Deep learning remote sensing image analysis method and system for automatic extraction of land use change

Abstract

The application relates to the technical field of image analysis, in particular to a deep learning remote sensing image analysis method and system for automatic extraction of land use change, which comprises the following steps: multi-scale spatial features of double-phase original remote sensing images of a target region are extracted through a convolutional neural network to obtain corresponding deep feature maps; an initial difference feature map is generated through feature vector difference calculation, and a change-sensitive feature tensor is obtained by attention-weighted fusion of the deep feature map; feature enhancement is realized through global average pooling and cross-scale feature integration; edge-guided correction is performed on enhanced feature representation in the up-sampling process according to the edge feature map extracted in the convolutional neural network to obtain a change detection probability map; and a land use change vector map is obtained through contour vectorization of the change detection probability map; and the application can improve the efficiency of deep learning remote sensing image analysis for automatic extraction of land use change.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to a deep learning remote sensing image analysis method and system for automatic extraction of land use change. Background Technology

[0002] Land use change extraction is one of the core requirements in the field of remote sensing applications. When conducting land use change analysis based on remote sensing images, existing technologies have significant shortcomings in the feature extraction stage. They are unable to fully explore the multi-scale spatial features of remote sensing images, and the fusion effect of shallow texture and deep semantic information is poor. They cannot accurately capture the subtle changes between images of different time phases, resulting in a bias in the identification of change areas.

[0003] Existing methods have many shortcomings in handling differential features and utilizing edge information. Differential feature fusion lacks an effective attention weighting mechanism, and cross-scale feature integration is ineffective. At the same time, the upsampling process does not incorporate accurate correction based on land cover edge features, resulting in low accuracy of the change detection probability map. The subsequent vectorized map features have blurred outlines and inaccurate boundaries. The overall accuracy and processing effect of land use change extraction are difficult to meet the requirements of practical applications. Therefore, how to improve the efficiency of automatic land use change extraction has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a deep learning remote sensing image analysis method and system for automatically extracting land use change, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a deep learning remote sensing image analysis method for automatic extraction of land use change, comprising: A1. Input the first and second phase original remote sensing images of the target area into a pre-constructed convolutional neural network to extract multi-scale spatial features, thereby obtaining the first and second depth feature maps of the target area. A2. Calculate the feature vector difference between the first depth feature map and the second depth feature map at corresponding pixel positions, use the feature vector difference as the pixel value to generate an initial difference feature map of the target region, and perform attention-weighted fusion of the initial difference feature map and the first depth feature map to obtain the change-sensitive feature tensor of the target region. A3. Perform global average pooling on the change-sensitive feature tensor to obtain a pooled feature map of the target region, and perform cross-scale feature integration on the pooled feature map and the change-sensitive feature tensor to obtain an enhanced feature representation of the target region; A4. Based on the edge feature map extracted from the convolutional neural network, perform edge-guided correction on the enhanced feature representation during the upsampling process to obtain the change detection probability map of the target region; A5. Perform contour vectorization on the change detection probability map to obtain the land use change vector plot of the target area.

[0006] In a preferred embodiment, the step of inputting the first and second temporal original remote sensing images of the target area into a pre-constructed convolutional neural network for multi-scale spatial feature extraction to obtain a first depth feature map and a second depth feature map of the target area includes: The first and second temporal original remote sensing images of the target area are respectively input into the first convolution group of the pre-constructed convolutional neural network for texture capture, so as to obtain the first shallow feature map and the second shallow feature map of the target area. The first shallow feature map and the second shallow feature map are respectively input into the second convolutional group in the convolutional neural network to extract deep semantic information, thereby obtaining the first intermediate feature map and the second intermediate feature map of the target region; Spatial pyramid pooling is performed on the first intermediate feature map and the second intermediate feature map respectively to obtain the first context-enhanced feature map and the second context-enhanced feature map of the target region. The first context-enhanced feature map and the first shallow feature map are added and fused pixel by pixel, and the fused features are upsampled and restored to obtain the first depth feature map of the target region. The second context-enhanced feature map and the second shallow feature map are subjected to multi-head attention interaction, and global information is embedded into the fused features to obtain the second deep feature map of the target region.

[0007] In a preferred embodiment, the step of calculating the feature vector difference between the first depth feature map and the second depth feature map at corresponding pixel positions, using the feature vector difference as a pixel value to generate an initial difference feature map of the target region, and performing attention-weighted fusion of the initial difference feature map and the first depth feature map to obtain a change-sensitive feature tensor of the target region includes: The feature vectors of the first depth feature map and the second depth feature map at the same spatial location are compared pixel by pixel. The difference value between the two feature vectors at each location is calculated, and the difference value is used as the pixel value at the location to construct the initial difference feature map of the target region. The initial differential feature map is encoded with spatial context information by using the convolutional layers in the pre-constructed attention weight generation network to obtain a spatial attention weight map with the same size as the first depth feature map. The spatial attention weight map is multiplied pixel by pixel with the first depth feature map, and the feature responses at different spatial locations in the first depth feature map are weighted to obtain the weighted first depth feature map of the target region. The weighted first depth feature map is concatenated with the initial difference feature map channel by channel, and the concatenated features are integrated with cross-channel information to obtain the change-sensitive feature tensor of the target region.

[0008] In a preferred embodiment, the formula for calculating the difference value is as follows: ; In the formula, Spatial location in the target area The difference value at that location, The preset scaling factor. This represents the total number of channels in the first depth feature map and the second depth feature map. The first depth feature map in spatial location First The response values ​​of each feature channel, The second depth feature map is located at the same spatial position. First The response values ​​of each feature channel, It is a very small constant. To take the absolute value.

[0009] In a preferred embodiment, the step of encoding spatial context information into the initial differential feature map through convolutional layers in a pre-constructed attention weight generation network to obtain a spatial attention weight map with the same size as the first depth feature map includes: The first convolutional layer in the network is generated by generating pre-constructed attention weights, and spatial features are extracted from the initial difference feature map to obtain the intermediate difference feature map of the initial difference feature map. The intermediate difference feature map is input into the second convolutional layer of the attention weight generation network for channel dimension compression to obtain the preliminary attention map of the intermediate difference feature map; Based on the attention weights, the upsampling layer in the network is generated, and the spatial size of the initial attention map is adjusted to be the same as that of the first depth feature map, so as to obtain an intermediate attention map that matches the size of the first depth feature map. The activation layer in the attention weight generation network normalizes the pixel position values ​​in the intermediate attention map to obtain a spatial attention weight map with the same size as the first depth feature map.

[0010] In a preferred embodiment, the step of performing global average pooling on the change-sensitive feature tensor to obtain a pooled feature map of the target region, and then performing cross-scale feature integration on the pooled feature map and the change-sensitive feature tensor to obtain an enhanced feature representation of the target region, includes: Obtain the spatial dimensions and number of channels of the change-sensitive feature tensor, wherein the spatial dimensions include height and width dimensions; In the height dimension, the change-sensitive feature tensor is summed and pooled to obtain the first intermediate pooled tensor of the change-sensitive feature tensor; Global aggregation is performed on the first intermediate pooling tensor along the width dimension to obtain the second intermediate pooling tensor of the change-sensitive feature tensor; Divide the second intermediate pooling tensor by the product of the height and width dimensions in the spatial dimensions to obtain the global average response value of each feature channel in the change-sensitive feature tensor, and combine the global average response values ​​according to the arrangement order of the feature channels to obtain the pooling feature map of the target region. The spatial size of the pooling feature map is expanded to the same spatial size as the change-sensitive feature tensor to obtain a spatially expanded feature map of the target region; The spatially expanded feature map and the change-sensitive feature tensor are fused element-wise to obtain the enhanced feature representation of the target region.

[0011] In a preferred embodiment, the step of performing edge-guided correction on the enhanced feature representation during the upsampling process based on the edge feature map extracted from the convolutional neural network to obtain the change detection probability map of the target region includes: Obtain the edge feature map output by the convolutional layer in the convolutional neural network, wherein the edge feature map contains spatial location information of the boundaries of ground features in the target region; Subpixel reconstruction is performed on the enhanced feature representation to obtain an upsampled feature map that matches the spatial size of the edge feature map; The edge feature map and the upsampled feature map are subjected to attention modulation to obtain the edge enhancement feature map of the target region; The edge enhancement feature map is iteratively upsampled until the spatial size of the edge enhancement feature map is restored to the original spatial size of the original remote sensing image, thus obtaining the alignment enhancement feature map of the target region; Based on a preset classifier, the feature vectors of pixel positions in the alignment enhancement feature map are used to predict the category probability, thereby obtaining a change detection probability map of the target region.

[0012] In a preferred embodiment, the step of performing attention modulation on the edge feature map and the upsampled feature map to obtain the edge enhancement feature map of the target region includes: Based on the first convolutional layer in the convolutional neural network, spatial features are extracted from the edge feature map to obtain the first intermediate edge feature map of the edge feature map; The activation layer in the convolutional neural network performs a nonlinear transformation on the pixel position values ​​in the first intermediate edge feature map to obtain the edge attention weight map of the first intermediate edge feature map. Based on the weight values ​​of each position in the edge attention weight map, the feature responses at the corresponding positions in the upsampled feature map are gating and adjusted to obtain the initial edge enhancement feature map of the target region; The initial edge enhancement feature map is input into the second convolutional layer of the convolutional neural network for channel recombination to obtain the edge enhancement feature map of the target region.

[0013] In a preferred embodiment, the step of contour vectorizing the change detection probability map to obtain the land use change vector patch of the target area includes: Extract the probability values ​​of all pixel positions in the change detection probability map, and determine the segmentation points of the change detection probability map based on the distribution characteristics of the probability values; The pixel positions with probability values ​​higher than the segmentation point in the change detection probability map are marked as changed regions, and the pixel positions with probability values ​​lower than the segmentation point are marked as unchanged regions, so as to obtain a binary change map of the target region. Perform connectivity analysis on the pixels marked as change regions in the binary change map to obtain the connected regions of the target region; Tracing boundary pixels along the boundary of the connected region and recording the spatial coordinates of the boundary pixels to generate a closed contour line corresponding to the connected region; The closed contour line is used to construct surface features to obtain a land use change vector map of the target area.

[0014] To address the aforementioned problems, this invention also provides a deep learning remote sensing image analysis system for automatic extraction of land use change, the system comprising: The multi-scale feature extraction module is used to input the first and second phase original remote sensing images of the target area into a pre-constructed convolutional neural network to extract multi-scale spatial features, thereby obtaining the first and second depth feature maps of the target area. The difference feature fusion module is used to calculate the feature vector difference between the first depth feature map and the second depth feature map at corresponding pixel positions, use the feature vector difference as the pixel value to generate an initial difference feature map of the target region, and perform attention-weighted fusion of the initial difference feature map and the first depth feature map to obtain the change-sensitive feature tensor of the target region. A cross-scale feature enhancement module is used to perform global average pooling on the change-sensitive feature tensor to obtain a pooled feature map of the target region, and to perform cross-scale feature integration on the pooled feature map and the change-sensitive feature tensor to obtain an enhanced feature representation of the target region. An edge-guided correction module is used to perform edge-guided correction on the enhanced feature representation during the upsampling process based on the edge feature map extracted from the convolutional neural network, so as to obtain a change detection probability map of the target region. The contour vectorization module is used to perform contour vectorization on the change detection probability map to obtain the land use change vector pattern of the target area.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention relies on a pre-constructed convolutional neural network to achieve accurate extraction of multi-scale spatial features. It combines spatial pyramid pooling and multi-head attention interaction to achieve deep feature fusion, which can fully explore the texture and semantic information of remote sensing images. At the same time, the change-sensitive feature tensor generated by feature vector difference calculation and attention weighted fusion can accurately capture the land use change features of the target area, making the representation of change features more targeted and effective, and improving the accuracy of change feature recognition.

[0016] 2. This invention enhances features through global average pooling and cross-scale feature integration. It also precisely guides and corrects the upsampling process by combining edge feature maps extracted by convolutional neural networks, effectively optimizing the accuracy of the change detection probability map and ensuring the clarity and accuracy of land feature boundaries. The land use change vector map obtained by contour vectorization has regular contours and accurate boundaries. The entire automated processing flow realizes the technical optimization of the entire land use change extraction process, significantly improving the overall quality of the extraction results and greatly improving the overall efficiency of land use change analysis. Attached Figure Description

[0017] Figure 1This is a flowchart illustrating a deep learning remote sensing image analysis method for automatic extraction of land use change according to an embodiment of the present invention. Figure 2 This is a functional block diagram of a deep learning remote sensing image analysis system for automatic extraction of land use change provided in an embodiment of the present invention; Figure 3 This is a graph showing the variation of the spatial location of the target area under different total numbers of channels, provided in an embodiment of the present invention. Figure 4 This is a visualization comparison chart of the entire process results of extracting land use change in a target area, provided by an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a deep learning remote sensing image analysis method for automatically extracting land use change. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the deep learning remote sensing image analysis method for automatically extracting land use change can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a deep learning remote sensing image analysis method for automatically extracting land use change according to an embodiment of the present invention. In this embodiment, the deep learning remote sensing image analysis method for automatically extracting land use change includes: A1. Input the first and second phase original remote sensing images of the target area into a pre-constructed convolutional neural network to extract multi-scale spatial features, thereby obtaining the first and second depth feature maps of the target area. In this embodiment of the invention, the step of inputting the first and second temporal original remote sensing images of the target area into a pre-constructed convolutional neural network for multi-scale spatial feature extraction to obtain the first and second depth feature maps of the target area includes: The first and second temporal original remote sensing images of the target area are respectively input into the first convolution group of the pre-constructed convolutional neural network for texture capture, so as to obtain the first shallow feature map and the second shallow feature map of the target area. The first shallow feature map and the second shallow feature map are respectively input into the second convolutional group in the convolutional neural network to extract deep semantic information, thereby obtaining the first intermediate feature map and the second intermediate feature map of the target region; Spatial pyramid pooling is performed on the first intermediate feature map and the second intermediate feature map respectively to obtain the first context-enhanced feature map and the second context-enhanced feature map of the target region. The first context-enhanced feature map and the first shallow feature map are added and fused pixel by pixel, and the fused features are upsampled and restored to obtain the first depth feature map of the target region. The second context-enhanced feature map and the second shallow feature map are subjected to multi-head attention interaction, and global information is embedded into the fused features to obtain the second deep feature map of the target region.

[0021] The first and second temporal original remote sensing images of the target area are respectively input into the first convolutional group of a pre-constructed convolutional neural network. The convolutional group is composed of multiple convolutional layers with fixed-size convolutional kernels connected in sequence. The convolutional kernels slide pixel by pixel in the height and width dimensions of the original remote sensing images. A weighted summation convolution operation is performed on all pixel values ​​within each sliding window. The calculated value is used as the feature response value of the corresponding spatial location. After the convolution operation of the entire original remote sensing image is completed, the first and second shallow feature maps of the target area are output respectively.

[0022] The first and second shallow feature maps are respectively input into the second convolutional group of the convolutional neural network. This convolutional group has more convolutional layers than the first convolutional group and the receptive field of the convolutional kernel is larger. Pixel-by-pixel convolution operation is performed on the input shallow feature maps. During the operation, feature reorganization and information integration are performed on the channel dimension of the feature maps. After the convolution operation of the entire shallow feature map is completed, the first and second intermediate feature maps containing deep semantic information of the ground features in the target area are output respectively.

[0023] Spatial pyramid pooling is performed on the first and second intermediate feature maps respectively, dividing the single feature map into a number of regular rectangular sub-regions according to four preset spatial scales. Aggregation operation is performed on all feature response values ​​in each sub-region to extract global feature information of the sub-region. Then, all sub-region feature information extracted at different scales is spliced ​​and fused according to channel dimension. The fused feature information is restored to the spatial size of the original intermediate feature map, and the first and second context-enhanced feature maps of the target region are output respectively.

[0024] A pixel-by-pixel addition and fusion operation is performed on the first context-enhanced feature map and the first shallow feature map. The feature response values ​​of the two feature maps at exactly the same spatial location are added together, so that the context information of the first context-enhanced feature map and the texture information of the first shallow feature map are fused to obtain the fused feature map. Then, the fused feature map is upsampled and restored by interpolation. The number of pixels in the height and width dimensions of the feature map is adjusted according to the spatial resolution of the original remote sensing image so that the spatial size of the feature map is completely consistent with the spatial size of the original remote sensing image. Finally, the first depth feature map of the target area is output.

[0025] Multi-head attention interaction is performed on the second context-enhanced feature map and the second shallow feature map. The feature information of both feature maps is divided into eight feature subsets according to the channel dimension. Spatial attention weights are calculated for each feature subset. Based on the calculated weight values, the feature responses of each spatial location in each feature subset are weighted and assigned. Then, the results of weight assignment of all feature subsets are concatenated and restored according to the channel dimension to complete the multi-head attention interaction of the two feature maps. Subsequently, a global information embedding operation is performed on the feature map after interaction and fusion. The feature response values ​​of all spatial locations of the feature map are traversed and the global feature information of the whole map is extracted. The global feature information is integrated into the feature vector of each spatial location of the feature map. Finally, the second depth feature map of the target region is output.

[0026] The beneficial effects are as follows: hierarchical convolutional operations enable the stepwise extraction and mining of shallow texture information and deep semantic information from remote sensing images; spatial pyramid pooling operations enrich the contextual information of feature maps at multiple scales; and differentiated fusion strategies are adopted for the two sets of feature maps. The first set of feature maps achieves direct fusion of texture and contextual information, while the second set of feature maps completes refined feature fusion through channel-specific attention interaction. Combined with targeted processing of upsampling restoration and global information embedding, the generated first and second depth feature maps simultaneously retain the shallow texture details of remote sensing images, the deep semantic information of land features, and the global contextual information of the entire image. This makes the depth feature maps more comprehensive in representing the land feature features of the target area, and provides a complete and reliable feature foundation for the accurate extraction of subsequent land use change features.

[0027] A2. Calculate the feature vector difference between the first depth feature map and the second depth feature map at corresponding pixel positions, use the feature vector difference as the pixel value to generate an initial difference feature map of the target region, and perform attention-weighted fusion of the initial difference feature map and the first depth feature map to obtain the change-sensitive feature tensor of the target region. In this embodiment of the invention, the step of calculating the feature vector difference between the first depth feature map and the second depth feature map at corresponding pixel positions, using the feature vector difference as a pixel value to generate an initial difference feature map of the target region, and performing attention-weighted fusion of the initial difference feature map and the first depth feature map to obtain a change-sensitive feature tensor of the target region includes: The feature vectors of the first depth feature map and the second depth feature map at the same spatial location are compared pixel by pixel. The difference value between the two feature vectors at each location is calculated, and the difference value is used as the pixel value at the location to construct the initial difference feature map of the target region. The initial differential feature map is encoded with spatial context information by using the convolutional layers in the pre-constructed attention weight generation network to obtain a spatial attention weight map with the same size as the first depth feature map. The spatial attention weight map is multiplied pixel by pixel with the first depth feature map, and the feature responses at different spatial locations in the first depth feature map are weighted to obtain the weighted first depth feature map of the target region. The weighted first depth feature map is concatenated with the initial difference feature map channel by channel, and the concatenated features are integrated with cross-channel information to obtain the change-sensitive feature tensor of the target region.

[0028] The formula for calculating the difference value is as follows: ; In the formula, Spatial location in the target area The difference value at that location, The preset scaling factor. This represents the total number of channels in the first depth feature map and the second depth feature map. The first depth feature map in spatial location First The response values ​​of each feature channel, The second depth feature map is located at the same spatial position. First The response values ​​of each feature channel, It is a very small constant. To take the absolute value.

[0029] The step of encoding spatial context information into the initial differential feature map through convolutional layers in a pre-constructed attention weight generation network to obtain a spatial attention weight map with the same size as the first depth feature map includes: The first convolutional layer in the network is generated by generating pre-constructed attention weights, and spatial features are extracted from the initial difference feature map to obtain the intermediate difference feature map of the initial difference feature map. The intermediate difference feature map is input into the second convolutional layer of the attention weight generation network for channel dimension compression to obtain the preliminary attention map of the intermediate difference feature map; Based on the attention weights, the upsampling layer in the network is generated, and the spatial size of the initial attention map is adjusted to be the same as that of the first depth feature map, so as to obtain an intermediate attention map that matches the size of the first depth feature map. The activation layer in the attention weight generation network normalizes the pixel position values ​​in the intermediate attention map to obtain a spatial attention weight map with the same size as the first depth feature map.

[0030] Target area spatial location The difference value is obtained by summing the absolute differences and feature response values ​​of all feature channels at the same spatial location in the first and second depth feature maps, multiplying the result by the scaling factor, and then summing the feature response values. This value is a precise quantification of the degree of difference between feature vectors at the same location in the two-temporal depth feature maps. The preset scaling factor is a fixed coefficient determined after calculating and verifying the feature difference values ​​of multiple samples, based on historical experimental data of land use change remote sensing image analysis of different regions, time phases, and land cover types, combined with the numerical distribution law of feature differences in actual remote sensing image analysis scenarios. This coefficient ensures that the calculated difference value range completely matches the input value requirements of the subsequent attention weight generation network. The total number of channels is determined by the network structure design of the pre-constructed convolutional neural network, jointly determined by the number of convolutional layer kernels, feature fusion method, and feature map dimension design. It is a fixed feature map channel parameter in the process of multi-scale spatial feature extraction of the original two-temporal remote sensing images by the convolutional neural network. The first depth feature map... The feature channel response value is a fixed feature representation value formed at the corresponding spatial location and feature channel after the pre-constructed convolutional neural network completes multi-scale spatial feature extraction of the first phase original remote sensing image of the target area. The feature channel response value of the second depth feature map is a fixed feature representation value formed at the corresponding spatial location and feature channel after the convolutional neural network completes multi-scale spatial feature extraction of the second phase original remote sensing image of the target area. The minimum constant is a fixed minimum value that is much smaller than the sum of the feature response values, determined through multiple experiments based on the requirement of avoiding zero in numerical calculations and combined with the actual numerical range of land use change remote sensing image feature value calculations. It is specifically used to avoid the meaningless calculation caused by the occurrence of zero values ​​in the denominator during the calculation of difference values. This formula is the core numerical basis for calculating the difference between feature vectors at the same spatial location in the first and second depth feature maps pixel by pixel. The difference value calculated by the formula is directly used as the pixel value at the corresponding pixel location in the initial difference feature map, becoming the core numerical basis for constructing the initial difference feature map. The formula achieves accurate quantification of the overall difference degree of feature vectors at the same location by performing absolute difference summation and summation operations on the feature response values ​​of all channels in the dual depth feature maps. The scaling factor and the setting of the minimum constant respectively ensure the adaptability of the numerical range of the difference value and the feasibility of the operation. The difference value calculated by this formula can accurately quantify the degree of land use feature change at the corresponding pixel location in the dual temporal remote sensing images, so that the initial difference feature map can truly and accurately reflect the distribution of land use feature differences in the target area, providing an accurate basis for the quantitative difference features for the subsequent attention-weighted fusion of the initial difference feature map and the first depth feature map to generate a change-sensitive feature tensor.

[0031] Iterate through all spatial pixel positions of the first and second depth feature maps sequentially along the height and width dimensions. Extract the feature vectors of the first and second depth feature maps at each same spatial position. That is, extract the feature vector formed by the response values ​​of all feature channels at that position. Based on the above formula, perform channel-by-channel numerical difference calculation on the two feature vectors and integrate them into a single difference value for that spatial position. Assign the difference value corresponding to each spatial position to the same pixel position of the initial difference feature map. After completing the assignment operation for all pixel positions, the initial difference feature map is formed. The height and width pixel count of the initial difference feature map are completely consistent with the height and width pixel count of the first depth feature map.

[0032] The initial difference feature map is input into the first convolutional layer of the pre-constructed attention weight generation network. This convolutional layer uses a 3×3 kernel to slide grid-by-grid across the pixel region of the initial difference feature map with a stride of 1. A weighted summation convolution operation is performed on all feature response values ​​within each sliding window; that is, the weight value of the convolution kernel is multiplied by the corresponding feature response value within the window and then summed to extract the spatial neighborhood feature association information of the initial difference feature map. After the operation, an intermediate difference feature map is output, whose spatial dimensions and number of channels are completely identical to the initial difference feature map. The intermediate difference feature map is then input into the second convolutional layer of the attention weight generation network. The number of convolutional kernels in this layer is precisely set to one-eighth of the number of channels in the intermediate difference feature map. Through convolution operations, the multi-channel features of the intermediate difference feature map are fused and aggregated, compressing the channel dimension of the feature map to one-eighth of the original number of channels, while maintaining the spatial dimensions of the feature map identical to the intermediate difference feature map. Figure 1 Finally, a preliminary attention map is output. After the preliminary attention map is input into the upsampling layer of the attention weight generation network, the upsampling layer uses bilinear interpolation to interpolate and complete the pixels of the preliminary attention map based on the height and width pixel count of the first depth feature map. That is, it calculates the coordinates and corresponding values ​​of the interpolated pixels according to the size of the first depth feature map and completes the completion. The height and width pixel values ​​of the preliminary attention map are gradually adjusted so that the adjusted feature map has a perfect match with the first depth feature map in terms of the number of pixels in height and width, resulting in an intermediate attention map. After the intermediate attention map is input into the activation layer of the attention weight generation network, the activation layer performs linear mapping on the values ​​of all pixel positions in the intermediate attention map, mapping all values ​​proportionally to a fixed value range of 0 to 1, ensuring that the value of each pixel position is within this range and there is no value overflow. After processing, a spatial attention weight map is obtained, and the spatial size of the spatial attention weight map is exactly the same as that of the first depth feature map.

[0033] Iterates through all spatial pixel positions in the spatial attention weight map and the first depth feature map sequentially along the height and width dimensions. For each spatial position, the value of the spatial attention weight map is multiplied by the feature response value of each channel in the corresponding position in the first depth feature map. This multiplication operation assigns targeted weights to the feature responses at different spatial positions in the first depth feature map. After all positions and channels have been processed, a weighted first depth feature map is output. This weighted first depth feature map has the same height, width, number of pixels, and number of channels as the original first depth feature map. All channels of the weighted first depth feature map are arranged in their original order, and then all channels of the initial difference feature map are arranged in their original order to complete the channel-by-channel continuous stitching. The height, width, and number of pixels of the stitched feature map remain unchanged, and the number of channels is the sum of the number of channels of the weighted first depth feature map and the initial difference feature map. Then, a 1×1 convolution kernel is used to perform convolution operation on the stitched feature map with a stride of 1. The feature response values ​​of all channels in each sliding window are weighted and summed to complete the cross-channel feature information interaction and integration. Finally, a change-sensitive feature tensor is output. This change-sensitive feature tensor has a three-dimensional structure, with dimensions of height × width × total number of channels.

[0034] The beneficial effects are as follows: A dedicated formula enables precise quantification of pixel-level differences in dual-temporal depth feature maps. The initial difference feature map constructed by combining pixel-by-pixel traversal and feature vector extraction operations exhibits a feature difference distribution that highly matches the actual land use feature changes in the target area, accurately reflecting pixel-level variations. Through the hierarchical progressive processing of the attention weight generation network, spatial neighborhood association information extraction, channel-dimensional feature simplification, and precise matching of feature map sizes are achieved. Channel compression simplifies feature information while preserving core difference features, bilinear interpolation size adjustment ensures the spatial accuracy of subsequent weight allocation operations, and activation layer normalization keeps attention weight values ​​within a reasonable range of 0 to 1, preventing excessively large or small weights from unduly impacting feature responses. Pixel-by-pixel and channel-by-channel multiplication operations further enhance the effectiveness of the system. The completed weight allocation accurately strengthens the feature responses of areas related to land use change in the first depth feature map, while reasonably weakening the feature responses of non-changed areas, making the feature information more targeted. Finally, through continuous channel-by-channel stitching, the weighted depth feature information and the initial difference feature information are completely preserved. Then, through convolution operation with a 1×1 convolution kernel, the cross-channel feature information interaction and integration are completed, allowing the two types of feature information to complement each other and deeply fuse. The resulting change-sensitive feature tensor not only fully preserves the basic texture and deep semantic feature information of land features, but also accurately highlights the difference features related to land use change. The feature representation is comprehensive and highly accurate, providing targeted and accurate feature support for the subsequent land use change detection process, and providing a solid guarantee for the accuracy of subsequent change detection from the feature level.

[0035] A3. Perform global average pooling on the change-sensitive feature tensor to obtain a pooled feature map of the target region, and perform cross-scale feature integration on the pooled feature map and the change-sensitive feature tensor to obtain an enhanced feature representation of the target region; In this embodiment of the invention, the step of performing global average pooling on the change-sensitive feature tensor to obtain a pooled feature map of the target region, and then performing cross-scale feature integration of the pooled feature map and the change-sensitive feature tensor to obtain an enhanced feature representation of the target region, includes: Obtain the spatial dimensions and number of channels of the change-sensitive feature tensor, wherein the spatial dimensions include height and width dimensions; In the height dimension, the change-sensitive feature tensor is summed and pooled to obtain the first intermediate pooled tensor of the change-sensitive feature tensor; Global aggregation is performed on the first intermediate pooling tensor along the width dimension to obtain the second intermediate pooling tensor of the change-sensitive feature tensor; Divide the second intermediate pooling tensor by the product of the height and width dimensions in the spatial dimensions to obtain the global average response value of each feature channel in the change-sensitive feature tensor, and combine the global average response values ​​according to the arrangement order of the feature channels to obtain the pooling feature map of the target region. The spatial size of the pooling feature map is expanded to the same spatial size as the change-sensitive feature tensor to obtain a spatially expanded feature map of the target region; The spatially expanded feature map and the change-sensitive feature tensor are fused element-wise to obtain the enhanced feature representation of the target region.

[0036] Extract the intrinsic dimensional parameters of the change-sensitive feature tensor to accurately obtain the spatial size and number of channels of the feature tensor. The spatial size is specifically the number of pixels in the height dimension and the number of pixels in the width dimension of the feature tensor, and the number of channels is the total number of features in the channel dimension of the feature tensor. All values ​​are directly taken from the three-dimensional structural properties of the change-sensitive feature tensor.

[0037] For the change-sensitive feature tensor, a summation pooling operation is performed on the height dimension. While keeping the number of channels and the number of pixels in the width dimension of the feature tensor unchanged, the feature response values ​​corresponding to all height dimension pixels at the same width position in each channel are accumulated one by one. The accumulated result is used as the unique feature response value at that width position of that channel. After completing the calculation for all channels and width positions, the first intermediate pooling tensor is obtained. The number of pixels in the height dimension of this tensor is 1, and the number of channels and the number of pixels in the width dimension are consistent with those of the change-sensitive feature tensor.

[0038] A global aggregation operation is performed on the width dimension of the first intermediate pooling tensor. While keeping the number of feature tensor channels unchanged, the feature response values ​​corresponding to all width positions in each channel are accumulated one by one. The accumulated result is used as the unique feature response value of that channel. After the calculation of all channels is completed, the second intermediate pooling tensor is obtained. The number of pixels in the height dimension and the width dimension of this tensor are both 1, and the number of channels is the same as that of the first intermediate pooling tensor.

[0039] The single feature response value corresponding to each channel in the second intermediate pooling tensor is divided by the product of the number of pixels in the height dimension and the number of pixels in the width dimension of the change-sensitive feature tensor. This calculation yields the global average response value of each feature channel in the change-sensitive feature tensor. Then, strictly following the original arrangement order of the feature channels in the change-sensitive feature tensor, the global average response values ​​of all feature channels are combined in an orderly manner to form a pooling feature map of the target region. The number of channels in this pooling feature map is consistent with that of the change-sensitive feature tensor, and the number of pixels in both the height and width dimensions is 1.

[0040] The spatial dimensions of the pooling feature map are adjusted by using dimensional expansion. The single global average response value corresponding to each feature channel in the pooling feature map is mapped unbiasedly to all spatial locations with the same number of pixels in the height and width dimensions as the change-sensitive feature tensor. This ensures that the adjusted feature map is completely consistent with the change-sensitive feature tensor in terms of the number of pixels in height and width, and the number of channels is also the same as the pooling feature map. Finally, a spatially expanded feature map of the target region is obtained.

[0041] An element-wise addition fusion operation is performed on the spatially expanded feature map and the change-sensitive feature tensor. Based on the complete matching of the three-dimensional structures of the two feature data, the feature response values ​​at the same feature channels and the same spatial locations are added one by one. The sum is used as the new feature response value at that channel and location. After completing the addition calculation of all feature channels and all spatial locations, the enhanced feature representation of the target region is obtained. The height, width, number of pixels, and number of channels of the enhanced feature representation are completely consistent with the change-sensitive feature tensor.

[0042] The beneficial effects are as follows: through multidimensional summation pooling and global aggregation operations, the global feature information of each feature channel of the change-sensitive feature tensor can be extracted accurately and completely. The construction of the pooling feature map realizes the condensation and integration of global feature information. The spatial expansion operation allows the condensed global feature information to be completely matched with the local feature information of the original change-sensitive feature tensor in spatial location. The element-wise addition and fusion deeply combines the global context feature information of the whole image with the pixel-level local change-sensitive feature information. This makes the generated enhanced feature representation not only completely retain the original pixel-level local change feature details of the change-sensitive feature tensor, but also incorporate the global feature information of the whole image. This makes the feature representation of land use change in the target area more comprehensive, while strengthening the overall consistency of feature response within the feature channel. This provides a more stable, comprehensive and accurate feature foundation for subsequent edge guidance correction and change detection probability map generation.

[0043] A4. Based on the edge feature map extracted from the convolutional neural network, perform edge-guided correction on the enhanced feature representation during the upsampling process to obtain the change detection probability map of the target region; In this embodiment of the invention, the step of performing edge-guided correction on the enhanced feature representation during the upsampling process based on the edge feature map extracted from the convolutional neural network to obtain the change detection probability map of the target region includes: Obtain the edge feature map output by the convolutional layer in the convolutional neural network, wherein the edge feature map contains spatial location information of the boundaries of ground features in the target region; Subpixel reconstruction is performed on the enhanced feature representation to obtain an upsampled feature map that matches the spatial size of the edge feature map; The edge feature map and the upsampled feature map are subjected to attention modulation to obtain the edge enhancement feature map of the target region; The edge enhancement feature map is iteratively upsampled until the spatial size of the edge enhancement feature map is restored to the original spatial size of the original remote sensing image, thus obtaining the alignment enhancement feature map of the target region; Based on a preset classifier, the feature vectors of pixel positions in the alignment enhancement feature map are used to predict the category probability, thereby obtaining a change detection probability map of the target region.

[0044] The step of performing attention modulation on the edge feature map and the upsampled feature map to obtain the edge enhancement feature map of the target region includes: Based on the first convolutional layer in the convolutional neural network, spatial features are extracted from the edge feature map to obtain the first intermediate edge feature map of the edge feature map; The activation layer in the convolutional neural network performs a nonlinear transformation on the pixel position values ​​in the first intermediate edge feature map to obtain the edge attention weight map of the first intermediate edge feature map. Based on the weight values ​​of each position in the edge attention weight map, the feature responses at the corresponding positions in the upsampled feature map are gating and adjusted to obtain the initial edge enhancement feature map of the target region; The initial edge enhancement feature map is input into the second convolutional layer of the convolutional neural network for channel recombination to obtain the edge enhancement feature map of the target region.

[0045] The edge feature map output by the convolutional layer in the pre-constructed convolutional neural network that is specifically responsible for ground feature edge detection is retrieved. This convolutional layer is the fixed output layer of the convolutional neural network in the multi-scale spatial feature extraction stage. The value of each pixel in the edge feature map is a feature representation value within a fixed range. Pixels with values ​​within the preset boundary threshold range accurately correspond to the actual spatial position of the ground feature boundary in the target area. Pixels in non-ground feature boundary areas are fixed reference values. This feature map completely preserves the spatial position, contour direction and boundary layer information of all ground feature boundaries in the target area.

[0046] Subpixel reconstruction is performed on the enhanced feature representation. The feature pixels of the enhanced feature representation are rearranged in an orderly manner according to the pixel arrangement rules corresponding to the channel dimension and the spatial dimension. During the rearrangement, the information integrity of each feature channel is maintained. Then, the feature value of the completed pixel is calculated by interpolation based on the feature response values ​​of adjacent pixels. The rearrangement and completion are performed simultaneously, and the number of pixels in the height and width of the enhanced feature representation is gradually adjusted until the number of pixels in the height and width of the feature map is completely consistent with the number of pixels in the height and width of the edge feature map, resulting in an upsampled feature map. The number of channels in this feature map is exactly the same as the number of channels in the enhanced feature representation.

[0047] The edge feature map is input into the first convolutional layer of the convolutional neural network. This convolutional layer uses a fixed-size convolutional kernel to slide grid by grid in the pixel region of the edge feature map with a stride of 1. A weighted summation convolution operation is performed on all pixel feature response values ​​within each sliding window. During the operation, only the feature information of the channel dimension is reorganized and enhanced, keeping the spatial size of the feature map unchanged. The spatial neighborhood association information and detailed features of the ground object edges of the edge feature map are extracted. After the convolution operation of all sliding windows is completed, the first intermediate edge feature map is output. The spatial size of this feature map is completely consistent with that of the edge feature map, and the number of channels is fixedly set by the number of convolutional kernels of the first convolutional layer.

[0048] The first intermediate edge feature map is input into the activation layer of the convolutional neural network. The activation layer performs a non-linear transformation operation on the values ​​of all spatial pixel positions in the feature map, mapping all pixel values ​​to a fixed value range of 0 to 1 according to a fixed transformation rule. During the transformation process, the pixel values ​​at the boundary of the ground object are enhanced, while the pixel values ​​at the non-boundary of the ground object are weakened, ensuring that the value of each pixel position after the transformation is in the range of 0 to 1 without any value overflow. After processing, an edge attention weight map is obtained, and the spatial size and number of channels of the weight map are completely consistent with the first intermediate edge feature map.

[0049] Iterates through all spatial pixel positions of the edge attention weight map and the upsampled feature map along the height and width dimensions, while simultaneously iterating through all feature channels. The weight value of each spatial position in the edge attention weight map is multiplied one by one with the feature response value of the corresponding spatial position and feature channel in the upsampled feature map. This multiplication operation achieves gating adjustment of the feature response of the upsampled feature map. The closer the weight value is to 1, the higher the enhancement of the feature response at the corresponding position; the closer the weight value is to 0, the higher the weakening of the feature response at the corresponding position. After completing the calculations for all spatial positions and feature channels, an initial edge enhancement feature map is obtained. The spatial size and number of channels of this feature map are completely consistent with the upsampled feature map.

[0050] The initial edge enhancement feature map is input into the pre-defined second convolutional layer in the convolutional neural network. This convolutional layer uses a 1×1 convolutional kernel to slide grid by grid across the pixel region of the initial edge enhancement feature map with a stride of 1. A weighted summation operation is performed on the multi-channel feature response values ​​within each sliding window. While keeping the spatial size of the feature map unchanged, the feature information in the channel dimension is recombined and integrated, redundant feature information in the channel dimension is removed, and channel feature information related to the edge of land cover and land use change is enhanced. After completing the convolution operation of all sliding windows, the edge enhancement feature map is output. The spatial size of this feature map is exactly the same as that of the initial edge enhancement feature map, and the number of channels is fixed by the number of convolutional kernels in the second convolutional layer.

[0051] An iterative upsampling operation is performed on the edge enhancement feature map. Each upsampling uses bilinear interpolation to adjust the height and width of the feature map by a fixed factor of 2. During each upsampling process, the feature values ​​of the supplementary pixels are calculated based on the feature response values ​​of adjacent pixels to maintain the integrity of the information of the edge features and land use change features without loss of features. This upsampling operation is continued until the height and width of the feature map are completely consistent with the height and width of the original remote sensing images of the first and second time phases of the target area. The iterative operation is then stopped and an aligned enhancement feature map is obtained. The number of channels in this feature map is exactly the same as the number of channels in the edge enhancement feature map.

[0052] The alignment-enhanced feature map is input into a pre-built preset classifier. This classifier is a classification model trained and validated using a large number of land use change remote sensing image samples from different regions, time periods, and land cover types. The classifier extracts the full-channel feature response values ​​of each pixel location in the alignment-enhanced feature map and combines them into a complete feature vector. The feature vector of each pixel location is used to calculate the dual-class probability of land use change and non-change, and outputs the probability values ​​of the change category and the non-change category for each pixel location respectively. The probability values ​​of the change category of all pixel locations are used as the pixel values ​​of the corresponding pixels and arranged in the spatial order of the original remote sensing image to obtain the change detection probability map of the target area. The spatial size of this probability map is completely consistent with the original remote sensing image, and the pixel values ​​are all land use change probability representation values ​​in the range of 0 to 1.

[0053] The beneficial effects are as follows: the edge feature maps directly retrieved from the convolutional neural network provide accurate spatial location information for land cover boundary analysis; sub-pixel reconstruction achieves precise matching between the enhanced feature representation and the spatial size of the edge feature map; the edge features are converted into attention weights in the range of 0 to 1 through the processing of convolutional and activation layers, allowing gating adjustment operations to accurately enhance the response of land cover edge features in the upsampled feature map; the channel reorganization of the second convolutional layer removes redundant features and enhances the core edges and change features; iterative upsampling gradually restores the feature map to the size of the original remote sensing image at a fixed magnification, maintaining the integrity of feature information without loss throughout the process; the preset classifier performs pixel-level category probability prediction on the aligned enhanced feature map based on the trained model; the spatial size of the generated change detection probability map is completely aligned with the original remote sensing image; the pixel values ​​can accurately represent the land use change probability at each location; the entire process deeply integrates land cover edge features and land use change features, making the change detection results more consistent with the actual situation of the target area, and providing a high-precision, spatially aligned probabilistic feature foundation for subsequent contour vectorization operations.

[0054] A5. Perform contour vectorization on the change detection probability map to obtain the land use change vector plot of the target area.

[0055] In this embodiment of the invention, the step of contour vectorizing the change detection probability map to obtain the land use change vector patch of the target area includes: Extract the probability values ​​of all pixel positions in the change detection probability map, and determine the segmentation points of the change detection probability map based on the distribution characteristics of the probability values; The pixel positions with probability values ​​higher than the segmentation point in the change detection probability map are marked as changed regions, and the pixel positions with probability values ​​lower than the segmentation point are marked as unchanged regions, so as to obtain a binary change map of the target region. Perform connectivity analysis on the pixels marked as change regions in the binary change map to obtain the connected regions of the target region; Tracing boundary pixels along the boundary of the connected region and recording the spatial coordinates of the boundary pixels to generate a closed contour line corresponding to the connected region; The closed contour line is used to construct surface features to obtain a land use change vector map of the target area.

[0056] Traverse all spatial pixel locations of the change detection probability map, extract the land use change probability value corresponding to each pixel location, arrange all extracted probability values ​​in order of numerical value and count the frequency of occurrence of different values ​​to form a complete probability value frequency distribution feature. Based on this distribution feature, select the value where the frequency distribution changes abruptly as the segmentation point of the change detection probability map. The value of this segmentation point is in the range of 0 to 1 of the pixel value of the change detection probability map, which is a fixed judgment value to distinguish the changed area from the non-changed area.

[0057] The probability value of each pixel in the change detection probability map is compared one by one with the value of the determined segmentation point. Pixels with probability values ​​higher than the segmentation point value are uniformly assigned a value of 1 and marked as changed areas, while pixels with probability values ​​lower than the segmentation point value are uniformly assigned a value of 0 and marked as unchanged areas. After completing the assignment of values ​​and marking of areas for all pixel locations, a binary change map of the target area is formed. The spatial size of this binary change map is completely consistent with the spatial size of the change detection probability map and the original remote sensing image.

[0058] The 4-neighborhood connectivity rule is used to perform connectivity analysis on all pixels marked as change regions and assigned a value of 1 in the binary change map. It iterates through all pixels assigned a value of 1 in the binary change map and checks the assignment results of the four adjacent pixels above, below, left, and right of each pixel. If the adjacent pixels are assigned a value of 1, they are determined to be connected. All mutually connected pixels assigned a value of 1 are grouped into the same region. Each independent connected region is assigned a unique region identifier. After completing the classification and identification of all pixels assigned a value of 1, all connected regions of the target region are obtained.

[0059] For each independently connected region that has been identified, perform boundary pixel tracking. Starting from the first boundary pixel in the upper left corner of the connected region, determine and track each point along the edge of the region in a clockwise direction. Use the existence of non-changing region pixels with a value of 0 among adjacent pixels as the basis for determining the boundary pixel. Record the planar spatial coordinates corresponding to each boundary pixel point. Continue tracking until the starting boundary pixel point is reached and then stop the operation. Arrange and connect all the recorded boundary pixel spatial coordinates in the actual tracking order to generate the closed contour line corresponding to the connected region.

[0060] Following the vector data surface feature construction specifications, for each connected region's corresponding closed contour line, the sequentially arranged spatial coordinate sequence of boundary pixels is converted into the geometric features of the vector surface. Each converted vector surface feature is assigned corresponding basic attribute information, including a unique regional identifier, contour spatial coordinate range, and number of pixels. After completing the construction of vector surface features and attribute information assignment for all connected regions, all vector surface features are integrated to obtain the land use change vector patch of the target region. The spatial coordinate system of this vector patch is completely consistent with the spatial coordinate system of the original remote sensing image.

[0061] The beneficial effects are as follows: determining the segmentation point by using the frequency distribution characteristics of statistical probability values ​​makes the binarization of changed and unchanged areas more closely match the actual numerical distribution characteristics of the change detection probability map, ensuring the accuracy of the binary change map division; the application of the 4-neighborhood connectivity rule enables connectivity analysis to accurately identify and divide all independent land use change areas, avoiding the situation of area adhesion or incomplete splitting; the clockwise boundary tracking combined with clear boundary pixel point determination criteria ensures the integrity, continuity and accuracy of the closed contour line; the construction of surface elements based on fixed specifications allows pixel-level change areas to be successfully converted into standardized vector patches, and the vector patches are aligned with the spatial coordinates of the original remote sensing image and have complete attribute information. The final generated land use change vector patches can accurately present the spatial location, actual range and morphological characteristics of land use change in the target area, providing a regular, accurate and editable standardized vector data foundation for subsequent statistics, analysis and application of land use change.

[0062] like Figure 2 The diagram shown is a functional block diagram of a deep learning remote sensing image analysis system for automatic extraction of land use change provided in an embodiment of the present invention.

[0063] The deep learning remote sensing image analysis system 100 for automatic land use change extraction described in this invention can be installed in an electronic device. Depending on the functions implemented, the deep learning remote sensing image analysis system 100 for automatic land use change extraction may include a multi-scale feature extraction module 101, a differential feature fusion module 102, a cross-scale feature enhancement module 103, an edge-guided correction module 104, and a contour vectorization module 105. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and which are stored in the memory of the electronic device.

[0064] In this embodiment, the functions of each module / unit are as follows: The multi-scale feature extraction module 101 is used to input the first temporal original remote sensing image and the second temporal original remote sensing image of the target area into a pre-constructed convolutional neural network to extract multi-scale spatial features, thereby obtaining the first depth feature map and the second depth feature map of the target area. The differential feature fusion module 102 is used to calculate the feature vector difference between the first depth feature map and the second depth feature map at corresponding pixel positions, use the feature vector difference as a pixel value to generate an initial differential feature map of the target region, and perform attention-weighted fusion of the initial differential feature map and the first depth feature map to obtain the change-sensitive feature tensor of the target region. The cross-scale feature enhancement module 103 is used to perform global average pooling on the change-sensitive feature tensor to obtain a pooled feature map of the target region, and to perform cross-scale feature integration on the pooled feature map and the change-sensitive feature tensor to obtain an enhanced feature representation of the target region. The edge guidance correction module 104 is used to perform edge guidance correction on the enhanced feature representation during the upsampling process based on the edge feature map extracted from the convolutional neural network, so as to obtain the change detection probability map of the target region; The contour vectorization module 105 is used to perform contour vectorization on the change detection probability map to obtain the land use change vector pattern of the target area.

[0065] like Figure 3 The figure shown is a curve illustrating the variation of spatial location differences in the target area under different total number of channels according to an embodiment of the present invention. It serves as a verification graph for key parameters in the land use change difference calculation process of the present invention. The horizontal axis represents the spatial location number of the target area from 0 to 30, and the vertical axis represents the pixel-level land use feature difference value calculated using the invention's proprietary formula. The three curves correspond to the three cases of the total number of channels in the depth feature map, B=8, B=16, and B=32, respectively. They intuitively show the influence and fluctuation characteristics of the different total number of channels, a core parameter of the convolutional neural network, on the quantitative value of the difference in land use characteristics in two temporal phases at different spatial locations in the target area. The experimental data presented also provide a scientific basis for the design of the number of channels in the convolutional neural network, helping to achieve accurate quantification of the difference value.

[0066] like Figure 4The figure shown is a visualization comparison of the entire process results of land use change extraction in a target area provided by an embodiment of the present invention. It is a core figure for verifying the effectiveness of the method of the present invention. The figure integrates three core layers: the original remote sensing image of the target area in two time phases, the change detection probability map, and the land use change vector patch. It sequentially presents the basic data for land use change extraction, the probability distribution of land use change in each pixel interval of 0-1 in the target area, and the final extraction result presented as closed surface elements. Through the progressive visualization display from the original image to the probability map and then to the vector patch, it intuitively demonstrates that the method of the present invention can generate a change detection probability map that is spatially aligned with the original image and has clear land feature boundaries. Moreover, the final land use change vector patch has a regular outline and accurate boundaries, which can accurately match the actual land use change situation in the target area. This fully verifies the effectiveness and superiority of the method of the present invention in automatic land use change extraction.

[0067] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0068] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0069] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0070] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0071] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based remote sensing image analysis method for automatic extraction of land use change, characterized in that, The method includes: A1. Input the first and second phase original remote sensing images of the target area into a pre-constructed convolutional neural network to extract multi-scale spatial features, thereby obtaining the first and second depth feature maps of the target area. A2. Calculate the feature vector difference between the first depth feature map and the second depth feature map at corresponding pixel positions, use the feature vector difference as the pixel value to generate an initial difference feature map of the target region, and perform attention-weighted fusion of the initial difference feature map and the first depth feature map to obtain the change-sensitive feature tensor of the target region. A3. Perform global average pooling on the change-sensitive feature tensor to obtain a pooled feature map of the target region, and perform cross-scale feature integration on the pooled feature map and the change-sensitive feature tensor to obtain an enhanced feature representation of the target region; A4. Based on the edge feature map extracted from the convolutional neural network, perform edge-guided correction on the enhanced feature representation during the upsampling process to obtain the change detection probability map of the target region; A5. Perform contour vectorization on the change detection probability map to obtain the land use change vector plot of the target area.

2. The deep learning remote sensing image analysis method for automatic extraction of land use change as described in claim 1, characterized in that, The process of inputting the first and second temporal original remote sensing images of the target area into a pre-constructed convolutional neural network for multi-scale spatial feature extraction, thereby obtaining a first depth feature map and a second depth feature map of the target area, includes: The first and second temporal original remote sensing images of the target area are respectively input into the first convolution group of the pre-constructed convolutional neural network for texture capture, so as to obtain the first shallow feature map and the second shallow feature map of the target area. The first shallow feature map and the second shallow feature map are respectively input into the second convolutional group in the convolutional neural network to extract deep semantic information, thereby obtaining the first intermediate feature map and the second intermediate feature map of the target region; Spatial pyramid pooling is performed on the first intermediate feature map and the second intermediate feature map respectively to obtain the first context-enhanced feature map and the second context-enhanced feature map of the target region. The first context-enhanced feature map and the first shallow feature map are added and fused pixel by pixel, and the fused features are upsampled and restored to obtain the first depth feature map of the target region. The second context-enhanced feature map and the second shallow feature map are subjected to multi-head attention interaction, and global information is embedded into the fused features to obtain the second deep feature map of the target region.

3. The deep learning remote sensing image analysis method for automatic extraction of land use change as described in claim 1, characterized in that, The process of calculating the feature vector difference between the first depth feature map and the second depth feature map at corresponding pixel positions, using the feature vector difference as a pixel value to generate an initial difference feature map of the target region, and then performing attention-weighted fusion of the initial difference feature map and the first depth feature map to obtain a change-sensitive feature tensor of the target region includes: The feature vectors of the first depth feature map and the second depth feature map at the same spatial location are compared pixel by pixel. The difference value between the two feature vectors at each location is calculated, and the difference value is used as the pixel value at the location to construct the initial difference feature map of the target region. The initial differential feature map is encoded with spatial context information by using the convolutional layers in the pre-constructed attention weight generation network to obtain a spatial attention weight map with the same size as the first depth feature map. The spatial attention weight map is multiplied pixel by pixel with the first depth feature map, and the feature responses at different spatial locations in the first depth feature map are weighted to obtain the weighted first depth feature map of the target region. The weighted first depth feature map is concatenated with the initial difference feature map channel by channel, and the concatenated features are integrated with cross-channel information to obtain the change-sensitive feature tensor of the target region.

4. The deep learning remote sensing image analysis method for automatic extraction of land use change as described in claim 3, characterized in that, The formula for calculating the difference value is as follows: ; In the formula, Spatial location in the target area The difference value at that location, The preset scaling factor. This represents the total number of channels in the first depth feature map and the second depth feature map. The first depth feature map in spatial location First The response values ​​of each feature channel, The second depth feature map is located at the same spatial position. First The response values ​​of each feature channel, It is a very small constant. To take the absolute value.

5. The deep learning remote sensing image analysis method for automatic extraction of land use change as described in claim 3, characterized in that, The step of encoding spatial context information into the initial differential feature map through convolutional layers in a pre-constructed attention weight generation network to obtain a spatial attention weight map with the same size as the first depth feature map includes: The first convolutional layer in the network is generated by generating pre-constructed attention weights, and spatial features are extracted from the initial difference feature map to obtain the intermediate difference feature map of the initial difference feature map. The intermediate difference feature map is input into the second convolutional layer of the attention weight generation network for channel dimension compression to obtain the preliminary attention map of the intermediate difference feature map. Based on the attention weights, the upsampling layer in the network is generated, and the spatial size of the initial attention map is adjusted to be the same as that of the first depth feature map, so as to obtain an intermediate attention map that matches the size of the first depth feature map. The activation layer in the attention weight generation network normalizes the pixel position values ​​in the intermediate attention map to obtain a spatial attention weight map with the same size as the first depth feature map.

6. The deep learning remote sensing image analysis method for automatic extraction of land use change as described in claim 1, characterized in that, The process involves performing global average pooling on the change-sensitive feature tensor to obtain a pooled feature map of the target region, and then integrating the pooled feature map with the change-sensitive feature tensor across scales to obtain an enhanced feature representation of the target region, including: Obtain the spatial dimensions and number of channels of the change-sensitive feature tensor, wherein the spatial dimensions include the height dimension and the width dimension; In the height dimension, the change-sensitive feature tensor is summed and pooled to obtain the first intermediate pooled tensor of the change-sensitive feature tensor; Global aggregation is performed on the first intermediate pooling tensor along the width dimension to obtain the second intermediate pooling tensor of the change-sensitive feature tensor; Divide the second intermediate pooling tensor by the product of the height and width dimensions in the spatial dimensions to obtain the global average response value of each feature channel in the change-sensitive feature tensor, and combine the global average response values ​​according to the arrangement order of the feature channels to obtain the pooling feature map of the target region. The spatial size of the pooling feature map is expanded to the same spatial size as the change-sensitive feature tensor to obtain a spatially expanded feature map of the target region; The spatially expanded feature map and the change-sensitive feature tensor are fused element-wise to obtain the enhanced feature representation of the target region.

7. The deep learning remote sensing image analysis method for automatic extraction of land use change as described in claim 1, characterized in that, The step of performing edge-guided correction on the enhanced feature representation during the upsampling process based on the edge feature map extracted from the convolutional neural network to obtain the change detection probability map of the target region includes: Obtain the edge feature map output by the convolutional layer in the convolutional neural network, wherein the edge feature map contains spatial location information of the boundaries of ground features in the target region; Subpixel reconstruction is performed on the enhanced feature representation to obtain an upsampled feature map that matches the spatial size of the edge feature map; The edge feature map and the upsampled feature map are subjected to attention modulation to obtain the edge enhancement feature map of the target region; The edge enhancement feature map is iteratively upsampled until the spatial size of the edge enhancement feature map is restored to the original spatial size of the original remote sensing image, thus obtaining the alignment enhancement feature map of the target region; Based on a preset classifier, the feature vectors of pixel positions in the alignment enhancement feature map are used to predict the category probability, thereby obtaining a change detection probability map of the target region.

8. The deep learning remote sensing image analysis method for automatic extraction of land use change as described in claim 7, characterized in that, The step of performing attention modulation on the edge feature map and the upsampled feature map to obtain the edge enhancement feature map of the target region includes: Based on the first convolutional layer in the convolutional neural network, spatial features are extracted from the edge feature map to obtain the first intermediate edge feature map of the edge feature map; The activation layer in the convolutional neural network performs a nonlinear transformation on the pixel position values ​​in the first intermediate edge feature map to obtain the edge attention weight map of the first intermediate edge feature map. Based on the weight values ​​of each position in the edge attention weight map, the feature responses at the corresponding positions in the upsampled feature map are gating and adjusted to obtain the initial edge enhancement feature map of the target region; The initial edge enhancement feature map is input into the second convolutional layer of the convolutional neural network for channel recombination to obtain the edge enhancement feature map of the target region.

9. The deep learning remote sensing image analysis method for automatic extraction of land use change as described in claim 1, characterized in that, The process of contour vectorizing the change detection probability map to obtain the land use change vector patch of the target area includes: Extract the probability values ​​of all pixel positions in the change detection probability map, and determine the segmentation points of the change detection probability map based on the distribution characteristics of the probability values; The pixel positions with probability values ​​higher than the segmentation point in the change detection probability map are marked as changed regions, and the pixel positions with probability values ​​lower than the segmentation point are marked as unchanged regions, so as to obtain a binary change map of the target region. Perform connectivity analysis on the pixels marked as change regions in the binary change map to obtain the connected regions of the target region; Tracing boundary pixels along the boundary of the connected region and recording the spatial coordinates of the boundary pixels to generate a closed contour line corresponding to the connected region; The closed contour line is used to construct surface features to obtain a land use change vector map of the target area.

10. A deep learning remote sensing image analysis system for automatic extraction of land use change, characterized in that, The system for implementing the deep learning remote sensing image analysis method for automatic extraction of land use change as described in claim 1 includes: The multi-scale feature extraction module is used to input the first and second phase original remote sensing images of the target area into a pre-constructed convolutional neural network to extract multi-scale spatial features, thereby obtaining the first and second depth feature maps of the target area. The difference feature fusion module is used to calculate the feature vector difference between the first depth feature map and the second depth feature map at corresponding pixel positions, use the feature vector difference as the pixel value to generate an initial difference feature map of the target region, and perform attention-weighted fusion of the initial difference feature map and the first depth feature map to obtain the change-sensitive feature tensor of the target region. A cross-scale feature enhancement module is used to perform global average pooling on the change-sensitive feature tensor to obtain a pooled feature map of the target region, and to perform cross-scale feature integration on the pooled feature map and the change-sensitive feature tensor to obtain an enhanced feature representation of the target region. An edge-guided correction module is used to perform edge-guided correction on the enhanced feature representation during the upsampling process based on the edge feature map extracted from the convolutional neural network, so as to obtain a change detection probability map of the target region. The contour vectorization module is used to perform contour vectorization on the change detection probability map to obtain the land use change vector pattern of the target area.

Images