Cloud layer blocking-based landslide disaster satellite image data generation method and device, equipment and medium

Abstract

The application discloses a cloud layer blocking-based landslide disaster satellite image data generation method and device, equipment and medium, relates to the technical field of satellite landslide disasters, and comprises the following steps: obtaining multi-source remote sensing, meteorological, topographic and cloud-shielded multi-temporal optical satellite image data of a target area, calculating multi-dimensional landslide physical characteristic vectors in combination with the topography; fusing multi-source data fine classification cloud layers, determining a priority processing area according to the cloud layer, meteorological and topographic data; inputting the area image and the physical characteristic vector into a landslide feature enhancement generation network containing a double-attention module, performing feature extraction, enhancement, fusion and up-sampling recovery, generating a cloud-free landslide satellite image and completing credibility evaluation, effectively suppressing the interference of cloud layer blocking, and improving the accuracy of the satellite landslide image data generated under the condition of cloud layer blocking.

Description

Technical Field

[0001] This invention relates to the field of satellite landslide disaster technology, and in particular to a method, apparatus, equipment and medium for generating satellite image data of landslide disasters based on cloud cover. Background Technology

[0002] Currently, traditional methods for addressing cloud cover issues mainly include the following: First, using single-frame image information for cloud detection and removal, and then filling in the cloud-covered areas using image inpainting algorithms; second, employing multi-temporal image fusion technology to complement information from images acquired at different times; and third, introducing deep learning networks for end-to-end cloud removal and image reconstruction. In addition, some methods combine active remote sensing data sources such as radar data, utilizing their ability to penetrate clouds to obtain surface information, or introducing meteorological data to assist in dynamic cloud prediction.

[0003] However, existing methods have the following problems: methods that only utilize single-frame image information cannot effectively handle large-area cloud cover, and the restored images lack realistic surface texture features; existing methods are mostly designed for general scenarios and are insufficient in extracting the special texture features of landslides, making it difficult to accurately identify the boundaries and morphology of landslide bodies; traditional methods ignore the spatiotemporal correlation between multi-temporal images and fail to fully utilize the temporal patterns of landslide evolution; existing technologies lack the ability to finely classify clouds by fusing multi-temporal optical satellite imagery data and multi-source remote sensing data, and cannot adaptively adjust processing strategies according to the dynamic changes of clouds; at the same time, existing methods do not fully consider the guiding role of meteorological and topographic data in cloud processing strategies, nor do they introduce a credibility judgment mechanism to evaluate the reliability of the processing results.

[0004] Therefore, how to effectively address the difficulty in identifying landslide hazards in satellite imagery under cloud cover conditions has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, device, and medium for generating satellite image data of landslide disasters based on cloud cover, aiming to solve the technical problem of how to improve the accuracy of satellite landslide image data generated under cloud cover conditions.

[0006] To achieve the above objectives, this application proposes a method for generating satellite image data of landslide disasters based on cloud cover, comprising:

[0007] Acquire multi-temporal optical satellite imagery, multi-source remote sensing data, meteorological data, and topographic data of the target area, and calculate the multi-dimensional landslide physical feature vector corresponding to each optical satellite imagery in combination with the topographic data. The optical satellite imagery is landslide imagery data obscured by cloud cover, and the multi-source remote sensing data includes radar data, hyperspectral data, and UAV low-altitude remote sensing data.

[0008] By fusing the multi-temporal optical satellite imagery data and the multi-source remote sensing data, the clouds are finely classified to obtain cloud classification results.

[0009] Based on the cloud classification results, the meteorological data, and the terrain data, priority processing areas are determined;

[0010] The multi-temporal optical satellite image data and multi-dimensional landslide physical feature vector of the priority processing area are input into the landslide feature enhancement generation network to obtain landslide disaster satellite image data. The landslide feature enhancement generation network includes an encoder, a landslide sensitive channel attention module, a temporal collaborative attention module, a decoder and a skip connection module.

[0011] The step of inputting the multi-temporal optical satellite image data and multi-dimensional landslide physical feature vector of the priority processing area into the landslide feature enhancement generation network to obtain landslide disaster satellite image data includes:

[0012] The encoder is used to extract multi-scale spatial features from the multi-temporal optical satellite image data to obtain initial features;

[0013] The landslide-sensitive channel attention module uses the multi-dimensional landslide physical feature vector to perform channel-level directional enhancement on the initial features, resulting in enhanced features for each time phase.

[0014] The temporal collaborative attention module performs multi-temporal spatiotemporal correlation feature fusion on the enhanced features of each temporal phase to obtain fused features;

[0015] The decoder and the skip connection module are used to upsample and restore details of the fused features to generate a cloudless landslide disaster satellite image.

[0016] The credibility assessment of the aforementioned cloudless landslide disaster satellite images was performed, and the credibility assessment results were obtained.

[0017] When the credibility assessment result exceeds the preset credibility threshold, the cloudless landslide disaster satellite image will be used as the final landslide disaster satellite image data.

[0018] In one embodiment, the step of fusing the multi-temporal optical satellite imagery data and the multi-source remote sensing data to perform fine classification of clouds and obtain cloud classification results includes:

[0019] Radiometric correction and geometric registration are performed on the multi-temporal optical satellite image data to obtain the first preprocessed image;

[0020] The radar data is subjected to penetration enhancement processing to obtain a second pre-processed image;

[0021] The hyperspectral data is demixed based on spectral features to obtain a third preprocessed image;

[0022] The low-altitude remote sensing data from the UAV is subjected to resolution alignment and texture enhancement to obtain a fourth preprocessed image.

[0023] The first preprocessed image, the second preprocessed image, the third preprocessed image, and the fourth preprocessed image are fused at the multimodal feature level to obtain a fused feature map;

[0024] The fused feature map is input into the multi-scale feature extraction branch of the cloud fine classification network to obtain cloud candidate features with different spatial resolutions.

[0025] The candidate cloud features with different spatial resolutions are input into the spectral-texture dual-modal attention module of the cloud fine classification network, and attention weights are generated using the spectral dimension of the hyperspectral data and the texture dimension of the UAV data.

[0026] The attention weights are weighted and fused with the cloud candidate features to obtain the enhanced cloud discrimination features;

[0027] The enhanced cloud discrimination features are input into the classification head of the cloud fine classification network to obtain the probability distribution of each pixel belonging to thin cloud, thick cloud, cumulus cloud, and cirrus cloud.

[0028] Based on the probability distribution, cloud classification results including thin clouds, thick clouds, cumulus clouds, and cirrus clouds are generated.

[0029] In one embodiment, the step of determining the priority processing area based on the cloud classification result, the meteorological data, and the terrain data includes:

[0030] Based on the terrain data, terrain features are extracted, and the landslide risk level of the target area is calculated by combining the landslide susceptibility assessment model. The terrain features include slope, aspect and topographic relief.

[0031] Meteorological features are extracted from the meteorological data, including cloud height, cloud thickness, cloud movement speed, and cloud coverage duration.

[0032] Based on the meteorological characteristics and the cloud classification results, cloud influence weights are assigned to different types of clouds.

[0033] The cloud layer influence weight and the corresponding landslide risk level of the area are weighted and calculated to obtain the coupled risk value of the target area.

[0034] The coupling risk value is compared with a preset coupling risk threshold to filter out candidate regions whose coupling risk value is higher than the preset coupling risk threshold;

[0035] Connectivity analysis and boundary optimization are performed on the candidate regions to obtain satellite image data of landslide disasters in the target region and generate priority processing regions.

[0036] In one embodiment, the step of inputting multi-temporal optical satellite image data of the priority processing area and multi-dimensional landslide physical feature vectors into a landslide feature enhancement generation network to obtain landslide disaster satellite image data includes:

[0037] The encoder is used to extract multi-scale spatial features from the multi-temporal optical satellite image data to obtain initial features;

[0038] The landslide-sensitive channel attention module uses the multi-dimensional landslide physical feature vector to perform channel-level directional enhancement on the initial features, resulting in enhanced features for each time phase.

[0039] The temporal collaborative attention module performs multi-temporal spatiotemporal correlation feature fusion on the enhanced features of each temporal phase to obtain fused features;

[0040] The decoder and the skip connection module are used to upsample and restore details of the fused features to generate a cloudless landslide disaster satellite image.

[0041] The credibility assessment of the aforementioned cloudless landslide disaster satellite images was performed, and the credibility assessment results were obtained.

[0042] When the credibility assessment result exceeds the preset credibility threshold, the cloudless landslide disaster satellite image will be used as the final landslide disaster satellite image data.

[0043] In one embodiment, the step of using the landslide-sensitive channel attention module to perform channel-level directional enhancement of the initial features using the multi-dimensional landslide physical feature vector to obtain enhanced features for each time phase includes:

[0044] The multi-dimensional landslide physical feature vector is input into a two-layer fully connected mapping network, and dimensionality increase and feature decoupling are performed sequentially to obtain an adapted feature vector, wherein the adapted feature vector is consistent with the initial feature channel number.

[0045] The initial features are subjected to dual pooling operations of global average pooling and global max pooling, and then fused to obtain a dual-channel statistical vector.

[0046] The adaptation feature vector and the dual-channel statistical vector are element-wise summed and fused to obtain the fused channel guidance vector.

[0047] The initial channel weights are obtained by performing a nonlinear mapping on the fusion channel guide vector using an activation function.

[0048] The initial channel weights are corrected by introducing a cloud occlusion mask, the weights of channels related to cloud interference are attenuated, and the weights of channels related to landslide features are strengthened to obtain a refined channel weight vector.

[0049] The refined channel weight vector and the initial feature are multiplied along the channel dimension in a channel-by-channel weighted manner to obtain the enhanced features for each time phase.

[0050] In one embodiment, the step of fusing multi-temporal spatiotemporal correlation features of the enhanced features of each temporal phase through the temporal collaborative attention module to obtain fused features includes:

[0051] The enhanced features of each time phase are spatially flattened and temporally serialized to generate a temporal feature sequence.

[0052] Add a learnable temporal position code to the temporal feature sequence to obtain a feature sequence with temporal information;

[0053] The cloud thickness variation values ​​of each time phase are extracted from the meteorological data to construct a temporal mask matrix that reflects the correlation of cloud shading between time phases;

[0054] Based on the temporal mask matrix, the feature sequence with temporal information is preprocessed by mask weighting to obtain the processed temporal feature sequence.

[0055] The processed temporal feature sequence is input into a multi-head self-attention layer, and the multi-attention head captures spatiotemporal correlation features of different scales of short and long temporal sequences between time phases to obtain multi-scale temporal correlation features.

[0056] The time series weight generator is used to assign time series weights to the multi-scale time series correlation features of each time phase and calculates them to obtain weighted multi-scale time series correlation features. The time series weights are positively correlated with the landslide physical features of the corresponding time phase. The time series weight generator is trained from the multi-dimensional landslide physical feature vector.

[0057] The weighted multi-scale temporal correlation features are reconstructed in dimension and restored in spatial feature. Then, the features are aggregated and the dimensions are unified through a convolutional fusion layer to obtain fused features that integrate spatiotemporal correlation information from multiple phases.

[0058] In one embodiment, the step of upsampling and detail restoration of the fused features through the decoder and the skip connection module to generate a cloudless landslide disaster satellite image includes:

[0059] The fused features are input into the initial upsampling layer of the decoder, and the first dimensionality increase is performed through transposed convolution to obtain full-scale upsampling features. At the same time, the full-scale jump features of the corresponding level of the encoder after being enhanced by the slippery sensitive channel attention module are extracted.

[0060] The cloud interference feature masking process is applied to the full-scale jump feature to obtain the processed full-scale jump feature.

[0061] The processed full-scale jump feature and the full-scale upsampling feature are channel-level stitched and fused to obtain the full-scale fused feature, wherein the full-scale fused feature has the same resolution as the multi-temporal optical satellite image data.

[0062] The full-scale fused features are input into the detail enhancement sub-network, and the features are enhanced through residual convolutional blocks to obtain the detail enhancement feature map;

[0063] The pixel values ​​of the enhanced feature map are normalized and the spectrum is restored to match the spectral features and color distribution of the original satellite image to generate a cloudless landslide disaster satellite image.

[0064] Furthermore, to achieve the above objectives, this application also proposes a satellite image data generation device for landslide disasters based on cloud cover, the device comprising:

[0065] The acquisition module is used to acquire multi-temporal optical satellite image data, multi-source remote sensing data, meteorological data and terrain data of the target area, and calculate the multi-dimensional landslide physical feature vector corresponding to each optical satellite image data in combination with the terrain data. The optical satellite image data is landslide image data obscured by clouds, and the multi-source remote sensing data includes radar data, hyperspectral data and UAV low-altitude remote sensing data.

[0066] The data enhancement module is used to fuse the multi-temporal optical satellite imagery data and the multi-source remote sensing data to perform fine classification of clouds and obtain cloud classification results;

[0067] The fusion module is used to determine the priority processing area based on the cloud classification results, the meteorological data, and the terrain data;

[0068] The result module is used to input the multi-temporal optical satellite image data and multi-dimensional landslide physical feature vector of the priority processing area into the landslide feature enhancement generation network to obtain landslide disaster satellite image data. The landslide feature enhancement generation network includes an encoder, a landslide sensitive channel attention module, a temporal collaborative attention module, a decoder, and a skip connection module.

[0069] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the method for generating satellite image data of landslide disasters based on cloud cover as described above.

[0070] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the method for generating satellite image data of landslide disasters based on cloud cover as described above.

[0071] This application acquires multi-source remote sensing, meteorological, topographic, and multi-temporal optical satellite imagery data of the target area, including data obscured by clouds. It then calculates multi-dimensional landslide physical feature vectors by combining topographic data. The application further integrates multi-source data for fine-grained cloud classification, determining priority processing areas based on cloud, meteorological, and topographic data. The imagery and physical feature vectors of this area are input into a landslide feature enhancement generation network containing a dual-attention module. Through feature extraction, enhancement, fusion, and upsampling restoration, a cloud-free landslide satellite image is generated and its reliability is assessed. This effectively suppresses interference from cloud obscuration and improves the accuracy of satellite landslide image data generated under cloud-obscured conditions. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a flowchart illustrating the first embodiment of the method for generating satellite image data of landslide disasters based on cloud cover in this application.

[0074] Figure 2 This is a flowchart illustrating the second embodiment of the method for generating satellite image data of landslide disasters based on cloud cover in this application.

[0075] Figure 3 This is a schematic diagram of the module structure of the satellite image data generation device for landslide disasters based on cloud cover, according to the first embodiment of the method for generating satellite image data for landslide disasters based on cloud cover in this application.

[0076] Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the method for generating satellite image data of landslide disasters based on cloud cover in the embodiments of this application.

[0077] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0078] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0079] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0080] In practical applications, the quality of optical satellite imagery is severely constrained by weather conditions, with cloud cover being one of the main factors affecting the usability of optical satellite images. When clouds cover landslide areas, satellites cannot acquire clear surface images, leading to interruptions in the identification and monitoring of landslide disasters and seriously affecting the timeliness and accuracy of disaster emergency response.

[0081] Currently, traditional methods for dealing with cloud cover issues mainly include the following: First, using single-frame image information for cloud detection and removal, and then filling in the cloud-covered areas using image inpainting algorithms; second, employing multi-temporal image fusion technology to complement information from images acquired at different times; and third, introducing deep learning networks for end-to-end cloud removal and image reconstruction. In addition, some methods combine active remote sensing data sources such as radar data, utilizing their ability to penetrate clouds to obtain surface information. However, methods relying solely on single-frame image information cannot effectively handle large-area cloud cover, and the restored images often lack realistic surface texture features. Existing methods are mostly designed for general scenarios and are insufficient in extracting the specific texture features of landslides, making it difficult to accurately identify the boundaries and morphology of landslide bodies. Traditional methods ignore the spatiotemporal correlation between multi-temporal images, failing to fully utilize the temporal patterns of landslide evolution. Furthermore, existing technologies lack refined processing mechanisms for changes in cloud thickness, cannot adaptively adjust processing strategies according to dynamic changes in clouds, and do not introduce uncertainty analysis mechanisms to evaluate the reliability of the processing results. Therefore, how to effectively address the difficulty in identifying landslide hazards in satellite imagery under cloud cover conditions has become an urgent problem to be solved.

[0082] Based on the above, this application also provides a method for generating satellite image data of landslide disasters based on cloud cover, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the method for generating satellite image data of landslide disasters based on cloud cover in this application.

[0083] In this embodiment, the method for generating satellite image data of landslide disasters based on cloud cover includes steps S10 to S40:

[0084] Step S10: Acquire multi-temporal optical satellite imagery data, multi-source remote sensing data, meteorological data, and topographic data of the target area, and calculate the multi-dimensional landslide physical feature vector corresponding to each optical satellite imagery data in combination with the topographic data.

[0085] It should be noted that the optical satellite imagery data is landslide imagery data obscured by cloud cover, while the multi-source remote sensing data includes radar data, hyperspectral data, and UAV low-altitude remote sensing data. The radar data is SAR data.

[0086] Specifically, multi-temporal optical satellite imagery data of the same geographical area collected at different times is acquired from a satellite data receiving platform, with the time interval preferably set to 5 to 10 days to ensure the capture of the dynamic process of landslide evolution. Simultaneously, radar data is acquired from radar satellites, hyperspectral data from hyperspectral satellites, and low-altitude remote sensing data from UAV remote sensing platforms, forming a multi-source data complementarity to compensate for the limitations of single optical sensors in cloud penetration and spectral resolution. Secondly, meteorological data such as cloud height, cloud thickness, cloud movement speed, and cloud type are acquired from meteorological department interfaces, and topographic data such as slope, aspect, elevation, and topographic relief are acquired from a geographic information system to establish a foundation for the correlation between dynamic cloud changes and topographic features. Then, for each temporal phase of optical satellite imagery data, a multi-dimensional landslide physical feature vector is calculated in conjunction with topographic data. This multi-dimensional landslide physical feature vector includes exposed soil characteristics, vegetation destruction characteristics, and textural directionality characteristics. Specifically, this includes: calculating the normalized bare soil index using surface reflectance in shortwave infrared band 1 and shortwave infrared band 2 to quantify the degree of soil exposure caused by landslides. The specific formula is as follows:

[0087]

[0088] in, This represents the surface reflectance of shortwave infrared band 1. This represents the surface reflectance of the shortwave infrared band 2. The shortwave infrared band is sensitive to soil moisture and mineral composition. Landslides causing vegetation destruction and soil exposure will significantly change this index value.

[0089] Multi-temporal normalized vegetation destruction characteristics are calculated using reflectance in the near-infrared and red light bands to capture vegetation destruction information caused by landslides. The specific formula is as follows:

[0090]

[0091] in Indicates near-infrared reflectivity. Indicates the reflectivity in the red light band. Indicates the time interval between adjacent time phases. Indicates phase Relative to time phase The rate of change of NDVI Indicates phase Normalized Difference Vegetation Index (NDVI) Indicates phase Normalized Difference Vegetation Index (NDVI).

[0092] A Gabor filter bank is used to extract texture directionality features to identify the directional fracturing texture pattern unique to landslides. Specifically, a two-dimensional Gabor filter bank is first constructed, which contains multiple filter units with different orientations and scales. Each Gabor filter is defined by the following parameters:

[0093] Directional parameters: Set N uniformly distributed directional angles θ, preferably N=4, corresponding to 0°, 45°, 90° and 135° directions respectively, to cover all major texture directions that may occur in landslides.

[0094] Scale parameters: Set M different spatial frequencies (wavelength λ), preferably M=2, corresponding to short wavelength (high frequency) and long wavelength (low frequency) scales respectively, so as to capture fine texture and macro texture features simultaneously.

[0095] The formula for expressing a filter:

[0096]

[0097] in, For filter space coordinates, For the filter direction, The wavelength of the sinusoidal component controls the filter size. Indicates phase shift, The aspect ratio of the space. The standard deviation of the Gaussian envelope is given. For the grayscale image of the input image, a two-dimensional convolution is performed with a Gabor filter to obtain a set of filter response maps. For each pixel location... Calculate its maximum response intensity in each direction, and denote the maximum absolute value of the scale-filtered response in that direction as . Let the response diagrams of N directions be denoted as Statistical features characterizing landslide texture are extracted from the directional response maps, and each directional response map is divided into several local regions. For each local region, the mean of each directional response map is calculated:

[0098]

[0099] in, Indicates a local area. Let be the number of pixels within the region. For each local area, calculate the standard deviation of the response map for each orientation:

[0100]

[0101] For each local area, calculate the directional consistency index:

[0102]

[0103] in It represents a relatively small positive number to prevent the denominator from being 0.

[0104] Based on the above, a multi-dimensional landslide physical feature vector containing bare soil characteristics, vegetation destruction characteristics, and texture directionality characteristics is constructed, providing landslide-sensitive prior knowledge guidance for subsequent network processing.

[0105] Step S20: The cloud layer is finely classified by fusing multi-temporal optical satellite imagery data and multi-source remote sensing data to obtain the cloud layer classification result.

[0106] It should be noted that step S20 includes: performing radiometric correction and geometric registration on multi-temporal optical satellite image data to obtain a first preprocessed image; performing penetration enhancement processing on radar data to obtain a second preprocessed image; performing spectral feature demixing on hyperspectral data to obtain a third preprocessed image; performing resolution alignment and texture enhancement on UAV low-altitude remote sensing data to obtain a fourth preprocessed image; performing multimodal feature-level fusion on the first, second, third, and fourth preprocessed images to obtain a fused feature map; and inputting the fused feature map into a multi-scale feature extraction network for cloud fine classification. Branching is performed to obtain candidate cloud features with different spatial resolutions. These candidate features are then input into the spectral-texture bimodal attention module of the cloud fine classification network. Attention weights are generated using the spectral dimension of the hyperspectral data and the texture dimension of the UAV data. The attention weights are then weighted and fused with the candidate cloud features to obtain enhanced cloud discrimination features. These enhanced cloud discrimination features are then input into the classification head of the cloud fine classification network to obtain the probability distribution of each pixel belonging to thin clouds, thick clouds, cumulus clouds, or cirrus clouds. Based on the probability distribution, cloud classification results containing thin clouds, thick clouds, cumulus clouds, and cirrus clouds are generated.

[0107] It's important to understand that radiometric correction eliminates radiometric errors caused by factors such as sensor characteristics, atmospheric scattering and absorption, and topographic lighting during satellite image acquisition. This process ensures that image pixel values ​​accurately reflect the actual radiometric brightness of ground features. The corrected image's radiometric information is accurate and comparable, laying the foundation for subsequent feature extraction. Geometric registration matches remote sensing images acquired from different sensors and at different times to the same geographic coordinate system. This corrects geometric distortions, ensuring that the spatial location, shape, and size of ground features in the image match the actual geographic space, guaranteeing spatial consistency in multi-source data fusion. Penetration enhancement, tailored to the characteristics of radar data, uses signal enhancement algorithms to improve the penetration ability of radar waves through clouds, vegetation, and other obstructions. This strengthens the effective signal of ground targets in the radar data, suppresses background noise, and allows the radar data to more clearly reflect the true surface features of the landslide area. Spectral feature demixing addresses the characteristics of hyperspectral data, which has multiple bands and rich spectral information. It decomposes mixed pixels into pure land cover endmembers and their corresponding abundances, separating the spectral features of different land cover elements in cloud and landslide areas, thus eliminating the interference of mixed pixels on cloud classification and landslide feature extraction. Resolution alignment adjusts the spatial resolution of UAV low-altitude remote sensing data to be consistent with that of optical satellite imagery, radar data, and hyperspectral data. Texture enhancement uses image processing algorithms to enhance the texture details of land cover elements in UAV low-altitude remote sensing data, highlighting unique textures such as directional fracturing textures and surface undulations of landslide bodies, making the texture information of land cover elements clearer and providing richer detailed information for cloud classification and landslide feature recognition. Multimodal feature-level fusion is the process of fusing remote sensing data acquired by different types of sensors at the feature level after feature extraction. It is not a simple pixel-level stitching, but rather integrates the unique features of each modality of data, namely the spectral texture features of optical satellites, the penetrating terrain features of radar, the fine spectral features of hyperspectral data, and the high-resolution detail features of UAV data, forming a more comprehensive fused feature. The Cloud Fine-Grain Classification Network is a deep learning network specifically designed for cloud classification in remote sensing imagery. It includes a multi-scale feature extraction branch, a spectral-texture bimodal attention module, and a classification head structure. It can accurately identify different types of clouds in images, achieving fine-grained cloud classification. The multi-scale feature extraction branch is the network structure in the Cloud Fine-Grain Classification Network used to extract cloud features at different spatial resolutions. Through convolutional kernels of different sizes and downsampling operations, it captures shallow, fine-grained features of clouds, such as edges and texture details, as well as deep, coarse-grained features, such as the overall shape and distribution patterns of the clouds. The spectral-texture bimodal attention module is the module in the Cloud Fine-Grain Classification Network used to enhance key features. It can automatically focus on the spectral dimension features of hyperspectral data and the texture dimension features of UAV data, assigning differentiated attention weights to different features to highlight valuable feature information for cloud classification.The classification head is the final layer of the cloud fine-grained classification network. It typically consists of fully connected layers and activation functions, mapping the enhanced cloud discrimination features to probability distributions for different cloud types, thus enabling the classification of the cloud type for each pixel. The probability distribution refers to the set of probability values ​​for each pixel in the image to be classified as one of four cloud types: thin cloud, thick cloud, cumulus cloud, or cirrus cloud. Each pixel corresponds to four probability values, and the sum of these probabilities is 1, directly reflecting the confidence level of the cloud type to which the pixel belongs.

[0108] Specifically, firstly, radiometric correction is performed on multi-temporal optical satellite imagery data using a radiative transfer model to eliminate atmospheric scattering and sensor response errors. Then, geometric registration is performed using polynomial transformation and control point matching to unify geographic coordinates and eliminate geometric distortion, resulting in the first pre-processed image. This ensures the radiometric accuracy and spatial consistency of the optical imagery. Secondly, for radar data (SAR data), speckle suppression and penetration enhancement algorithms are used. Lee filtering removes speckle noise, and a penetration gain function enhances the ability to penetrate cloud layers and vegetation, resulting in the second pre-processed image. This strengthens the radar data's ability to characterize obscured areas. Next, spectral feature demixing is performed on the hyperspectral data. Endmember extraction and abundance inversion algorithms are used to separate the pure spectral components from the mixed pixels, resulting in the third pre-processed image to obtain rich spectral diagnostic information for distinguishing different cloud types. Finally, resolution alignment and texture enhancement are performed on UAV low-altitude remote sensing data. Super-resolution reconstruction improves spatial resolution, and edge enhancement algorithms highlight texture details, resulting in the fourth pre-processed image, providing high spatial accuracy local texture information. Subsequently, the first, second, third, and fourth preprocessed images are fused using multimodal feature-level fusion. By integrating complementary information from optical, radar, hyperspectral, and UAV data through feature stitching and cross-modal attention mechanisms, a fused feature map is obtained to construct a comprehensive representation encompassing spectral, texture, penetration, and spatial details. This fused feature map is then input into the multi-scale feature extraction branch of the cloud fine-grained classification network. Parallel convolution and pooling operations capture cloud morphology information at different scales, resulting in candidate cloud features with different spatial resolutions to accommodate the high variability in cloud size and shape. These candidate cloud features with different spatial resolutions are then input into the spectral-texture bimodal attention module of the cloud fine-grained classification network. The spectral dimension information from the hyperspectral data is used to identify the spectral features of cloud types, and the texture dimension information from the UAV data is used to capture the fine structure of cloud edges, generating attention weights to highlight cloud features and suppress background interference. Finally, the attention weights are weighted and fused with the candidate cloud features to obtain enhanced cloud discrimination features, improving the separability between cloud types. The enhanced cloud discrimination features are then input into the classification head of the cloud fine-grained classification network. A fully connected layer and a normalized exponential function are used to calculate the probability distribution of each pixel belonging to thin clouds, thick clouds, cumulus clouds, or cirrus clouds, quantifying classification uncertainty. Finally, based on the probability distribution, the category corresponding to the highest probability is selected as the cloud type label for each pixel, generating cloud classification results including thin clouds, thick clouds, cumulus clouds, and cirrus clouds. This provides refined cloud spatial distribution information for determining subsequent priority processing areas.

[0109] Step S30: Based on the cloud classification results, meteorological data, and topographic data, determine the priority processing areas.

[0110] It should be noted that step S30 includes: extracting topographic features based on topographic data and calculating the landslide risk level of the target area in conjunction with the landslide susceptibility assessment model, wherein the topographic features include slope, aspect, and topographic relief; extracting meteorological features from meteorological data, wherein the meteorological features include cloud height, cloud thickness, cloud movement speed, and cloud coverage duration; configuring cloud influence weights for different types of clouds based on the meteorological features and the cloud classification results; performing a weighted calculation on the cloud influence weights and the corresponding landslide risk level of the area to obtain the coupled risk value of the target area; comparing the coupled risk value with a preset coupled risk threshold to select candidate areas whose coupled risk values ​​are higher than the preset coupled risk threshold; performing connectivity analysis and boundary optimization on the candidate areas to obtain satellite image data of landslide disasters in the target area and generating priority processing areas.

[0111] It is important to understand that topographic data refers to various types of data reflecting the surface topography and geomorphological characteristics of the target area, and is the basic data source for extracting topographic features such as slope and aspect. Topographic features refer to the characteristic information extracted from topographic data that reflects the topographic attributes of the region. A landslide susceptibility assessment model is a professional model used to assess the likelihood of landslide disasters occurring within a target area, and can be combined with topographic features to calculate the landslide risk level. The landslide risk level refers to the classification result calculated by the landslide susceptibility assessment model, characterizing the level of likelihood of landslide disasters occurring in the target area. Meteorological data refers to various types of data reflecting atmospheric conditions and weather changes in the target area, and is the basic data source for extracting meteorological features such as cloud height. Meteorological features refer to the characteristic information extracted from meteorological data that reflects cloud state and changes. The coupled risk value is a numerical value that characterizes the degree of coupling between the cloud cover effect and the likelihood of landslide disasters in the target area, obtained by weighting the cloud influence weight with the landslide risk level. The preset coupled risk threshold is a pre-set critical value for coupled risk to screen high-risk areas, used to define the screening criteria for candidate areas. Candidate regions refer to areas with coupling risk values ​​higher than a preset coupling risk threshold, serving as preliminary screening areas for priority processing of landslide images. Connectivity analysis refers to image processing and analysis operations performed on candidate regions to identify sets of adjacent connected pixels within the region, used to integrate adjacent candidate regions. Boundary optimization refers to processing operations performed on the regions after connectivity analysis to correct the boundary contours of the regions, making the region boundaries more closely resemble the actual geographic space. Priority processing regions refer to the target regions finally determined after completing connectivity analysis and boundary optimization of candidate regions, which require priority processing of landslide disaster satellite image data.

[0112] Specifically, firstly, three types of terrain features—slope, aspect, and topographic relief—are extracted from the terrain data. These extracted features are then input into a landslide susceptibility assessment model to calculate the landslide risk level of the target area. This, combined with the basic terrain attributes, helps determine the likelihood of landslides occurring in the region. Simultaneously, four types of meteorological features—cloud height, cloud thickness, cloud movement speed, and cloud cover duration—are extracted from the meteorological data. Based on these extracted meteorological features and the obtained cloud classification results, corresponding cloud influence weights are assigned to different types of clouds to quantify the actual impact of different clouds on landslide images. Then, the cloud influence weights for each region are weighted according to preset rules and the calculated landslide risk level for that region to obtain the coupled risk value for each region within the target area. This comprehensively measures the combined effect of cloud cover and landslide risk. Finally, the calculated coupled risk values ​​are compared one by one with a pre-set coupled risk threshold. Regions with coupled risk values ​​higher than this threshold are selected as candidate regions, thus identifying high-coupled-risk areas requiring focused attention. Finally, connected component analysis and boundary optimization operations are carried out sequentially on the selected candidate regions. Connected component analysis integrates adjacent candidate regions, and boundary optimization corrects the boundary contours of the regions. Finally, the priority processing areas of the landslide disaster satellite image data of the target region are determined and generated, so that the division of the processing area is more in line with the actual landslide monitoring needs and the targeting and efficiency of subsequent image processing are improved.

[0113] Step S40: Input the multi-temporal optical satellite image data of the priority processing area and the multi-dimensional landslide physical feature vector into the landslide feature enhancement generation network to obtain landslide disaster satellite image data.

[0114] It should be noted that the landslide feature enhancement generation network includes an encoder, a landslide-sensitive channel attention module, a temporal collaborative attention module, a decoder, and a skip connection module. Multi-temporal optical satellite imagery data of the priority processing area is input into the encoder of the landslide feature enhancement generation network. Hierarchical feature extraction is performed through multiple consecutive downsampling convolutional layers to obtain multi-scale depth spatial features under different receptive fields. This is done to effectively capture semantic information at all levels, from macroscopic terrain contours to microscopic surface textures, while compressing spatial redundancy. Secondly, the landslide-sensitive channel attention module is used to fuse multi-dimensional landslide physical feature vectors with depth spatial features at each level. A fully connected layer maps physical attributes into channel dimension weights, thereby dynamically amplifying feature channels related to landslide evolution and suppressing noise channels caused by cloud cover. This is done to introduce external physical priors as strong guidance, correcting the network's bias in landslide feature extraction under thin cloud interference. Then, the enhanced feature map sequences from each time phase are fed into the temporal collaborative attention module. A multi-head self-attention mechanism is used to model the spatiotemporal correlation between images at different time points, and key features are weighted and aggregated based on temporal evolution logic. This is done to fully exploit the spatiotemporal redundancy between multi-time phase data and utilize the dynamic changes in cloud layers to complement and fill in the static landslide features in obscured areas. Finally, the decoder progressively upsamples and restores the fused features, and a skip connection module is used to cascade and fuse the shallow high-frequency details retained in the encoder stage with deep semantic features, thereby generating the final resolution-aligned and significantly enhanced satellite image data of landslide disasters. This is done to maximize the restoration of key micro-topographic information such as the boundaries and cracks of the landslide body while reconstructing the global semantics of the image, ensuring that the generated data can support high-precision disaster identification.

[0115] This embodiment acquires multi-source remote sensing, meteorological, topographic, and multi-temporal optical satellite imagery data of the target area, including data obscured by clouds. It then calculates multi-dimensional landslide physical feature vectors based on the topographic data. The system further integrates multi-source data for fine-grained cloud classification, determining priority processing areas based on cloud, meteorological, and topographic data. The imagery and physical feature vectors of this area are input into a landslide feature enhancement generation network containing a dual-attention module. Through feature extraction, enhancement, fusion, and upsampling restoration, a cloud-free landslide satellite image is generated and its reliability is assessed. This effectively suppresses interference from cloud obscuration and improves the accuracy of satellite landslide image data generated under cloud-obscured conditions.

[0116] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The method for generating satellite image data of landslide disasters based on cloud cover, step S40, further includes steps S201 to S206:

[0117] Step S201: Multi-scale spatial feature extraction is performed on multi-temporal optical satellite image data using an encoder to obtain initial features.

[0118] It should be noted that step S201 includes: performing block-based and dimensionality-standardized processing on the multi-temporal optical satellite image data of the priority processing area to obtain normalized temporal image data; inputting the normalized temporal image data into the multi-scale convolutional group of the encoder, and extracting shallow fine-grained features and deep coarse-grained features through parallel convolution with different kernel sizes; performing cross-scale fusion on the shallow fine-grained features and deep coarse-grained features, and obtaining multi-scale fused features through feature stitching and compression; performing temporal dimension feature aggregation on the multi-scale fused features to obtain aggregated multi-scale fused features; and inputting the aggregated multi-scale fused features into the downsampling layer of the encoder for dimensionality reduction processing to obtain the initial features of each level.

[0119] Understandingly, block processing refers to dividing ultra-large format remote sensing satellite imagery into several small image patches with fixed pixel sizes to facilitate batch computation and memory management by deep learning models. Dimensionality normalization refers to the preprocessing process of adjusting image data from different sources or batches to a uniform spatial resolution, image height, and width through techniques such as scaling, padding, or resampling. Multi-scale convolutional groups refer to structures that set multiple convolutional kernel branches with different receptive field sizes in the same layer of a neural network to simultaneously capture feature information of objects at different scales in the image. Fine-grained features refer to the feature representation of high-frequency information such as object edges, textures, and minute details contained in the image. Coarse-grained features refer to the low-frequency semantic information in the image that reflects the overall outline, shape, and macroscopic spatial positional relationships of objects. Cross-scale fusion refers to the technique of stitching or weighting feature maps from different receptive fields or different depth levels to achieve complementarity between spatial detail information and deep semantic information. Downsampling layers refer to network modules that significantly reduce the spatial dimension of feature maps while extracting more abstract, higher-level features by increasing the convolution stride or using pooling operators. Initial features refer to the feature representations that have been initially extracted by the encoder but have not yet undergone weight calibration by the landslide-sensitive channel attention module or spatiotemporal correlation modeling by the temporal co-attention module.

[0120] Specifically, firstly, the multi-temporal optical satellite imagery data of the priority processing area is segmented and standardized at a preset spatial resolution (e.g., 256×256) to obtain regularized temporal imagery data. This is done to reduce the pressure on hardware computing resources from ultra-large images and ensure that the data at different time points have a completely consistent shape at the network input. Secondly, the regularized temporal imagery data is input into the encoder's multi-scale convolutional group. Using a preset first convolutional kernel size (3×3 in this embodiment) and a preset second convolutional kernel size (5×5 in this embodiment), shallow fine-grained features and deep coarse-grained features are extracted respectively. These features are then concatenated and compressed at a preset compression rate to achieve cross-scale fusion, resulting in multi-scale fused features. This is done to... While capturing microscopic details such as cracks at the edge of landslide bodies, the model also acquires macroscopic features of large-scale terrain contours to improve its adaptability to disasters of different scales. Then, the multi-scale fusion features are aggregated in the form of mean or maximum value along the time axis to obtain aggregated multi-scale fusion features. This is done to utilize redundant information between multiple time phases to initially suppress random noise and transient thin cloud interference in single-frame images. Finally, the aggregated multi-scale fusion features are input into the downsampling layer of the encoder and dimensionality is reduced through convolution operations with a preset stride (e.g., 2) to obtain the initial features of each level. This is done to gradually reduce the spatial size of the feature map and increase the channel depth while preserving key semantic information, laying the foundation for subsequent fine-grained enhancement of deep features.

[0121] Step S202: The initial features are enhanced at the channel level using the landslide-sensitive channel attention module with multi-dimensional landslide physical feature vectors to obtain the enhanced features for each time phase.

[0122] It should be noted that step S202 includes: inputting the multi-dimensional landslide physical feature vector into a two-layer fully connected mapping network, performing dimensionality upscaling and feature decoupling sequentially to obtain an adapted feature vector, wherein the adapted feature vector has the same number of initial feature channels; performing global average pooling and global max pooling on the initial features to fuse them into a dual-channel statistical vector; summing and fusing the adapted feature vector and the dual-channel statistical vector element-wise to obtain a fused channel guiding vector; performing nonlinear mapping on the fused channel guiding vector through an activation function to obtain initial channel weights; introducing a cloud occlusion mask to correct the initial channel weights, attenuating the weights of channels related to cloud interference, and strengthening the weights of channels related to landslide features to obtain a refined channel weight vector; and performing a weighted multiplication of the refined channel weight vector and the initial features along the channel dimension to obtain the enhanced features for each time phase.

[0123] It's important to understand that a two-layer fully connected mapping network refers to a computational structure composed of two layers of neurons connected in a fully connected manner. This structure maps an input vector to another feature space through linear transformation and non-linear activation, achieving dimensionality adjustment and feature representation transformation. Feature decoupling refers to decomposing a complex, fused feature vector into independent, non-interfering components. An adapted feature vector, after dimensionality transformation and feature processing, is a guiding vector whose length matches the channel dimension of the target feature map and whose semantic information is optimized. Dual pooling refers to the operation of simultaneously using global average pooling and global max pooling techniques to extract global background information and the most salient feature information from the feature map. A dual-channel statistical vector is a vector generated by dual pooling that contains channel-level mean and maximum statistical information, used to characterize the activation intensity of each feature channel globally. A fusion channel guiding vector is a vector obtained by fusing the physical prior features of the external input with the statistical features of the image itself, used to guide the attention mechanism to generate precise weight allocation. A cloud occlusion mask is a pixel-level binarization or probability matrix used to identify the specific spatial location and degree of occlusion of cloud-occluded areas in an image, serving as a reference for subsequent feature restoration. Channel dimension refers to the third dimension of a convolutional neural network feature map beyond spatial coordinates; each channel typically represents a specific feature attribute extracted by a particular filter. Channel-wise weighted multiplication applies each element of the calculated weight vector to the corresponding feature channel in the feature map, achieving targeted scaling or suppression of different feature attributes.

[0124] Specifically, firstly, the multi-dimensional landslide physical feature vector is input into a two-layer fully connected mapping network consisting of a preset first upscaling dimension (e.g., 1, 2, 3) and a preset second decoupling dimension (e.g., 2, 5, 6). The upscaling transformation of the feature space and the decoupling calculation of the latent features are performed sequentially, outputting an adapted feature vector with the same number of initial feature channels. This is done to map low-dimensional physical prior attributes to a high-dimensional semantic space, aligning it with the deep features extracted by the convolutional neural network in both dimension and logic. Secondly, parallel dual-pooling operations of global average pooling and global max pooling are performed simultaneously on the initial features, and the obtained mean and maximum values ​​are fused to obtain a dual-channel statistical vector reflecting the global features of each channel. This is done to simultaneously capture the background consistency and local saliency of image features, thereby more comprehensively describing the importance of each feature channel. Then, the adapted feature vector and the dual-channel statistical vector are fused element-wise to obtain a fused channel guidance vector. This vector is then non-linearly mapped using an activation function to output initial channel weights that can initially distinguish the contribution of features. This is done to use external physical guidance signals to filter internal features and initially suppress redundant information unrelated to landslides. Finally, the initial channel weights are dynamically corrected based on the spatial location association according to the cloud occlusion mask. By reducing the gain of channels affected by cloud interference and strengthening the gain of landslide-sensitive channels, a refined channel weight vector is obtained. This vector is then multiplied pixel-wise with the initial features along the channel dimension to obtain the enhanced features for each phase. This is done to use the occlusion information in the spatial dimension to further filter noise and ensure that the network can accurately supplement information and enhance the saliency of landslide areas obscured by clouds at the feature level.

[0125] Step S203: The temporal collaborative attention module is used to perform multi-temporal spatiotemporal correlation feature fusion on the enhanced features of each temporal phase to obtain fused features.

[0126] It should be noted that step S203 includes: spatially flattening and temporally serializing the enhanced features of each time phase to generate a temporal feature sequence. Specifically, the enhanced features of each time phase... Flattened in spatial dimension ,in Given the total number of spatial locations, we obtain the temporal feature sequence.

[0127] Learnable temporal position codes are added to the temporal feature sequences to obtain feature sequences with temporal information. Specifically, for each... Add learnable time-location coding To obtain features containing temporal sequence information Then, all the temporally encoded features are concatenated along the sequence dimension to obtain... .

[0128] The cloud thickness variation values ​​for each time phase are extracted from meteorological data to construct a temporal mask matrix reflecting the correlation of cloud occlusion between time phases. Specifically, the cloud thickness variation values ​​corresponding to each time phase are extracted from the acquired meteorological data to obtain information on the variation of cloud occlusion intensity at different time phases. Next, the extracted cloud thickness variation values ​​for each time phase are arranged in chronological order to construct a temporal mask matrix that reflects the correlation between the strength of cloud occlusion between time phases. Finally, a temporal mask matrix that can be used for subsequent weighted preprocessing of temporal features is obtained. , Each element Reflecting the phase When the time comes Changes in cloud thickness. From a temporal perspective... When the time comes As the cloud layer thins, Set it to a large positive value; otherwise, set it to a small value. This mask is expanded and added to the attention score matrix, thereby guiding the model to pay more attention to past or future phases with thinner cloud layers and clearer information compared to the current phase when fusing information.

[0129] Then, based on the temporal mask matrix, the feature sequence with temporal information is preprocessed with mask weighting to obtain the processed temporal feature sequence. The processed temporal feature sequence is input into a multi-head self-attention layer, which captures different scale spatiotemporal correlation features between short and long time phases through multiple attention heads, resulting in multi-scale temporal correlation features. A temporal weight generator is used to assign temporal weights to the multi-scale temporal correlation features of each time phase and calculates them to obtain weighted multi-scale temporal correlation features, where the temporal weights are positively correlated with the corresponding time phase landslide physical features. The temporal weight generator is trained from multi-dimensional landslide physical feature vectors. The weighted multi-scale temporal correlation features are then reconstructed in dimension and restored in spatial features, and then aggregated and unified in dimension through a convolutional fusion layer to obtain fused features that integrate spatiotemporal correlation information from multiple time phases. Specifically, based on the temporal mask matrix, mask weighting preprocessing is performed on the feature sequence with temporal information to reduce the feature weights of time phases with large cloud thickness variations and increase the feature weights of stable time phases, thereby reducing the interference of cloud fluctuations on the temporal features. Secondly, the processed temporal feature sequence is input into a multi-head self-attention layer. Multiple attention heads are used to extract spatiotemporal correlation features at different scales for short and long time periods, resulting in multi-scale temporal correlation features. This comprehensively captures the changing patterns of landslide features across different time spans. Then, a temporal weight generator trained from multi-dimensional landslide physical feature vectors is used to assign temporal weights positively correlated with the physical features of the landslide at each time phase to the multi-scale temporal correlation features, and weighted calculations are performed to obtain weighted multi-scale temporal correlation features. This allows the network to focus more on temporal data where landslide information is significant. Finally, the weighted multi-scale temporal correlation features undergo dimensional reconstruction and spatial feature restoration, and are then input into a convolutional fusion layer to complete feature aggregation and dimensional unification, resulting in fused features that integrate spatiotemporal correlation information from multiple time phases. This restores the temporal features to a spatial feature form suitable for image reconstruction.

[0130] Step S204: Upsample and restore details of the fused features through the decoder and skip connection module to generate a cloudless landslide disaster satellite image.

[0131] It should be noted that step S204 includes: inputting the fused features into the initial upsampling layer of the decoder, performing the first dimensionality increase through transposed convolution to obtain full-scale upsampling features, and simultaneously extracting the full-scale jump features of the corresponding layer of the encoder after enhancement by the landslide-sensitive channel attention module; performing cloud interference feature masking removal on the full-scale jump features to obtain processed full-scale jump features; performing channel-level stitching and fusion of the processed full-scale jump features and the full-scale upsampling features to obtain full-scale fused features, wherein the full-scale fused features have the same resolution as the multi-temporal optical satellite image data; inputting the full-scale fused features into the detail enhancement sub-network, enhancing the features through residual convolution blocks to obtain a detail enhancement feature map; performing pixel value normalization and spectral restoration processing on the detail enhancement feature map, matching the spectral features and color distribution of the original satellite image to generate a cloudless landslide disaster satellite image.

[0132] It's important to understand that, firstly, the fused features are input into the decoder's initial upsampling layer. An initial dimensionality increase is performed using transposed convolutions with a preset kernel size (e.g., 4×4) and a preset stride (e.g., 2), yielding full-scale upsampling features. Simultaneously, full-scale skip features enhanced by the landslide-sensitive channel attention module at the corresponding encoder level are extracted. This is done to restore the image's spatial resolution while recovering high-frequency geometric structures and texture information lost during encoding downsampling through cross-layer connections. Secondly, using a mask matrix generated from the cloud fine-classification results, cloud interference feature masking is applied to the full-scale skip features to obtain processed full-scale skip features. These processed features are then concatenated with the full-scale upsampling features along the channel dimension to obtain full-scale fused features. This is done to filter out residual cloud noise in the original features, ensuring the decoder uses only pure landslide surface features for image reconstruction. Then, the full-scale fused features are input into the detail enhancement subnetwork. The features are non-linearly enhanced using residual convolutional blocks with multiple preset depths (e.g., 3 layers) to obtain a detail enhancement feature map. This is done to compensate for the blurring effect caused by the upsampling process using a residual learning mechanism, further highlighting key micro-features of the landslide body such as cracks and deposits. Finally, the detail enhancement feature map is normalized for pixel values ​​and spectral restoration is performed using the spectral statistical features of the original image to generate the final cloudless landslide disaster satellite image. This ensures that the generated image not only visually eliminates occlusion but also maintains consistency with the original satellite image in color distribution and spectral radiation characteristics, thus providing a highly realistic base map for quantitative disaster analysis.

[0133] Step S205: Conduct a credibility assessment on the cloudless landslide disaster satellite image to obtain the credibility assessment result.

[0134] It should be noted that step S205 includes: constructing a landslide feature benchmark library, extracting benchmark feature sets of exposed soil, vegetation destruction, and texture directionality from cloudless real landslide images in the library; performing same-dimensional feature extraction on cloudless landslide disaster satellite images to obtain a landslide feature set to be evaluated, and calculating the pixel-by-pixel feature similarity with the benchmark feature set; configuring differentiated weights for the affected areas of different cloud types based on cloud classification results, and performing weighted correction on the pixel-by-pixel feature similarity to obtain a weighted similarity matrix; comparing the weighted similarity matrix with a preset similarity threshold, dividing high-confidence areas and low-confidence areas and calculating the proportion of each area; and integrating the global mean of the weighted similarity matrix, the proportion of high and low confidence areas, and the landslide feature matching degree to calculate a comprehensive confidence value as the confidence evaluation result.

[0135] Specifically, a landslide feature benchmark library is constructed. This is achieved by retrieving a large number of historical, cloudless, real landslide images from the library and extracting features such as exposed soil characteristics, vegetation destruction characteristics, and textural directionality. This establishes a benchmark feature set, which aims to establish objective identification standards using real geological disaster samples, providing a ground truth for subsequent assessments. Secondly, features are extracted in the same dimension from the generated cloudless landslide disaster satellite images to obtain the landslide feature set to be assessed. The pixel-by-pixel feature similarity between this set and the benchmark feature set is calculated. Then, combined with the output cloud classification results, differentiated correction weights are assigned to areas with different cloud cover intensities, such as thin clouds and thick clouds, to further refine the pixel-by-pixel feature similarity. A weighted similarity matrix is ​​obtained by performing weighted correction. This is done to quantify the accuracy of the generated image in restoring the physical characteristics of landslides under different occlusion backgrounds and to effectively distinguish different areas affected by reconstruction difficulty. Finally, the weighted similarity matrix is ​​compared with a preset similarity threshold to divide high-confidence and low-confidence regions and calculate their area proportions. Then, the global mean of the weighted similarity matrix, the region proportions, and the landslide feature matching degree are integrated to calculate a comprehensive confidence value as the final evaluation result. This is done to provide users with an intuitive and quantitative image reliability score through the weighted integration of multi-dimensional indicators, thereby guiding the decision-making logic of subsequent disaster monitoring work.

[0136] Step S206: When the credibility assessment result exceeds the preset credibility threshold, the cloudless landslide disaster satellite image is used as the final landslide disaster satellite image data.

[0137] Specifically, firstly, the obtained credibility assessment results and preset credibility thresholds are retrieved and compared one by one to determine whether the credibility assessment result exceeds the threshold. This is done to filter out cloudless landslide disaster satellite images that meet the quality requirements. Secondly, when it is confirmed that the credibility assessment result exceeds the preset credibility threshold, the cloudless landslide disaster satellite image is designated as the final landslide disaster satellite image data, and the credibility assessment information of the image is marked. This is done to confirm that the image quality meets the standards and to ensure that subsequent landslide monitoring and analysis can be based on reliable data. Finally, the determined final landslide disaster satellite image data is formatted and stored for easy retrieval, viewing, and application later. This is done to ensure the standardization and accessibility of the data, providing stable data support for subsequent work such as landslide disaster early warning and prevention.

[0138] This embodiment extracts multi-scale spatial features from multi-temporal cloud-obscured optical satellite images using an encoder to obtain initial features; it then uses a landslide-sensitive channel attention module combined with multi-dimensional landslide physical feature vectors to enhance the initial features; the temporal collaborative attention module fuses multi-temporal spatiotemporal correlation features, and the decoder and skip connection module upsample and restore details to generate cloudless images; the image is then evaluated for credibility, and if it meets the standards, it is used as the final landslide disaster satellite image data, accurately extracting and enhancing landslide features, suppressing cloud interference, and improving the quality and credibility of cloudless image generation.

[0139] Based on the first embodiment of this application, this application also provides a device for generating satellite image data of landslide disasters based on cloud cover. Please refer to [link / reference]. Figure 3 The device includes:

[0140] The acquisition module 10 is used to acquire multi-temporal optical satellite image data, multi-source remote sensing data, meteorological data and topographic data of the target area, and calculate the multi-dimensional landslide physical feature vector corresponding to each optical satellite image data in combination with the topographic data. The optical satellite image data is landslide image data obscured by clouds, and the multi-source remote sensing data includes radar data, hyperspectral data and UAV low-altitude remote sensing data.

[0141] The data augmentation module 20 is used to fuse multi-temporal optical satellite imagery data and multi-source remote sensing data to perform fine classification of clouds and obtain cloud classification results.

[0142] The fusion module 30 is used to determine the priority processing area based on cloud classification results, meteorological data, and terrain data.

[0143] The result module 40 is used to input multi-temporal optical satellite image data of the priority processing area and multi-dimensional landslide physical feature vectors into the landslide feature enhancement generation network to obtain landslide disaster satellite image data. The landslide feature enhancement generation network includes an encoder, a landslide sensitive channel attention module, a temporal collaborative attention module, a decoder, and a skip connection module.

[0144] The cloud-obscured satellite image data generation device for landslide disasters provided in this application employs the cloud-obscured satellite image data generation method described in the above embodiments, and can solve the technical problem of how to improve the accuracy of satellite landslide image data generated under cloud obscuration conditions. Compared with the prior art, the beneficial effects of the cloud-obscured satellite image data generation device for landslide disasters provided in this application are the same as those of the cloud-obscured satellite image data generation method described in the above embodiments, and other technical features in the cloud-obscured satellite image data generation device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0145] This application provides a satellite image data generation device for landslide disasters based on cloud cover. The device includes: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the satellite image data generation method for landslide disasters based on cloud cover described in Embodiment 1 above.

[0146] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a cloud-based landslide disaster satellite image data generation device suitable for implementing embodiments of this application. The cloud-based landslide disaster satellite image data generation device in this application embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 4 The illustrated satellite imagery data generation device for landslide disasters based on cloud cover is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments in this application.

[0147] like Figure 4As shown, the cloud-obscured landslide disaster satellite imagery data generation device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the cloud-obscured landslide disaster satellite imagery data generation device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the cloud-based landslide disaster satellite imagery data generation device to wirelessly or wiredly communicate with other devices to exchange data. Although various cloud-based landslide disaster satellite imagery data generation devices are shown in the figures, it should be understood that it is not required to implement or possess all of them. More or fewer may be implemented alternatively.

[0148] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0149] The cloud-obscured satellite image data generation device for landslide disasters provided in this application employs the cloud-obscured satellite image data generation method described in the above embodiments, and can solve the technical problem of how to improve the accuracy of satellite landslide image data generated under cloud obscuration conditions. Compared with the prior art, the beneficial effects of the cloud-obscured satellite image data generation device for landslide disasters provided in this application are the same as those of the cloud-obscured satellite image data generation method described in the above embodiments, and other technical features in this cloud-obscured satellite image data generation device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0150] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0151] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0152] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the method for generating satellite image data of landslide disasters based on cloud cover in the above embodiments.

[0153] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0154] The aforementioned computer-readable medium may be included in a cloud-shaded satellite imagery data generation device for landslide disasters; or it may exist independently and not assembled into a cloud-shaded satellite imagery data generation device for landslide disasters.

[0155] The aforementioned computer-readable medium carries one or more programs that, when executed by a cloud-obscured landslide disaster satellite imagery data generation device, enable the cloud-obscured landslide disaster satellite imagery data generation device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0156] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0157] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0158] The readable medium provided in this application is a computer-readable medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described method for generating satellite image data of landslide disasters based on cloud cover. This method can solve the technical problem of how to improve the accuracy of satellite landslide image data generated under cloud cover conditions. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the method for generating satellite image data of landslide disasters based on cloud cover provided in the above embodiments, and will not be repeated here.

[0159] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for generating satellite image data of landslide disasters based on cloud cover.

[0160] The computer program product provided in this application can solve the technical problem of how to improve the accuracy of satellite landslide image data generated under cloud cover conditions. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the cloud cover-based satellite image data generation method for landslide disasters provided in the above embodiments, and will not be repeated here.

[0161] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for generating satellite image data of landslide disasters based on cloud cover, characterized in that, The method includes: Acquire multi-temporal optical satellite imagery, multi-source remote sensing data, meteorological data, and topographic data of the target area, and calculate the multi-dimensional landslide physical feature vector corresponding to each optical satellite imagery in combination with the topographic data. The optical satellite imagery is landslide imagery data obscured by cloud cover, and the multi-source remote sensing data includes radar data, hyperspectral data, and UAV low-altitude remote sensing data. By fusing the multi-temporal optical satellite imagery data and the multi-source remote sensing data, the clouds are finely classified to obtain cloud classification results. Based on the cloud classification results, the meteorological data, and the terrain data, priority processing areas are determined; The multi-temporal optical satellite image data and multi-dimensional landslide physical feature vector of the priority processing area are input into the landslide feature enhancement generation network to obtain landslide disaster satellite image data. The landslide feature enhancement generation network includes an encoder, a landslide sensitive channel attention module, a temporal collaborative attention module, a decoder and a skip connection module. The step of fusing the multi-temporal optical satellite imagery data and the multi-source remote sensing data to perform fine classification of clouds and obtain cloud classification results includes: Radiometric correction and geometric registration are performed on the multi-temporal optical satellite image data to obtain the first preprocessed image; The radar data is subjected to penetration enhancement processing to obtain a second pre-processed image; The hyperspectral data is demixed based on spectral features to obtain a third preprocessed image; The low-altitude remote sensing data from the UAV is subjected to resolution alignment and texture enhancement to obtain a fourth preprocessed image. The first preprocessed image, the second preprocessed image, the third preprocessed image, and the fourth preprocessed image are fused at the multimodal feature level to obtain a fused feature map; The fused feature map is input into the multi-scale feature extraction branch of the cloud fine classification network to obtain cloud candidate features with different spatial resolutions. The candidate cloud features with different spatial resolutions are input into the spectral-texture dual-modal attention module of the cloud fine classification network, and attention weights are generated using the spectral dimension of the hyperspectral data and the texture dimension of the UAV data. The attention weights are weighted and fused with the cloud candidate features to obtain the enhanced cloud discrimination features; The enhanced cloud discrimination features are input into the classification head of the cloud fine classification network to obtain the probability distribution of each pixel belonging to thin cloud, thick cloud, cumulus cloud, and cirrus cloud. Based on the probability distribution, cloud classification results including thin clouds, thick clouds, cumulus clouds, and cirrus clouds are generated; The step of determining the priority processing area based on the cloud classification results, the meteorological data, and the terrain data includes: Based on the terrain data, terrain features are extracted, and the landslide risk level of the target area is calculated by combining the landslide susceptibility assessment model. The terrain features include slope, aspect and topographic relief. Meteorological features are extracted from the meteorological data, including cloud height, cloud thickness, cloud movement speed, and cloud coverage duration. Based on the meteorological characteristics and the cloud classification results, cloud influence weights are assigned to different types of clouds. The cloud layer influence weight and the corresponding landslide risk level of the area are weighted and calculated to obtain the coupled risk value of the target area. The coupling risk value is compared with a preset coupling risk threshold to filter out candidate regions whose coupling risk value is higher than the preset coupling risk threshold; Connectivity analysis and boundary optimization are performed on the candidate regions to obtain satellite image data of landslide disasters in the target region and generate priority processing regions; The step of inputting the multi-temporal optical satellite image data and multi-dimensional landslide physical feature vector of the priority processing area into the landslide feature enhancement generation network to obtain landslide disaster satellite image data includes: The encoder is used to extract multi-scale spatial features from the multi-temporal optical satellite image data to obtain initial features; The landslide-sensitive channel attention module uses the multi-dimensional landslide physical feature vector to perform channel-level directional enhancement on the initial features, resulting in enhanced features for each time phase. The temporal collaborative attention module performs multi-temporal spatiotemporal correlation feature fusion on the enhanced features of each temporal phase to obtain fused features; The decoder and the skip connection module are used to upsample and restore details of the fused features to generate a cloudless landslide disaster satellite image. The credibility assessment of the aforementioned cloudless landslide disaster satellite images was performed, and the credibility assessment results were obtained. When the credibility assessment result exceeds the preset credibility threshold, the cloudless landslide disaster satellite image will be used as the final landslide disaster satellite image data.

2. The method as described in claim 1, characterized in that, The step of using the landslide-sensitive channel attention module to perform channel-level directional enhancement of the initial features using the multi-dimensional landslide physical feature vector to obtain enhanced features for each time phase includes: The multi-dimensional landslide physical feature vector is input into a two-layer fully connected mapping network, and dimensionality increase and feature decoupling are performed sequentially to obtain an adapted feature vector, wherein the adapted feature vector is consistent with the initial feature channel number. The initial features are subjected to dual pooling operations of global average pooling and global max pooling, and then fused to obtain a dual-channel statistical vector. The adaptation feature vector and the dual-channel statistical vector are element-wise summed and fused to obtain the fused channel guidance vector. The initial channel weights are obtained by performing a nonlinear mapping on the fusion channel guide vector using an activation function. The initial channel weights are corrected by introducing a cloud occlusion mask, the weights of channels related to cloud interference are attenuated, and the weights of channels related to landslide features are strengthened to obtain a refined channel weight vector. The refined channel weight vector and the initial feature are multiplied along the channel dimension in a channel-by-channel weighted manner to obtain the enhanced features for each time phase.

3. The method as described in claim 1, characterized in that, The step of fusing multi-temporal spatiotemporal correlation features of the enhanced features of each temporal phase through the temporal collaborative attention module to obtain fused features includes: The enhanced features of each time phase are spatially flattened and temporally serialized to generate a temporal feature sequence. Add a learnable temporal position code to the temporal feature sequence to obtain a feature sequence with temporal information; The cloud thickness variation values ​​of each time phase are extracted from the meteorological data to construct a temporal mask matrix that reflects the correlation of cloud shading between time phases; Based on the temporal mask matrix, the feature sequence with temporal information is preprocessed by mask weighting to obtain the processed temporal feature sequence. The processed temporal feature sequence is input into a multi-head self-attention layer, and the multi-attention head captures spatiotemporal correlation features of different scales of short and long temporal sequences between time phases to obtain multi-scale temporal correlation features. The time series weight generator is used to assign time series weights to the multi-scale time series correlation features of each time phase and calculates them to obtain weighted multi-scale time series correlation features. The time series weights are positively correlated with the landslide physical features of the corresponding time phase. The time series weight generator is trained from the multi-dimensional landslide physical feature vector. The weighted multi-scale temporal correlation features are reconstructed in dimension and restored in spatial feature. Then, the features are aggregated and the dimensions are unified through a convolutional fusion layer to obtain fused features that integrate spatiotemporal correlation information from multiple phases.

4. The method as described in claim 1, characterized in that, The step of upsampling and detail restoration of the fused features through the decoder and the skip connection module to generate a cloudless landslide disaster satellite image includes: The fused features are input into the initial upsampling layer of the decoder, and the first dimensionality increase is performed through transposed convolution to obtain full-scale upsampling features. At the same time, the full-scale jump features of the corresponding level of the encoder after being enhanced by the slippery sensitive channel attention module are extracted. The cloud interference feature masking process is applied to the full-scale jump feature to obtain the processed full-scale jump feature. The processed full-scale jump feature and the full-scale upsampling feature are channel-level stitched and fused to obtain the full-scale fused feature, wherein the full-scale fused feature has the same resolution as the multi-temporal optical satellite image data. The full-scale fused features are input into the detail enhancement sub-network, and the features are enhanced through residual convolutional blocks to obtain the detail enhancement feature map; The pixel values ​​of the enhanced feature map are normalized and the spectrum is restored to match the spectral features and color distribution of the original satellite image to generate a cloudless landslide disaster satellite image.

5. A device for generating satellite image data of landslide disasters based on cloud cover, characterized in that, The apparatus is applied to the method for generating satellite image data of landslide disasters based on cloud cover as described in any one of claims 1-4, and the apparatus comprises: The acquisition module is used to acquire multi-temporal optical satellite image data, multi-source remote sensing data, meteorological data and topographic data of the target area, and calculate the multi-dimensional landslide physical feature vector corresponding to each optical satellite image data in combination with the topographic data. The optical satellite image data is landslide image data obscured by clouds, and the multi-source remote sensing data includes radar data, hyperspectral data and UAV low-altitude remote sensing data. The data enhancement module is used to fuse the multi-temporal optical satellite imagery data and the multi-source remote sensing data to perform fine classification of clouds and obtain cloud classification results; The fusion module is used to determine the priority processing area based on the cloud classification results, the meteorological data, and the terrain data; The result module is used to input the multi-temporal optical satellite image data and multi-dimensional landslide physical feature vector of the priority processing area into the landslide feature enhancement generation network to obtain landslide disaster satellite image data. The landslide feature enhancement generation network includes an encoder, a landslide sensitive channel attention module, a temporal collaborative attention module, a decoder, and a skip connection module.

6. A device for generating satellite image data of landslide disasters based on cloud cover, characterized in that, The device includes: a memory, a processor, and a cloud-based landslide disaster satellite image data generation program stored in the memory and running on the processor, the cloud-based landslide disaster satellite image data generation program being configured to implement the steps of the cloud-based landslide disaster satellite image data generation method as described in any one of claims 1-4.

7. A storage medium, characterized in that, The storage medium stores a satellite image data generation program for landslide disasters based on cloud cover. When the cloud cover-based satellite image data generation program is executed by the processor, it implements the steps of the cloud cover-based satellite image data generation method for landslide disasters as described in any one of claims 1-4.

Images