2, and 3 to perform multi-scale feature extraction: Among them, I RS_1 This represents the extracted multi-scale remote sensing image features, Contact represents the connection operation, and F RS_MR1 F RS_MR2 F RS_MR3 These represent dilated convolutional features in remote sensing images with dilation rates of 1, 2, and 3, respectively; I GI_1 F represents the extracted multi-scale geoscientific information features. GI_MR1 F GI_MR2 F GI_MR3 These represent geoscientific information feature dilated convolutions with dilation rates of 1, 2, and 3, respectively. Step S14: ... and The input to the decoder undergoes upsampling and multi-level feature fusion. During this process, to fuse features from different levels, the skip connection operation of the U-Net model is preserved to stitch together deep and shallow features of the same resolution. The decoding process is as follows: Among them, I RS_1 For the input multi-scale remote sensing image features, I RS_en_2 For the remote sensing image features extracted from the second layer structure of the encoder, connect represents a skip connection operation, Trans represents a deconvolution operation, Conv represents a convolution operation, and ReLU represents the ReLU activation function; I RS_de_1 This represents the remote sensing image features extracted by the first layer structure of the decoder; I RS_en_1 I represents the remote sensing image features extracted from the first layer structure of the encoder. RS_de_2 This represents the remote sensing image features extracted by the second layer structure of the decoder; I GI_1 For the input multi-scale geoscience information features, I GI_en_2 For the geoscientific information features extracted from the second layer structure of the encoder, I GI_de_1 I represents the geoscientific information features extracted by the first layer structure of the decoder. GI_en_1 For the geoscientific information features extracted from the first layer structure of the encoder, I GI_de_2 This represents the geoscientific information features extracted by the second layer structure of the decoder; Let the final output of the decoding part be represented as follows: Among them, I RS Indicates remote sensing image features, I GI Representing the characteristics of geoscientific information; Step S2: Input the remote sensing image features and geoscientific information features into the attention mechanism location weight calculation module to calculate the weights of the geoscientific information features and construct a geoscientific information feature map with location weights; Step S3: The geoscientific information feature map with location weights is superimposed with the remote sensing image features through the feature fusion module to generate an abstract feature map that fuses geoscientific information and image features; Step S4: Perform continuous upsampling operations on the abstract feature map to obtain a feature map with higher spatial resolution at a specific scale, and then use the activation function to obtain the classification results of each sub-pixel, i.e., the super-resolution mapping results.

2. The CNN super-resolution land cover mapping method integrating geoscientific information according to claim 1, characterized in that, Step S2: Inputting the remote sensing image features and geoscientific information features into the attention mechanism location weight calculation module, calculating the weights of the geoscientific information features, and constructing a geoscientific information feature map with location weights, specifically includes: Step S21: For I RS and I GI Further feature extraction is performed using two consecutive convolutions: in, This indicates the first convolution operation with a kernel size of 32 and a convolution size of 3×3. This indicates a second convolution operation with a kernel size of 1 and a convolution size of 3×3. Step S22: For and Perform location-by-location multiplication to determine the correlation between remote sensing image features and geoscientific information features at each location. SA : Step S23: Use the Sigmoid activation function to activate I. SA Mapped to a value between 0 and 1, the weight value SA represents the weight of the remote sensing image features and the geoscientific information features at each location: Step S24: Transfer the geoscientific information features Multiplying with SA generates a geoscientific feature map W with location weights. GA : Here, Multiply represents the position-by-position multiplication operation.

3. The CNN super-resolution land cover mapping method integrating geoscientific information according to claim 2, characterized in that, Step S3: The location-weighted geoscientific information feature map is superimposed on the remote sensing image features through a feature fusion module to generate an abstract feature map that fuses geoscientific information and image features, specifically including: The remote sensing image feature I RS With W GA Input the feature fusion module: Among them, I RS This represents the features of remote sensing images. Add represents the overlay operation, and Fconc represents the abstract feature map that integrates geoscientific information and image features.

4. The CNN super-resolution land cover mapping method integrating geoscientific information according to claim 3, characterized in that, Step S4: Perform continuous upsampling operations on the abstract feature map to obtain a feature map with higher spatial resolution at a specific scale, and then use an activation function to obtain the classification results of each sub-pixel, i.e., the super-resolution mapping results, specifically including: The abstract feature map, which integrates geoscientific information and image features, is upsampled using the pixel shuffle method to form a high-resolution feature map. The high-resolution feature map is then classified using the sigmoid function to obtain land cover mapping results for each land cover type at the sub-pixel scale. in, For pixel shuffle upsampling operation, r represents the upsampling amplification factor, H Fconc F represents the high-resolution abstract feature map after upsampling. pixel_class This represents the classification result for each sub-pixel of the high-resolution abstract feature map.

5. A CNN super-resolution land cover mapping system integrating geoscientific information, characterized in that, Includes the following modules: A multi-scale U-Net semantic feature extraction module is used to acquire remote sensing image data and geoscientific information data with geographic location information. The remote sensing image data and the geoscientific information data are input into an improved multi-scale U-Net model to extract corresponding remote sensing image features and geoscientific information features, specifically including: Step S11: Construct the improved multi-scale U-Net model, including an encoder, a dilated convolution module, and a decoder. The encoder has a three-layer structure: the first two layers consist of a convolutional layer, a ReLU layer, and a pooling layer; the third layer consists of a convolutional layer and a ReLU layer. The dilated convolution module consists of three dilated convolutions with different dilation coefficients. The decoder has a two-layer structure: each layer consists of a deconvolutional layer, a convolutional layer, and a ReLU layer. Step S12: Input the remote sensing image data and the geoscientific information data into the encoder of the improved multi-scale U-Net model, and perform feature extraction through the three-layer structure of the encoder: Among them, I RS_en_3 F represents the remote sensing image features extracted by the third layer structure of the encoder. RS_en_1 F represents the first layer structure through which image features are extracted by the encoder. RS_en_2 F represents the second layer structure of the encoder for extracting image features. RS_en_3 This represents the third layer structure from which image features are extracted by the encoder; I rs This represents the input remote sensing image data; I GI_en_3 F represents the geoscientific information features extracted by the third layer structure of the encoder. GI_en_1 F represents the first layer structure of the encoder for extracting geoscientific information features. GI_en_2 F represents the second-layer structure of the encoder for extracting geoscientific information features. GI_en_3 I represents the third layer structure of the encoder for extracting geoscientific information features. gi This represents the input geoscientific information data; Step S13: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and The input to the dilated convolution module utilizes three layers of dilated convolution with dilation coefficients of 1, 2, and 3 to perform multi-scale feature extraction: Among them, I RS_1 This represents the extracted multi-scale remote sensing image features, Contact represents the connection operation, and F RS_MR1 F RS_MR2 F RS_MR3 These represent dilated convolutional features in remote sensing images with dilation rates of 1, 2, and 3, respectively; I GI_1 F represents the extracted multi-scale geoscientific information features. GI_MR1 F GI_MR2 F GI_MR3 These represent geoscientific information feature dilated convolutions with dilation rates of 1, 2, and 3, respectively. Step S14: ... and The input to the decoder undergoes upsampling and multi-level feature fusion. During this process, to fuse features from different levels, the skip connection operation of the U-Net model is preserved to stitch together deep and shallow features of the same resolution. The decoding process is as follows: Among them, I RS_1 For the input multi-scale remote sensing image features, I RS_en_2 For the remote sensing image features extracted from the second layer structure of the encoder, connect represents a skip connection operation, Trans represents a deconvolution operation, Conv represents a convolution operation, and ReLU represents the ReLU activation function; I RS_de_1 This represents the remote sensing image features extracted by the first layer structure of the decoder; I RS_en_1 I represents the remote sensing image features extracted from the first layer structure of the encoder. RS_de_2 This represents the remote sensing image features extracted by the second layer structure of the decoder; I GI_1 For the input multi-scale geoscience information features, I GI_en_2 For the geoscientific information features extracted from the second layer structure of the encoder, I GI_de_1 I represents the geoscientific information features extracted by the first layer structure of the decoder. GI_en_1 For the geoscientific information features extracted from the first layer structure of the encoder, I GI_de_2 This represents the geoscientific information features extracted by the second layer structure of the decoder; Let the final output of the decoding part be represented as follows: Among them, I RS Indicates remote sensing image features, I GI Representing the characteristics of geoscientific information; The attention mechanism position weight calculation module is used to input the remote sensing image features and geoscientific information features into the attention mechanism position weight calculation module, calculate the weight of the geoscientific information features, and construct a geoscientific information feature map with position weights. The feature fusion module is used to overlay the location-weighted geoscientific information feature map with the remote sensing image features to generate an abstract feature map that integrates geoscientific information and image features. The super-resolution classification module is used to perform continuous upsampling operations on the abstract feature map to obtain a feature map with higher spatial resolution at a specific scale, and then use an activation function to obtain the classification results of each sub-pixel, i.e., the super-resolution mapping results.

Images