A remote sensing image change detection method based on deep learning

Abstract

The present invention relates to the field of computer technology, and specifically discloses a remote sensing image change detection method based on deep learning. The method optimizes a target remote sensing image set through a first optimization model in a deep learning method, obtains attribute parameters of the optimized target remote sensing image set, evaluates the data quality of the target remote sensing image set, extracts features based on the target remote sensing image set with high data quality, analyzes the feature accuracy of a pixel-level feature set, and warns of abnormal changes in the target optimized remote sensing image set based on high-precision features, thereby achieving accurate warning of abnormal changes in the target optimized remote sensing image set, providing strong technical support for remote sensing monitoring and analysis, and effectively improving the accuracy and efficiency of change detection.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a remote sensing image change detection method based on deep learning. Background Art

[0002] With the development of remote sensing technology, different types of satellites have been launched into space, enabling us to obtain multi-platform, multi-sensor, and multi-angle remote sensing images. Remote sensing images have the advantages of high temporal resolution, easy computer processing, and a wide range of spatial and spectral resolution options. In the past few decades, they have become the main data source for different change detection applications.

[0003] For example, the invention patent with announcement number CN114155200B announces a remote sensing image change detection method, device, equipment and computer-readable storage medium based on a convolutional neural network. The method includes obtaining continuous multi-period remote sensing images; inputting the continuous multi-period remote sensing images into a change detection model to determine the change area of ​​the target object; the change area is an area including the increase and decrease attributes of the target object; and calibrating the change area to obtain the detection result of the remote sensing image.

[0004] For example, the invention patent with publication number CN117494765A discloses a twin network and method for change detection of remote sensing images with ultra-high spatial resolution, and provides a change detection network MDSANet, which uses asymmetric convolution residual blocks to extract deeper features with more discriminative change information, and uses the spatiotemporal attention mechanism to enhance the fused multi-scale difference features in the spatiotemporal domain to enhance the areas of real changes and suppress the areas of unreal changes. At the same time, the correlation between the channels of feature maps of different scales is established to improve the representation ability of features of different scales.

[0005] However, in the process of implementing the embodiments of the present application, it was found that the above-mentioned technology has at least the following technical problems: in the process of analyzing the changes of remote sensing images, the existing technology directly extracts the features in the remote sensing images and performs change analysis. The default processing method of extracting without error may cause the accuracy of subsequent image change analysis to be affected, thereby causing inaccurate analysis results. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a remote sensing image change detection method based on deep learning, which can effectively solve the problems involved in the above-mentioned background technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a remote sensing image change detection method based on deep learning, comprising: step 1, optimizing the target remote sensing image set through the first optimization model in the deep learning method, obtaining the attribute parameters of the optimized target remote sensing image set, evaluating the data quality index of the target remote sensing image set, and comparing it with the data quality threshold; step 2, if the data quality index of the target remote sensing image set is greater than the data quality threshold, marking the optimized target remote sensing image set as the target optimized remote sensing image set, performing feature detection and extraction on the target optimized remote sensing image set through the second feature extraction model in the deep learning method, and marking it as the pixel-level feature set of the target optimized remote sensing image set, obtaining the attribute parameters of the pixel-level feature set, analyzing the feature accuracy factor of the pixel-level feature set, and comparing it with the feature accuracy threshold; step 3, if the feature accuracy factor of the pixel-level feature set is greater than the feature accuracy threshold, analyzing the abnormal change factor of the target optimized remote sensing image set according to the pixel-level feature set of the target optimized remote sensing image set, thereby giving an early warning of the abnormal change of the target optimized remote sensing image set.

[0008] As a further method, the attribute parameters of the optimized target remote sensing image set specifically include the grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, the spectral band probability curve of each sample image belonging to the target remote sensing image set, and the discrete value of the pixel value of each sample area of ​​each sample image belonging to the target remote sensing image set.

[0009] As a further method, the comparison is performed with the data quality threshold, and the specific comparison process is:

[0010] If the data quality index of the target remote sensing image set is less than or equal to the data quality threshold, the data quality threshold and the data quality index of the target remote sensing image set are subjected to difference processing, and the difference processing result is marked as the data quality index deviation value of the target remote sensing image set; according to the data quality index deviation value of the target remote sensing image set, an adjustment set of the first optimization model is matched from the spatiotemporal management information library, the first optimization model is adjusted according to the adjustment set of the first optimization model, and the target remote sensing image set is optimized by the adjusted first optimization model until the data quality index of the target remote sensing image set is greater than the data quality threshold, and the optimized target remote sensing image set is marked as the target optimized remote sensing image set.

[0011] As a further method, the comparison is performed with a feature accuracy threshold, and the specific analysis process is as follows: if the feature accuracy factor of the pixel-level feature set is less than or equal to the feature accuracy threshold, the feature accuracy threshold and the feature accuracy factor of the pixel-level feature set are subjected to difference processing, and the difference processing result is marked as the feature accuracy factor deviation value of the pixel-level feature set; according to the feature accuracy factor deviation value of the pixel-level feature set, an adjustment set of a second feature extraction model is matched from a spatiotemporal management information library, the second feature extraction model is adjusted according to the adjustment set of the second feature extraction model, and feature detection and extraction are performed on the target optimized remote sensing image set through the adjusted second feature extraction model, and the feature accuracy factor of the pixel-level feature set is adjusted to be greater than the feature accuracy threshold, and the abnormal change factor of the target optimized remote sensing image set is analyzed according to the pixel-level feature set of the target optimized remote sensing image set.

[0012] As a further method, the abnormal changes of the target optimized remote sensing image set are warned. The specific analysis process is: comparing the abnormal change factor of the target optimized remote sensing image set with the abnormal change threshold. If the abnormal change factor of the target optimized remote sensing image set is greater than the abnormal change threshold, then a warning is issued for the abnormal changes of the target optimized remote sensing image set.

[0013] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0014] (1) The present invention provides a remote sensing image change detection method based on deep learning. The target remote sensing image set is optimized through the first optimization model in the deep learning method, and the attribute parameters of the optimized target remote sensing image set are obtained, the data quality of the target remote sensing image set is evaluated, and the features therein are extracted based on the target remote sensing image set with high data quality, and the feature accuracy of the pixel-level feature set is analyzed. Based on the high-precision features, an early warning is issued for abnormal changes in the target optimized remote sensing image set, thereby achieving accurate early warning of abnormal changes in the target optimized remote sensing image set, providing strong technical support for remote sensing monitoring and analysis, and effectively improving the accuracy and efficiency of change detection.

[0015] (2) The present invention optimizes the data quality of the target remote sensing image set. The improvement in data quality directly enhances the accuracy and clarity of the image information, allowing the details and key information in the image to be better presented. This optimization not only provides a more reliable basis for the subsequent processing of remote sensing data, but also significantly improves the accuracy and reliability of data analysis.

[0016] (3) By analyzing and optimizing the feature detection and extraction of the target optimized remote sensing image set, the present invention can more accurately identify and extract key features in the image, such as texture shape, etc., providing a richer and more accurate information basis for subsequent image analysis and recognition. Secondly, in the feature extraction process, through the detailed analysis and optimization of the features, the stability and robustness of the features can be significantly improved, and the influence of noise and interference on feature extraction can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0018] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0019] Figure 2 It is a schematic diagram of the comparison diagram before and after optimization involved in the present invention.

[0020] Figure 3 The figure is a schematic diagram of the information extraction process involved in the present invention.

[0021] Figure 4 It is a schematic diagram of the management system structure of the spatiotemporal information database involved in the present invention.

[0022] Figure 5 It is a schematic diagram of a probability curve of a certain spectral band involved in the present invention. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0024] Reference Figure 1 As shown, the present invention provides a remote sensing image change detection method based on deep learning, including: step one, optimizing the target remote sensing image set through the first optimization model in the deep learning method, obtaining the attribute parameters of the optimized target remote sensing image set, evaluating the data quality index of the target remote sensing image set, and comparing it with the data quality threshold.

[0025] In an example embodiment, the first optimization model in the above-mentioned deep learning method specifically refers to: first, using the Self-Attentive Generative Adversarial Network model, combining the generative adversarial network and the attention mechanism, effectively distinguishing the features of cloud images and cloud-free images in the target remote sensing image set, and generating realistic cloud-free images by introducing the visual attention mechanism, thereby solving the problem of missing farmland information caused by cloud occlusion. The training of the self-attention generative adversarial network follows the guiding principles of expanding the attention to replace the cloud area, reducing the attention to match the cloud boundary, and optimizing the attention to handle extreme situations, and only requires images and image-level labels as input, which reduces costs and improves efficiency. In addition, the adversarial network is transformed and a cloud distortion model (CR-GAN-PM) semi-supervised method is used. This method combines The combination of generative adversarial networks and cloud distortion physical models can remove thin clouds in remote sensing images, ensuring that the generated background layer is both cloud-free and relevant to the input image, further improving the clarity of the cultivated land data. Finally, for the GF-2 multispectral image, a cloud detection method based on the radiation transmission model is used. This method uses radiation transmission to estimate the scattering coefficient and combines geometric and high-resolution texture features to achieve high-precision cloud detection, effectively reducing the impact of bright buildings on cloud detection. Even in the case of fewer bands, it can achieve an accuracy of more than 94.70%, thereby solving the problem of farmland information extraction caused by cloud occlusion and thin cloud cover, and significantly improving the quality of cultivated land data. Figure 2 The comparison diagram before and after optimization shown clearly shows the optimized image obtained after the cloud occlusion image is processed by the first optimization model.

[0026] The above-mentioned optimization of the target remote sensing image set through the first optimization model in the deep learning method refers to inputting the target remote sensing image set into the first optimization model in the deep learning method, and the data output by the model is the attribute parameters of the optimized target remote sensing image set.

[0027] In a specific embodiment, the present invention optimizes the data quality of the target remote sensing image set. The improvement in data quality directly enhances the accuracy and clarity of the image information, so that the details and key information in the image can be better presented. This optimization not only provides a more reliable foundation for the subsequent processing of remote sensing data, but also significantly improves the accuracy and reliability of data analysis.

[0028] Specifically, the comparison with the data quality threshold is as follows: if the data quality index of the target remote sensing image set is less than or equal to the data quality threshold, the data quality threshold and the data quality index of the target remote sensing image set are subjected to difference processing, and the difference processing result is marked as the data quality index deviation value of the target remote sensing image set; according to the data quality index deviation value of the target remote sensing image set, an adjustment set of the first optimization model is matched from the spatiotemporal management information library, the first optimization model is adjusted according to the adjustment set of the first optimization model, and the target remote sensing image set is optimized by the adjusted first optimization model until the data quality index of the target remote sensing image set is greater than the data quality threshold, and the optimized target remote sensing image set is marked as the target optimized remote sensing image set. The above-mentioned data quality threshold refers to the minimum value of the reasonable range of the data quality index of the target remote sensing image set, which is extracted from the spatiotemporal management information library. The spatiotemporal management information library stores the adjustment set of the first optimization model corresponding to each data quality index deviation value interval. In an example embodiment, assuming that the data quality index deviation value of a certain target remote sensing image set is G, belonging to The adjustment set of the first optimization model corresponding to the data quality index deviation value interval [G-13%, G+12%] stored in the spatiotemporal management information database and the data quality index deviation value interval [G-13%, G+12%] includes: gradually increasing the model learning rate, up to 0.005, gradually increasing the number of model iterations, up to 3 times the existing number of iterations, and gradually increasing the convolution kernel size, up to 2.5 times the existing convolution kernel. Then, the adjustment set of the first optimization model corresponding to the target remote sensing image set includes: gradually increasing the model learning rate, up to 0.005, gradually increasing the number of model iterations, up to 3 times the existing number of iterations, and gradually increasing the convolution kernel size, up to 2.5 times the existing convolution kernel. It should be noted that the adjustment set adopts a highly accurate and targeted adjustment strategy, which can quickly identify and correct problems in the first optimization model according to the shortcomings of the current data quality index. Therefore, it is only necessary to make fewer adjustments until the data quality index of the target remote sensing image set is greater than the data quality threshold, and there is no situation where the data quality index of the target remote sensing image set cannot be adjusted to be greater than the data quality threshold.

[0029] Specifically, the attribute parameters of the optimized target remote sensing image set specifically include the grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, the spectral band probability curve of each sample image belonging to the target remote sensing image set, and the discrete value of the pixel value of each sample area of ​​each sample image belonging to the target remote sensing image set. In an example embodiment, the target remote sensing image set includes optical images, radar images and data elevation images of cultivated land, and the sample images are randomly extracted from the optical images, radar images and data elevation images of cultivated land through sampling tools such as ArcGIS; the grayscale value of each spectral band of each sample image belonging to the above-mentioned target remote sensing image set is detected by the detection tool belonging to the complete remote sensing image processing platform. The detailed information includes the grayscale value and the wavelength range. Each spectral band contains different grayscale values. The different grayscale values ​​are averaged, and the processing results are marked as the grayscale values ​​of each spectral band of each sample image belonging to the target remote sensing image set; the spectral band probability curve of each sample image belonging to the above-mentioned target remote sensing image set refers to The grayscale values ​​of each spectral band of each sample image in the target remote sensing image set are counted to obtain the distribution of grayscale values ​​of each sample image. The number of occurrences of each grayscale value is divided by the total number of grayscale values, which is the occurrence probability of each grayscale value. The occurrence probability of each grayscale value is marked as the occurrence probability of each spectral band. The matrix laboratory software is used to draw it into a coherent smooth curve, that is, the spectral band probability curve of each sample image in the target remote sensing image set is obtained; the discrete values ​​of the pixel values ​​of each sample area of ​​each sample image in the target remote sensing image set are obtained by remote sensing image processing software The pixel value of each pixel point in each sample area of ​​each sample image belonging to the target remote sensing image set is detected by software (such as a geographic information system platform), and the pixel value of each pixel point is processed by standard deviation. The processing result is marked as a discrete value of pixel value, thereby obtaining the discrete value of pixel value of each sample area of ​​each sample image belonging to the target remote sensing image set. The sample area refers to an area composed of a certain number of pixels, and these pixels have similarities in spectral characteristics and spatial distribution. These areas usually correspond to specific land object types, such as cultivated land, woodland, built-up area and water body.

[0030] Furthermore, the data quality index of the target remote sensing image set is evaluated, and the specific evaluation process is: obtaining the reference grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, performing corresponding difference processing with the grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, obtaining the deviation grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, and performing accumulation processing to obtain the total deviation grayscale value of each sample image belonging to the target remote sensing image set; the reference grayscale value of each spectral band of each sample image belonging to the target remote sensing image set is obtained by The data analyst analyzes the grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, and marks it as the reference grayscale value of each spectral band of each sample image belonging to the remote sensing image set. It should be explained that compared with the number of images in the target remote sensing image set, the number of each sample image is relatively small, such as only dozens of images, while the number of images in the target remote sensing image set may be millions. The data analyst only needs to analyze the information of each sample image for comparison, which reflects all the image information in the target remote sensing image set and makes the data analysis fit the actual situation.

[0031] The probability corresponding to each spectral band is located and extracted from the spectral band probability curve, and data processing is performed, and the processing result is marked as information entropy, thereby obtaining the information entropy of each sample image belonging to the target remote sensing image set through this method; in an exemplary embodiment, a spectral band probability curve is as follows Figure 5 As shown, the horizontal axis is the spectral band, the unit is nanometers, and the vertical axis is the probability of occurrence, the unit is percentage, showing the change of spectral band probability, locating and extracting the corresponding probability of occurrence of each spectral band, and performing the following data processing: ,in is the probability of the cth spectral band, c is the number of each spectral band, , k is the total number of spectral bands, and the processing result is the information entropy. This method can be used to analyze and calculate the information entropy of each sample image in the target remote sensing image set.

[0032] The discrete values ​​of pixel values ​​of each sample area of ​​each sample image belonging to the target remote sensing image set are averaged, and the processing results are marked as the average pixel discrete values ​​of each sample image belonging to the target remote sensing image set; the reference information entropy of each sample image belonging to the target remote sensing image set is obtained; the reference information entropy of each sample image belonging to the target remote sensing image set is consistent with the reference grayscale value of each spectral band of each sample image belonging to the target remote sensing image set.

[0033] The total deviation gray value of each sample image belonging to the target remote sensing image set, the information entropy of each sample image belonging to the target remote sensing image set, the average pixel discrete value of each sample image belonging to the target remote sensing image set, and the reference information entropy of each sample image belonging to the target remote sensing image set are comprehensively analyzed to obtain an evaluation method for the data quality index of the target remote sensing image set. In this embodiment, the data quality index of the target remote sensing image set refers to the numerical value of the data quality of the target remote sensing image set. The specific evaluation method is:

[0034] ;

[0035] It is the total deviation gray value of the d-th sample image of the target remote sensing image set, which refers to the sum of the deviations between the pixel gray value and the reference gray value in the d-th sample image of the target remote sensing image set. This value reflects the accuracy of the gray distribution of the image.

[0036] It is the information entropy of the d-th sample image in the target remote sensing image set, and refers to the gray value distribution of the d-th sample image in the target remote sensing image set.

[0037] It is the average pixel discrete value of the dth sample image belonging to the target remote sensing image set, which refers to the average difference between the pixel values ​​of the sample area of ​​the dth sample image belonging to the target remote sensing image set.

[0038] is the reference information entropy of the d-th sample image in the target remote sensing image set, indicating the reference value of the information entropy of the d-th sample image in the target remote sensing image set. The remote sensing image of the cultivated field must contain different pixel contents, so the value of the reference information entropy is not 0.

[0039] It is the data quality index weight factor corresponding to the total deviation gray value unit value preset in the spatiotemporal management information library, which represents the influence degree of the total deviation gray value unit value on the data quality index of the target remote sensing image set. The corresponding relationship of the weight factors is stored in the spatiotemporal management information library. By inputting the total deviation gray value into the spatiotemporal management information library, the data quality index weight factor corresponding to the total deviation gray value unit value can be obtained.

[0040] It is the data quality index weight factor corresponding to the average pixel discrete value unit value preset in the spatiotemporal management information library, which represents the influence degree of the average pixel discrete value unit value on the data quality index of the target remote sensing image set. The corresponding relationship of the weight factors is stored in the spatiotemporal management information library. By inputting the average pixel discrete value into the spatiotemporal management information library, the data quality index weight factor corresponding to the average pixel discrete value unit value can be obtained.

[0041] e is a natural constant, d is the number of each sample image, , g is the total number of sample images.

[0042] in, is the data quality index of the target remote sensing image set. The total deviation grayscale value reflects the degree of deviation of the image grayscale value. When the value is large, it means that the images in the target remote sensing image set may contain a large amount of noise, outliers or detail information, which makes the grayscale values ​​of each spectral band of the images in the target remote sensing image set significantly deviate from the corresponding reference grayscale values. Therefore, the target remote sensing image set will appear blurred when displaying the cultivated land area to be detected, and key information such as the boundary and texture of the cultivated land may be difficult to accurately identify. When the image cannot accurately display the cultivated land area, the distribution law of its grayscale value will also change, resulting in a significant deviation between the information entropy and the reference information entropy (i.e., the ideal information entropy). The deviation of information entropy may mean that the image contains It contains more unnecessary noise information or lacks important cultivated land information, thereby reducing the practical value of the image. In the consistent area of ​​the image (that is, the area that should have similar grayscale values ​​in theory, such as the inside of cultivated land), the difference between pixel values ​​will increase due to the increase of the total deviation grayscale value. The increase of the average pixel discrete value reflects that the grayscale value of the local area in the image is more discrete, which may cause the information such as the texture details inside the cultivated land area to become disorganized, further affecting the clarity and readability of the image, and thus jointly lead to the poor data quality of the target remote sensing image set. If the data quality is poor, the extracted features may be inaccurate, incomplete or misleading, thus affecting the accuracy and efficiency of the entire remote sensing image processing and analysis.

[0043] Step 2: If the data quality index of the target remote sensing image set is greater than the data quality threshold, the optimized target remote sensing image set is marked as the target optimized remote sensing image set, and feature detection and extraction are performed on the target optimized remote sensing image set through the second feature extraction model in the deep learning method, and the target optimized remote sensing image set is marked as the pixel-level feature set of the target optimized remote sensing image set. The attribute parameters of the pixel-level feature set are obtained, the feature accuracy factor of the pixel-level feature set is analyzed, and compared with the feature accuracy threshold.

[0044] The second feature extraction model in the above-mentioned deep learning method specifically refers to the integration of multi-source remote sensing information, including multispectral data of Landsat and Sentinel, polarimetric radar data, and digital elevation information. In terms of technical implementation, the object-oriented classification method is referred to, and superpixel technology is introduced, which significantly improves the classification efficiency and effectively reduces the segmentation error that may be generated by the traditional object-oriented method in the segmentation process, thereby ensuring the high accuracy of cultivated land identification. In an example embodiment, the second feature extraction model extracts information from the target remote sensing image set of cultivated land, such as Figure 3 As shown in the schematic diagram of the information extraction process, the feature detection and extraction process of the cultivated land target remote sensing image set is demonstrated. Specifically, the cultivated land target remote sensing image set includes optical images, radar images and digital elevation models. The second feature extraction model performs data preprocessing on the optical images, radar images and digital elevation models, including geometric correction and radiation correction, etc., and uses multispectral data to perform optical data fusion, radar polarization feature extraction and data elevation feature extraction through fusion algorithms (such as principal component analysis, wavelet transform, etc.), so as to obtain the spectral information (such as vegetation coverage), change information (such as surface temperature change, etc.), derived parameters (such as soil moisture and surface roughness, etc.), signal parameters (such as radar signal strength, etc.), altitude, slope and slope aspect of the cultivated land, and store them in the pixel-level feature set.

[0045] The above-mentioned feature detection and extraction of the target optimized remote sensing image set through the second feature extraction model in the deep learning method refers to inputting the target optimized remote sensing image set into the second feature extraction model in the deep learning method, and the data output by the model is generally marked as a pixel-level feature set of the target optimized remote sensing image set.

[0046] In a specific embodiment, the present invention can more accurately identify and extract key features in the image, such as texture shape, by analyzing and optimizing the feature detection and extraction of the target optimized remote sensing image set, thereby providing a richer and more accurate information basis for subsequent image analysis and recognition. Secondly, in the feature extraction process, through the detailed analysis and optimization of the features, the stability and robustness of the features can be significantly improved, and the influence of noise and interference on feature extraction can be reduced.

[0047] Specifically, the comparison with the feature accuracy threshold is performed, and the specific analysis process is as follows:

[0048] If the feature accuracy factor of the pixel-level feature set is less than or equal to the feature accuracy threshold, the feature accuracy threshold and the feature accuracy factor of the pixel-level feature set are subjected to difference processing, and the difference processing result is marked as the feature accuracy factor deviation value of the pixel-level feature set; according to the feature accuracy factor deviation value of the pixel-level feature set, an adjustment set of the second feature extraction model is matched from the spatiotemporal management information library, the second feature extraction model is adjusted according to the adjustment set of the second feature extraction model, and feature detection and extraction are performed on the target optimized remote sensing image set through the adjusted second feature extraction model, and the feature is adjusted to the pixel-level feature set. When the precision factor is greater than the feature precision threshold, the abnormal change factor of the target optimized remote sensing image set is analyzed according to the pixel-level feature set of the target optimized remote sensing image set; the above-mentioned feature precision threshold refers to the minimum value of the reasonable range of the feature precision factor of the pixel-level feature set, which is extracted from the spatiotemporal management information library; the spatiotemporal management information library stores the adjustment set of the second feature extraction model corresponding to each feature precision factor deviation value interval. In an example embodiment, assuming that the feature precision factor deviation value of a pixel-level feature set is X, it belongs to the feature precision factor deviation value interval [X-13%, X+ 7%] The adjustment set of the second feature extraction model corresponding to the feature precision factor deviation value interval [X-13%, X+7%]] includes: gradually increasing the number of network layers of the second feature extraction model, up to 6 times the existing number of layers, gradually increasing the number of neurons in the fully connected layer, up to 5 times the existing number of neurons, adding the sum of squares of weights as a penalty term through L2 regularization, which can reduce the value of the weight and prevent the model from overfitting. Then the adjustment set of the second feature extraction model corresponding to the feature precision factor deviation value of the pixel-level feature set includes: gradually increasing the number of network layers of the second feature extraction model, up to 6 times the existing number of layers , gradually increase the number of neurons in the fully connected layer, up to 5 times the existing number of neurons, and add the sum of squares of weights as a penalty term through L2 regularization to reduce the value of weights and prevent model overfitting; It should be noted that the adjustment set adopts a highly accurate and targeted adjustment strategy, which can quickly identify and correct problems in the second feature extraction model according to the shortcomings of the current feature precision factor. Therefore, when the feature precision factor of the pixel-level feature set is adjusted to be greater than the feature precision threshold, only a few adjustments are required, and there is no situation where the feature precision factor of the pixel-level feature set cannot be adjusted to be greater than the feature precision threshold.

[0049] Specifically, the feature precision factor of the pixel-level feature set, in this embodiment, refers to a value obtained by comprehensively analyzing the data quality index of the target optimization remote sensing image set, the superpixel compactness of each sample pixel-level feature, the average contour boundary curvature of each sample pixel-level feature, the spatial resolution of each sample pixel-level feature, the reference superpixel compactness of each sample pixel-level feature, the average contour boundary curvature of each sample pixel-level feature, and the reference spatial resolution of each sample pixel-level feature, which is used to determine the feature recognition and extraction accuracy of the pixel-level feature set. The specific analysis process is:

[0050] ;

[0051] The data quality index of the target optimized remote sensing image set is the prerequisite for ensuring that the feature accuracy of the pixel-level feature set is at a high level in the feature accuracy factor of the pixel-level feature set. Therefore, the data quality index of the target optimized remote sensing image set is integrated into the feature accuracy index of the pixel-level feature set, wherein the data quality index of the target remote sensing image set has the same meaning as the data quality index of the target remote sensing image set.

[0052] It is the weight factor corresponding to the data quality index preset in the spatiotemporal management information library, which represents the proportion of the data quality index of the target optimized remote sensing image set to the feature accuracy factor of the pixel-level feature set. The corresponding relationship of the weight factors is stored in the spatiotemporal management information library. By inputting the data quality index of the target optimized remote sensing image set into the spatiotemporal management information library, the weight factor corresponding to the data quality index can be obtained.

[0053] is the superpixel compactness of the h-th sample pixel-level feature, which refers to the superpixel compactness of the h-th sample pixel-level feature.

[0054] It is the average contour boundary curvature of the h-th sample pixel-level feature, which refers to the curvature of the contour boundary of the h-th sample pixel-level feature. In remote sensing images, the contour boundary curvature of a superpixel can reflect its shape complexity and edge characteristics.

[0055] It is the spatial resolution of the h-th sample pixel-level feature, which refers to the size of the smallest ground object unit that can be distinguished by the h-th sample pixel-level feature. For example, if the spatial resolution of a remote sensing image is 1 meter, then each pixel in the image represents an area of ​​1 meter × 1 meter on the actual ground.

[0056] is the reference superpixel compactness of the h-th sample pixel-level feature, which represents the reference value of the superpixel compactness of the h-th sample pixel-level feature. The pixel-level feature must contain cultivated land feature information, so the value of the reference superpixel compactness is a non-zero value.

[0057] The reference average contour boundary curvature of the hth sample pixel-level feature, indicating the reference value of the average contour boundary curvature of the hth sample pixel-level feature. The pixel-level feature must contain cultivated land feature information, so the value of the reference average contour boundary curvature is a non-zero value.

[0058] It is the reference spatial resolution of the h-th sample pixel-level feature, indicating the reference value of the spatial resolution of the h-th sample pixel-level feature. The pixel-level feature must contain cultivated land feature information, so the value of the reference spatial resolution is a non-zero value.

[0059] e is a natural constant, h is the number of each sample pixel-level feature, , y is the total number of sample pixel-level features.

[0060] in, is the feature accuracy factor of the pixel-level feature set. The data quality of the higher optimized target remote sensing image set is the cornerstone to ensure the feature accuracy of the pixel-level feature set. In remote sensing image processing, superpixel compactness is a crucial pixel-level feature, which is directly related to the accuracy of image segmentation and the clarity of object recognition. When the superpixel compactness deviates significantly from its reference value, especially when the actual compactness is much higher than the reference value, it will cause serious problems in image segmentation. In this case, some objects that should have been clearly separated, especially cultivated fields with similar shapes or adjacent to each other, may be mistakenly merged into a whole. This merger not only destroys the integrity of the cultivated field boundary, but also introduces new segmentation errors, making subsequent object recognition and classification difficult. Furthermore, due to the abnormal increase in superpixel compactness, the spatial resolution of the sample pixel-level features will also be significantly affected, and the original ability to clearly distinguish different cultivated fields will be affected. The image of the farmland area may become blurred when the compactness is too high, resulting in a significant deviation of the spatial resolution from its reference value. This deviation greatly reduces the expressiveness of the image at the detail level, making it impossible to accurately distinguish adjacent farmland areas, and may even lead to misjudgment of the farmland boundary. Boundary curvature, as an important feature for measuring the smoothness of the object boundary, is also affected when the superpixel compactness is abnormal. When the compactness is too high, the farmland boundary may become too smooth, resulting in a boundary curvature that is significantly lower than the reference contour boundary curvature. This reduction not only masks the original boundary details, but also may cause the slight differences between adjacent farmlands to be ignored, further reducing the feature accuracy of the pixel-level feature set. In summary, in the process of remote sensing image processing and analysis, it is necessary to analyze key features such as superpixel compactness to ensure the stable improvement of feature accuracy and provide a strong guarantee for the accurate identification and classification of farmland and other land objects.

[0061] Specifically, the attribute parameters of the pixel-level feature set include the superpixel compactness of each sample pixel-level feature, the contour boundary curvature of each detection position point to which each sample pixel-level feature belongs, and the spatial resolution of each sample pixel-level feature; the superpixel compactness of each sample pixel-level feature is obtained as follows: a superpixel segmentation algorithm (such as simple linear iterative clustering, etc.) is applied to segment the sample pixel-level feature into multiple superpixels, and for each superpixel, its perimeter and area are calculated, thereby calculating the ratio of the total area of ​​all superpixels in the sample pixel-level feature to the square of its total perimeter (multiplied by 4π for normalization) to obtain the compactness; The contour boundary curvature of each detection position point belonging to each sample pixel-level feature refers to first performing edge detection on each sample pixel-level feature (using the Canny edge detection algorithm), and then using a mathematical method (such as discrete point difference method and linear regression method, etc.) to calculate the curvature of each detection position point on the contour, thereby obtaining the contour boundary curvature of each detection position point belonging to each sample pixel-level feature, wherein each detection position point is a point randomly arranged on the contour boundary of the sample pixel-level feature; the spatial resolution of the above-mentioned each sample pixel-level feature can be obtained by querying the attribute information of each sample pixel-level feature through remote sensing image processing software (such as remote sensing image processing platform).

[0062] Furthermore, the feature precision factor of the pixel-level feature set is analyzed, and the specific analysis process is: the contour boundary curvature of each detection position point belonging to each sample pixel-level feature is averaged, and the processing result is marked as the average contour boundary curvature of each sample pixel-level feature; the above-mentioned sample pixel-level feature refers to an image randomly extracted from the pixel-level feature set through a sampling tool such as ArcGIS.

[0063] The data quality index of the target optimized remote sensing image set, the reference superpixel compactness of each sample pixel-level feature, the average contour boundary curvature of each sample pixel-level feature and the reference spatial resolution of each sample pixel-level feature are obtained; the data quality index of the target optimized remote sensing image set is the data quality index of the target remote sensing image set corresponding to the data quality index of the target remote sensing image set when the data quality index of the target remote sensing image set is greater than the data quality threshold; the reference superpixel compactness of each sample pixel-level feature, the average contour boundary curvature of each sample pixel-level feature and the reference spatial resolution of each sample pixel-level feature are obtained in the same manner as the reference grayscale value of each spectral band of each sample image belonging to the target remote sensing image set.

[0064] The data quality index of the target optimized remote sensing image set, the superpixel compactness of each sample pixel-level feature, the average contour boundary curvature of each sample pixel-level feature, the spatial resolution of each sample pixel-level feature, the reference superpixel compactness of each sample pixel-level feature, the average contour boundary curvature of each sample pixel-level feature and the reference spatial resolution of each sample pixel-level feature are comprehensively analyzed to obtain the feature precision factor of the pixel-level feature set.

[0065] Step 3: If the feature accuracy factor of the pixel-level feature set is greater than the feature accuracy threshold, the abnormal change factor of the target optimized remote sensing image set is analyzed based on the pixel-level feature set of the target optimized remote sensing image set, thereby issuing an early warning for abnormal changes in the target optimized remote sensing image set.

[0066] Specifically, the above-mentioned early warning of abnormal changes in the target optimized remote sensing image set is analyzed in the following steps: comparing the abnormal change factor of the target optimized remote sensing image set with the abnormal change threshold; if the abnormal change factor of the target optimized remote sensing image set is greater than the abnormal change threshold, an early warning of abnormal changes in the target optimized remote sensing image set is issued; the above-mentioned abnormal change threshold refers to the maximum value of a reasonable range of the abnormal change factor of the target optimized remote sensing image set, which is extracted from a spatiotemporal management information library; the above-mentioned early warning of abnormal changes in the target optimized remote sensing image set, in an example embodiment, the detailed information of the target optimized remote sensing image set and the abnormal change factor of the target optimized remote sensing image set can be packaged into an early warning log, and the early warning log can be visualized and displayed to a data analyst, thereby completing the early warning of abnormal changes in the target optimized remote sensing image set; if the abnormal change factor of the target optimized remote sensing image set is less than or equal to the abnormal change threshold, the target optimized remote sensing image set analyzed this time is marked as a historical target optimized remote sensing image set, and the uploaded target remote sensing image set continues to be received for analysis.

[0067] Furthermore, the abnormal change factor of the target optimized remote sensing image set is analyzed, and the specific analysis process is: extracting the local eigenvalue of each pixel-level feature, the information entropy of each pixel-level feature and the contour boundary curve of each pixel-level feature from the pixel-level feature set of the target optimized remote sensing image set; the local eigenvalue of each pixel-level feature is obtained by using the number of edge pixels of each pixel-level feature extracted by an edge detection algorithm (such as Canny, etc.) as the local eigenvalue of the shape feature of each pixel-level feature; the information entropy of each pixel-level feature is consistent with the information entropy acquisition method of each sample image belonging to the target remote sensing image set; the contour boundary curve of each pixel-level feature is identified by the Canny edge detection algorithm.

[0068] The historical contour boundary curves of each pixel-level feature are extracted from the spatiotemporal management information library, and the curves are overlapped and compared with the contour boundary curves corresponding to each pixel-level feature, and the total deviation value of the contour boundary curve of each pixel-level feature is obtained by analysis; the above-mentioned historical contour boundary curves of each pixel-level feature refer to the latest historical contour boundary curves of each pixel-level feature stored in the spatiotemporal management information library. The above analysis obtains the total deviation value of the contour boundary curve of each pixel-level feature. The specific analysis process is: the contour boundary curve of each pixel-level feature is overlapped with the corresponding historical contour boundary curve (that is, the corresponding coordinate points are overlapped), the distance between each point on the contour boundary curve and the corresponding point of the corresponding historical contour boundary curve is calculated and accumulated, so as to obtain the total deviation value of the contour boundary curve of each pixel-level feature.

[0069] The historical local feature values ​​of each pixel-level feature and the historical information entropy of each pixel-level feature are extracted from the spatiotemporal management information library, and the analysis method of the abnormal change factor of the target optimized remote sensing image set is obtained by comprehensively analyzing the feature precision factor of the pixel-level feature set, the local feature values ​​of each pixel-level feature, the information entropy of each pixel-level feature, the total deviation value of the contour boundary curve of each pixel-level feature, the historical local feature values ​​of each pixel-level feature and the historical information entropy of each pixel-level feature. In this embodiment, the abnormal change factor of the target optimized remote sensing image set refers to the numerical value of the abnormal change degree of the target optimized remote sensing image set. The specific analysis method is as follows:

[0070] ;

[0071] is the feature precision factor of the pixel-level feature set, is the feature accuracy threshold. In the abnormal change factor of the target optimized remote sensing image set, if the feature accuracy factor of the pixel-level feature set is significantly greater than the feature accuracy threshold, it means that the feature accuracy of the pixel-level feature set is at a high level. When comparing pixel-level features with historical conditions, this high-precision feature set helps reduce the risk of abnormal errors caused by fuzzy or unclear identification. Therefore, the feature accuracy factor and feature accuracy threshold of the pixel-level feature set are integrated into the abnormal change factor of the target optimized remote sensing image set.

[0072] is the local feature value of the t-th pixel-level feature, which refers to the local quantized value extracted from a specific feature (such as pixel, color, texture, gradient, etc.) of the t-th pixel-level feature. In this embodiment, it refers to the number of edge pixels of the t-th pixel-level feature.

[0073] It is the historical local eigenvalue of the t-th pixel-level feature, indicating the latest historical local eigenvalue of the t-th pixel-level feature. The historical local eigenvalue is used to display the historical information of the pixel-level feature, and its value is non-zero.

[0074] is the information entropy of the t-th pixel-level feature, which refers to the grayscale value distribution of the t-th pixel-level feature.

[0075] is the historical information entropy of the t-th pixel-level feature, indicating the latest historical information entropy of the t-th pixel-level feature. The historical information entropy is used to display the historical information of the pixel-level feature, and its value is non-zero.

[0076] It is the total deviation value of the contour boundary curve of the t-th pixel-level feature, which refers to the total deviation between the contour boundary curve of the feature corresponding to the t-th pixel-level feature and the historical contour boundary curve.

[0077] G is the abnormal change weight factor corresponding to the unit value of the total deviation value of the contour boundary curve preset in the space-time management information library, which represents the influence degree of the unit value of the total deviation value of the contour boundary curve on the abnormal change factor of the target optimized remote sensing image set. The corresponding relationship of the weight factors is stored in the space-time management information library. By inputting the total deviation value of the contour boundary curve into the space-time management information library, the data quality index weight factor corresponding to the unit value of the average pixel discrete value can be obtained.

[0078] t is the number of each pixel-level feature, , r is the total number of pixel-level features.

[0079] in, Optimize the abnormal change factor of the remote sensing image set for the target. If the feature accuracy factor is higher than the set feature accuracy threshold, it usually means that the overall feature accuracy of the pixel-level feature set is in good condition. In this case, if there is a significant difference between the current local feature value of the pixel-level feature and its historical local feature value, this may indicate that these pixel-level features, especially those related to cultivated land, have undergone significant changes. This change in local feature values ​​is particularly important for cultivated land because it may reflect changes in crop growth status, changes in land cover types (such as conversion of cultivated land to construction land), or changes in soil conditions. When these changes occur, the total deviation value of the contour boundary curve of the pixel-level feature tends to increase. For cultivated land, it means The blurring or displacement of cultivated land boundaries may be caused by changes in farming activities, natural disasters (such as floods and droughts) or human factors (such as illegal occupation of cultivated land). At the same time, changes in local eigenvalues ​​may also cause the information entropy of pixel-level features to deviate significantly from its historical information entropy. For cultivated land, changes in information entropy may reflect changes in the complexity of internal information in the cultivated land, such as an increase or decrease in the diversity of crop species and changes in soil moisture, which is of great significance for monitoring the health of cultivated land and predicting crop yields. Based on the above analysis, if multiple changes are accumulated on multiple pixel-level features, that is, multiple features related to cultivated land show significant changes, then the changes in the target optimized remote sensing image set can be considered abnormal changes.

[0080] It should be explained that the spatiotemporal information database is a database that stores the data required for remote sensing image change detection. It integrates the spatiotemporal information basic data, new spatiotemporal information mapping data, spatiotemporal information thematic data, and multi-source heterogeneous sensor data. It realizes the spatialization and centralized storage of massive multi-source heterogeneous data, unified query and visualization display. The management architecture of the spatiotemporal information database is as follows: Figure 4As shown in the figure, it is divided into application layer, engine layer and data layer. The application layer passes the user's operation request to the engine layer by calling the API or interface provided by the engine layer. After receiving the request from the application layer, the engine layer parses and executes the corresponding database operation, and then returns the result to the application layer. When executing the database operation, the engine layer needs to access the spatiotemporal information data in the data layer. The application layer includes data aggregation, data processing, data quality inspection, data directory, data storage, data display, service publishing, data update, historical data management, data extraction, data query, data statistics, data mapping and security management. The engine layer includes Kettle (data integration engine), GeoEvent (real-time geographic event processing engine), The data layer includes public thematic data, multi-source heterogeneous sensor data, vector data, place name and address data, geographic entity data, natural resource thematic data, image data, elevation model data, three-dimensional terrain data, new surveying and mapping product data and electronic map data. The spatiotemporal information database realizes the management of massive multi-source heterogeneous data, and adopts the way of warehousing and hanging to bring together rich spatiotemporal big data. The data currently managed include basic geographic data, new product surveying and mapping data, natural resource thematic data, intelligent perception data, etc., with a total data volume of 9TB. It is based on ArcGIS Engine and integrates the components of the three-dimensional data SkyLine platform (skyline three-dimensional platform) to integrate traditional surveying and mapping data with new surveying and mapping data. All data can be processed, browsed, queried and managed through database management.

[0081] In a specific embodiment, the present invention provides a remote sensing image change detection method based on deep learning, which optimizes the target remote sensing image set through the first optimization model in the deep learning method, obtains the attribute parameters of the optimized target remote sensing image set, evaluates the data quality of the target remote sensing image set, and extracts the features based on the target remote sensing image set with high data quality, analyzes the feature accuracy of the pixel-level feature set, and warns of abnormal changes in the target optimized remote sensing image set based on the high-precision features, thereby achieving accurate early warning of abnormal changes in the target optimized remote sensing image set, providing strong technical support for remote sensing monitoring and analysis, and effectively improving the accuracy and efficiency of change detection.

[0082] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A remote sensing image change detection method based on deep learning, characterized in that: include: Step 1: Optimize the target remote sensing image set through the first optimization model in the deep learning method, obtain the attribute parameters of the optimized target remote sensing image set, evaluate the data quality index of the target remote sensing image set, and compare it with the data quality threshold; Step 2: If the data quality index of the target remote sensing image set is greater than the data quality threshold, the optimized target remote sensing image set is marked as the target optimized remote sensing image set, and the target optimized remote sensing image set is subjected to feature detection and extraction by the second feature extraction model in the deep learning method, and is marked as the pixel-level feature set of the target optimized remote sensing image set, and the attribute parameters of the pixel-level feature set are obtained, and the feature precision factor of the pixel-level feature set is analyzed, and compared with the feature precision threshold; Step 3: If the feature precision factor of the pixel-level feature set is greater than the feature precision threshold, the abnormal change factor of the target optimized remote sensing image set is analyzed according to the pixel-level feature set of the target optimized remote sensing image set, thereby giving an early warning of the abnormal change of the target optimized remote sensing image set; The attribute parameters of the optimized target remote sensing image set specifically include the grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, the spectral band probability curve of each sample image belonging to the target remote sensing image set, and the discrete value of the pixel value of each sample area of ​​each sample image belonging to the target remote sensing image set; The attribute parameters of the pixel-level feature set specifically include the superpixel compactness of each sample pixel-level feature, the contour boundary curvature of each detection position point to which each sample pixel-level feature belongs, and the spatial resolution of each sample pixel-level feature; The feature precision factor of the pixel-level feature set is analyzed, and the specific analysis process is as follows: Performing mean processing on the contour boundary curvatures of each detection position point to which each sample pixel-level feature belongs, and marking the processing result as the average contour boundary curvature of each sample pixel-level feature; Obtain the data quality index of the target optimized remote sensing image set, the reference superpixel compactness of each sample pixel-level feature, the average contour boundary curvature of each sample pixel-level feature, and the reference spatial resolution of each sample pixel-level feature; Comprehensively analyze the data quality index of the target optimized remote sensing image set, the superpixel compactness of each sample pixel-level feature, the average contour boundary curvature of each sample pixel-level feature, the spatial resolution of each sample pixel-level feature, the reference superpixel compactness of each sample pixel-level feature, the average contour boundary curvature of each sample pixel-level feature, and the reference spatial resolution of each sample pixel-level feature to obtain the feature precision factor of the pixel-level feature set; The characteristic precision factor of the pixel-level feature set is specifically analyzed as follows: ； in, is the feature precision factor of the pixel-level feature set, To optimize the data quality index of remote sensing image sets, is the weight factor corresponding to the data quality index preset in the spatiotemporal management information database, is the superpixel compactness of the h-th sample pixel-level feature, is the average contour boundary curvature of the h-th sample pixel-level feature, is the spatial resolution of the h-th sample pixel-level feature, is the reference superpixel compactness of the h-th sample pixel-level feature, is the reference average contour boundary curvature of the hth sample pixel-level feature, is the reference spatial resolution of the hth sample pixel-level feature, e is a natural constant, h is the number of each sample pixel-level feature, , y is the total number of sample pixel-level features.

2. The remote sensing image change detection method based on deep learning according to claim 1, characterized in that: The data quality index of the target remote sensing image set is evaluated, and the specific evaluation process is as follows: Obtaining the reference grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, performing corresponding difference processing with the grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, obtaining the deviation grayscale value of each spectral band of each sample image belonging to the target remote sensing image set, and performing accumulation processing to obtain the total deviation grayscale value of each sample image belonging to the target remote sensing image set; Locate and extract the probability corresponding to each spectral band from the spectral band probability curve, perform data processing, and mark the processing result as information entropy, thereby obtaining the information entropy of each sample image belonging to the target remote sensing image set through this method; The pixel value discrete values ​​of each sample area of ​​each sample image belonging to the target remote sensing image set are averaged, and the processing results are marked as the average pixel discrete values ​​of each sample image belonging to the target remote sensing image set; Obtain the reference information entropy of each sample image in the target remote sensing image set; The total deviation grayscale value of each sample image in the target remote sensing image set, the information entropy of each sample image in the target remote sensing image set, the average pixel discrete value of each sample image in the target remote sensing image set and the reference information entropy of each sample image in the target remote sensing image set are comprehensively analyzed to obtain the data quality index of the target remote sensing image set.

3. The remote sensing image change detection method based on deep learning according to claim 1, characterized in that: The comparison is performed with the data quality threshold, and the specific comparison process is as follows: If the data quality index of the target remote sensing image set is less than or equal to the data quality threshold, the data quality threshold and the data quality index of the target remote sensing image set are subjected to difference processing, and the difference processing result is marked as the data quality index deviation value of the target remote sensing image set; According to the data quality index deviation value of the target remote sensing image set, an adjustment set of the first optimization model is matched from the spatiotemporal management information library, the first optimization model is adjusted according to the adjustment set of the first optimization model, and the target remote sensing image set is optimized by the adjusted first optimization model until the data quality index of the target remote sensing image set is greater than the data quality threshold, and the optimized target remote sensing image set is marked as the target optimized remote sensing image set.

4. The remote sensing image change detection method based on deep learning according to claim 1, characterized in that: The comparison is performed with the feature accuracy threshold, and the specific analysis process is as follows: If the feature precision factor of the pixel-level feature set is less than or equal to the feature precision threshold, the feature precision threshold is subjected to difference processing with the feature precision factor of the pixel-level feature set, and the difference processing result is marked as the feature precision factor deviation value of the pixel-level feature set; According to the feature precision factor deviation value of the pixel-level feature set, an adjustment set of the second feature extraction model is matched from the spatiotemporal management information library, the second feature extraction model is adjusted according to the adjustment set of the second feature extraction model, and feature detection and extraction are performed on the target optimized remote sensing image set through the adjusted second feature extraction model. When the feature precision factor of the pixel-level feature set is adjusted to be greater than the feature precision threshold, the abnormal change factor of the target optimized remote sensing image set is analyzed based on the pixel-level feature set of the target optimized remote sensing image set.

5. The remote sensing image change detection method based on deep learning according to claim 4 is characterized in that: The specific analysis process of analyzing the abnormal change factors of the target optimized remote sensing image set is as follows: Extracting the local eigenvalue of each pixel-level feature, the information entropy of each pixel-level feature and the contour boundary curve of each pixel-level feature from the pixel-level feature set of the target optimized remote sensing image set; Extract the historical contour boundary curve of each pixel-level feature from the spatiotemporal management information database, and compare the curve coincidence with the contour boundary curve corresponding to each pixel-level feature, and analyze and obtain the total deviation value of the contour boundary curve of each pixel-level feature; The historical local eigenvalues ​​of each pixel-level feature and the historical information entropy of each pixel-level feature are extracted from the spatiotemporal management information database. The abnormal change factor of the target optimized remote sensing image set is obtained by comprehensively analyzing the feature precision factor of the pixel-level feature set, the local eigenvalues ​​of each pixel-level feature, the information entropy of each pixel-level feature, the total deviation value of the contour boundary curve of each pixel-level feature, the historical local eigenvalues ​​of each pixel-level feature and the historical information entropy of each pixel-level feature.

6. The remote sensing image change detection method based on deep learning according to claim 1, characterized in that: The specific analysis process of early warning of abnormal changes in the target optimized remote sensing image set is as follows: The abnormal change factor of the target optimized remote sensing image set is compared with the abnormal change threshold. If the abnormal change factor of the target optimized remote sensing image set is greater than the abnormal change threshold, an early warning is issued for the abnormal change of the target optimized remote sensing image set.

Images