Satellite landslide data generation method, device, equipment, and medium based on multi-scale adversarial neural network.

Abstract

This application discloses a method, apparatus, device, and medium for generating satellite landslide data based on a multi-scale adversarial neural network, relating to the field of satellite landslide disaster technology. The method includes: fusing optical remote sensing images, DEM and auxiliary disaster perception data such as InSAR deformation and thermal infrared data to construct an adaptive attention fusion mechanism for topographic gradient field, curvature field, disaster intensity field, and thermal anomaly field, generating multi-source disaster perception features; using a multi-scale generator combined with DEM multi-scale features to generate images step-by-step, and using a multi-task discriminator to jointly judge optical authenticity and topographic plausibility, ultimately obtaining target satellite landslide data. Through precise alignment of multi-source disaster perception data and optical features, satellite landslide data with both topographic structural plausibility and optical texture authenticity is generated, improving the accuracy and reliability of the generated complex terrain satellite landslide data.

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 landslide data based on a multi-scale adversarial neural network. Background Technology

[0002] To address the scarcity of remote sensing data, current optical remote sensing data generation technologies primarily employ generative adversarial networks (GANs)-based methods. These methods typically use random noise or simple conditional vectors as input, generating optical remote sensing images through a deep convolutional GAN ​​architecture. Specifically, existing approaches often focus on generating data for a single optical modality, utilizing only the statistical features of optical remote sensing images for generation and modeling, lacking deep integration with terrain data. Regarding discriminator design, existing techniques mainly assess the optical realism of the generated images, optimizing network parameters by calculating the differences in optical distribution between generated and real samples, without fully considering the consistency constraints between the generated images and terrain conditions.

[0003] Therefore, current technologies suffer from the following shortcomings: First, they lack comprehensive utilization of multi-source disaster perception data, relying solely on a single optical modality. Second, the generator structure lacks multi-scale terrain constraints; existing methods fail to adaptively adjust generation strategies based on different regional characteristics of complex terrain. There is a lack of clear division of labor and coordination at levels such as large-scale terrain skeleton modeling, mid-scale cross-modal semantic alignment, and small-scale optical detail refinement, leading to problems such as disjointed terrain structures and blurred landslide boundaries in generated images. Furthermore, existing discriminators only assess optical realism, failing to introduce multi-source constraints based on DEM data. This makes it impossible to jointly evaluate the rationality of the generated image's terrain structure and the realism of its optical texture, hindering the consistency and reliability of the generated data in complex terrain scenarios.

[0004] Therefore, how to construct an adaptive attention fusion mechanism using multi-source disaster perception data and generate satellite landslide data through a generative adversarial network with multi-scale terrain constraints 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 landslide data based on a multi-scale adversarial neural network, aiming to solve the technical problem of how to improve the accuracy of the generated satellite landslide data.

[0006] To achieve the above objectives, this application proposes a satellite landslide data generation method based on a multi-scale adversarial neural network, comprising:

[0007] Acquire optical remote sensing images, digital elevation model data, and auxiliary disaster perception data for the same geographical area, wherein the auxiliary disaster perception data includes InSAR deformation data and thermal infrared image data;

[0008] The terrain gradient field and terrain curvature field are calculated based on the digital elevation model data, and the disaster intensity field and thermal anomaly field are calculated based on the auxiliary disaster perception data.

[0009] The terrain gradient field, the terrain curvature field, the catastrophic intensity field, and the thermal anomaly field are used to calculate the weight coefficients of each modal data through an adaptive attention mechanism to generate multi-source disaster perception features;

[0010] The multi-source disaster perception features and the optical remote sensing images are stitched together across modalities to form a joint input feature that includes terrain and disaster information;

[0011] The joint input features are input into a multi-scale generator, and combined with the multi-scale features of the digital elevation model data, a three-channel optical remote sensing image is generated step by step.

[0012] The three-channel optical remote sensing images are stitched together with the corresponding batch of digital elevation model data to form generated samples, and the optical remote sensing images of the real landslide area are stitched together with the corresponding batch of real digital elevation model data as real samples.

[0013] The generated samples and the real samples are input into the discriminator for adversarial discrimination. The multi-scale generator and the multi-task discriminator are trained alternately and iteratively until the number of iterations reaches the preset maximum number of iterations. The current multi-scale generator is then output as the target multi-scale generator.

[0014] The optical remote sensing image of the area to be enhanced, the digital elevation model data, and the auxiliary disaster perception data are input into the target multi-scale generator to obtain the target satellite landslide data.

[0015] In one embodiment, the steps of calculating the terrain gradient field and terrain curvature field based on the digital elevation model data, and calculating the disaster intensity field and thermal anomaly field based on the auxiliary disaster perception data, include:

[0016] The digital elevation model data is subjected to Gaussian filtering for noise reduction to obtain the noise-reduced digital elevation model data.

[0017] Based on the denoised digital elevation model data, the finite difference method is used to calculate the terrain slope components in the x and y directions, and the terrain gradient field is obtained through vector synthesis.

[0018] The profile curvature and plane curvature of the denoised digital elevation model data are solved by second-order partial derivatives, and then weighted and fused according to preset weights to form a terrain curvature field, wherein the preset weights are adaptively adjusted according to the complexity of the regional terrain.

[0019] Spatiotemporal baseline correction is performed on the InSAR deformation data of the auxiliary disaster perception data to extract the deformation rate of the potential landslide area, and the disaster intensity field is obtained by normalization based on the deformation rate.

[0020] The thermal infrared image data of the auxiliary disaster perception data is subjected to atmospheric correction and radiometric calibration to convert it into real temperature data;

[0021] Based on the real temperature data, thermal anomaly regions are extracted using an adaptive threshold segmentation algorithm to generate a thermal anomaly field.

[0022] In one embodiment, the step of calculating the weight coefficients of each modal data—the terrain gradient field, the terrain curvature field, the catastrophic intensity field, and the thermal anomaly field—using an adaptive attention mechanism to generate multi-source disaster perception features includes:

[0023] Construct a dual-branch attention network, which includes a channel attention branch and a spatial attention branch;

[0024] The terrain gradient field, the terrain curvature field, the catastrophic intensity field, and the thermal anomaly field are input into the channel attention branch. The channel statistical features of each modality data are extracted through global mean pooling and global maximum pooling. The channel weight coefficients of each modality data are output through a fully connected layer and a Softmax activation function.

[0025] After multiplying each modal data by its corresponding channel weight coefficient, we obtain the weighted modal data.

[0026] The weighted modal data are input into the spatial attention branch, and a spatial attention map is obtained through cross-modal feature interaction calculation. The modal data are the terrain gradient field, curvature field, catastrophic intensity field and thermal anomaly field. The spatial attention map is used to characterize the importance of catastrophic features at different spatial locations.

[0027] The spatial attention map and the weighted modal data are multiplied pixel by pixel to obtain the double-weighted modal features;

[0028] Modality correlation constraint loss is introduced to iteratively optimize the dual-weighted modal features to obtain the optimized modal features.

[0029] The optimized modal features are concatenated along the channel dimension, and feature dimension fusion and dimensionality reduction are performed through a 1×1 convolutional layer to generate multi-source disaster perception features.

[0030] In one embodiment, the step of cross-modal stitching of the multi-source disaster perception features and the optical remote sensing image to form a joint input feature containing terrain and disaster information includes:

[0031] The multi-source disaster perception features are subjected to feature enhancement processing to extract deep disaster correlation features;

[0032] Multi-scale feature extraction is performed on the optical remote sensing image to obtain optical texture features at different levels;

[0033] Based on the deep disaster association features and optical texture features at each level, a semantic similarity matrix is ​​obtained.

[0034] Based on the semantic similarity matrix, cross-modal fusion is performed using a two-dimensional fusion method to obtain fused features, wherein the two-dimensional fusion method is channel dimension splicing and spatial dimension superposition.

[0035] By introducing modality consistency constraints and optimizing the fused features through a comparative loss function, the optimized fused features are obtained.

[0036] Based on the optimized fusion features and unified feature distribution range, a joint input feature containing terrain and disaster information is formed.

[0037] In one embodiment, the step of inputting the joint input features into a multi-scale generator and combining them with the multi-scale features of the digital elevation model data to generate a three-channel optical remote sensing image step by step includes:

[0038] Construct a multi-scale generator, wherein the multi-scale generator includes a large-scale terrain constraint branch, a medium-scale cross-modal fusion branch, and a small-scale detail completion branch;

[0039] Multi-scale pyramid decomposition is performed on the digital elevation model data to obtain large-scale terrain skeleton features, meso-scale terrain transition features, and small-scale terrain detail features in sequence.

[0040] The joint input features are input into the large-scale terrain constraint branch, and gated fusion is performed with the large-scale terrain skeleton features. The global spatial structure is modeled through the terrain-guided generation strategy, and a large-scale base image is output.

[0041] The large-scale base image is upsampled and input into the mesoscale cross-modal fusion branch. The optical texture and terrain features are aligned through a cross-modal attention mechanism by combining the mesoscale terrain transition features with the deep semantic information in the joint input features to obtain a mesoscale refined image.

[0042] The mesoscale refined image is upsampled and then input into the small-scale detail completion branch. The small-scale terrain detail features are fused with the high-frequency texture information of the optical remote sensing image. Key details are enhanced through an adaptive edge enhancement mechanism to obtain the full-scale initial image.

[0043] The initial full-scale image is subjected to channel filtering and feature calibration, and mapped to a three-channel feature map through a 1×1 convolutional layer. After normalization, a three-channel optical remote sensing image is obtained.

[0044] In one embodiment, the step of inputting the generated samples and the real samples into a discriminator for adversarial discrimination, alternately iteratively training the multi-scale generator and the multi-task discriminator until the number of iterations reaches a preset maximum number of iterations, and outputting the current multi-scale generator as the target multi-scale generator includes:

[0045] A multi-task discriminator is constructed, wherein the multi-task discriminator includes a feature extraction network, a terrain structure discrimination branch, a disaster feature discrimination branch, and a global authenticity discrimination branch. The terrain structure discrimination branch loads prior parameters of digital elevation model features, and the disaster feature discrimination branch loads a disaster intensity and thermal anomaly feature template library.

[0046] The generated samples and the real samples are input into the low-level feature extraction network of the multi-task discriminator to obtain a shared feature map;

[0047] The shared feature map is input into the terrain structure discrimination branch, and the terrain rationality score is obtained by matching and calculating with the prior parameters of the digital elevation model features.

[0048] The shared feature map is input into the disaster feature discrimination branch, and the disaster typicality score is obtained by comparing it with the disaster intensity and thermal anomaly feature template library.

[0049] The shared feature map is input into the global authenticity discrimination branch, and the overall authenticity score is obtained by mapping through a fully connected layer and activating with Sigmoid.

[0050] A joint loss function is constructed based on the terrain rationality score, disaster typicality score, and overall sample authenticity score. The weight coefficients of the joint loss function include global authenticity weight coefficients, terrain structure weight coefficients, and disaster feature weight coefficients. The global authenticity weight coefficients, terrain structure weight coefficients, and disaster feature weight coefficients are adaptively adjusted according to the current training iteration round.

[0051] The parameters of the multi-scale generator are fixed, and the calculation results of the joint loss function are used to update the parameters of the multi-task discriminator according to a preset learning rate through the backpropagation algorithm to obtain the optimized discriminator. The parameters of the multi-task discriminator include the feature extraction network parameters, terrain structure discrimination branch parameters, disaster feature discrimination branch parameters, and global authenticity discrimination branch parameters.

[0052] By fixing the parameters of the optimized discriminator, the generated sample is input into the optimized discriminator to obtain an updated terrain rationality score, an updated disaster typicality score, and an updated overall authenticity score;

[0053] Based on the updated terrain rationality score, updated disaster typicality score, and updated overall realism score, combined with the joint loss function, the generator loss value is calculated, and the parameters of the multi-scale generator are updated through the gradient descent algorithm to obtain an optimized generator.

[0054] Repeatedly execute the alternating iterative process of discriminator training and generator training. When the number of iterations reaches the preset maximum number of iterations, output the current optimized generator as the target multi-scale generator.

[0055] In one embodiment, the step of inputting the optical remote sensing image of the area to be enhanced, digital elevation model data, and auxiliary disaster perception data into the target multi-scale generator to obtain target satellite landslide data includes:

[0056] A multi-dimensional prior model library is constructed based on historical real satellite landslide data. The multi-dimensional prior model library includes the data distribution patterns of different terrain complexities and the data distribution patterns corresponding to different disaster development stages.

[0057] The multi-dimensional prior model library is fused with the target multi-scale generator to obtain an optimized generator;

[0058] The optical remote sensing images, digital elevation model data, and auxiliary disaster perception data of the area to be enhanced are preprocessed to obtain standardized data.

[0059] The standardized data is input into the optimization generator to obtain initial satellite landslide data, wherein the initial satellite landslide data includes terrain structure information, disaster feature information and optical texture information;

[0060] The initial satellite landslide data is compared dimension by dimension with the source data in the real satellite landslide data sample library to obtain the terrain structure deviation value, disaster characteristic deviation value and optical texture deviation value.

[0061] Based on the terrain structure deviation value, disaster feature deviation value, and optical texture deviation value, the initial satellite landslide data is optimized using a gradient descent algorithm to obtain optimized satellite landslide data.

[0062] Based on the optimized satellite landslide data, a multi-dimensional quality assessment was conducted to obtain data reliability, feature consistency, and scenario adaptability.

[0063] When the data credibility, feature consistency, and scene adaptability all meet the corresponding preset thresholds, the optimized satellite landslide data will be used as the target satellite landslide data.

[0064] Furthermore, to achieve the above objectives, this application also proposes a satellite landslide data generation device based on a multi-scale adversarial neural network, the satellite landslide data generation device based on the multi-scale adversarial neural network comprising:

[0065] The acquisition module is used to acquire optical remote sensing images, digital elevation model data and auxiliary disaster perception data of the same geographical area, wherein the auxiliary disaster perception data includes InSAR deformation data and thermal infrared image data.

[0066] The calculation module is used to calculate the terrain gradient field and terrain curvature field based on the digital elevation model data, and to calculate the disaster intensity field and thermal anomaly field based on the auxiliary disaster perception data.

[0067] The multimodal fusion module is used to calculate the weight coefficients of each modal data by means of the terrain gradient field, the terrain curvature field, the disaster intensity field and the thermal anomaly field through an adaptive attention mechanism, so as to generate multi-source disaster perception features;

[0068] The cross-modal stitching module is used to stitch the multi-source disaster perception features and the optical remote sensing image across modalities to form a joint input feature that includes terrain and disaster information;

[0069] The image generation module is used to input the joint input features into the multi-scale generator, and combine the multi-scale features of the digital elevation model data to generate a three-channel optical remote sensing image step by step.

[0070] The sample construction module is used to stitch the three-channel optical remote sensing images with the corresponding batch of digital elevation model data to form generated samples, and to stitch the optical remote sensing images of the real landslide area with the corresponding batch of real digital elevation model data as real samples.

[0071] The adversarial training module is used to input the generated samples and the real samples into the discriminator for adversarial discrimination, and to alternately iterate the training of the multi-scale generator and the multi-task discriminator until the number of iterations reaches the preset maximum number of iterations, and output the current multi-scale generator as the target multi-scale generator.

[0072] The results module is used to input the optical remote sensing image of the area to be enhanced, digital elevation model data, and auxiliary disaster perception data into the target multi-scale generator to obtain target satellite landslide data.

[0073] 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 satellite landslide data generation method based on multi-scale adversarial neural networks as described above.

[0074] 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 satellite landslide data generation method based on a multi-scale adversarial neural network as described above.

[0075] This application integrates optical remote sensing images, DEM and InSAR deformation, thermal infrared and other auxiliary disaster perception data to construct an adaptive attention fusion mechanism for topographic gradient field, curvature field, disaster intensity field and thermal anomaly field, generating multi-source disaster perception features. A multi-scale generator is used to combine DEM multi-scale features to generate images step by step, and a multi-task discriminator is used to jointly judge optical realism and topographic plausibility, ultimately obtaining target satellite landslide data. Through precise alignment of multi-source disaster perception data and optical features, satellite landslide data with both topographic structural plausibility and optical texture realism is generated, improving the accuracy and reliability of the generated complex terrain satellite landslide data. Attached Figure Description

[0076] 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.

[0077] Figure 1 This is a flowchart illustrating the first embodiment of the satellite landslide data generation method based on a multi-scale adversarial neural network of this application.

[0078] Figure 2 This is a flowchart illustrating the second embodiment of the satellite landslide data generation method based on a multi-scale adversarial neural network of this application.

[0079] Figure 3 This is a schematic diagram of the module structure of the satellite landslide data generation device based on a multi-scale adversarial neural network, according to the first embodiment of the satellite landslide data generation method based on a multi-scale adversarial neural network of this application.

[0080] Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the satellite landslide data generation method based on multi-scale adversarial neural networks in the embodiments of this application.

[0081] 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

[0082] 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.

[0083] 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.

[0084] To address the scarcity of remote sensing data, current optical remote sensing data generation technologies primarily employ generative adversarial networks (GANs)-based methods. These methods typically use random noise or simple conditional vectors as input, generating optical remote sensing images through a deep convolutional GAN ​​architecture. Specifically, existing approaches often focus on generating data for a single optical modality, utilizing only the statistical features of optical remote sensing images for generation and modeling, lacking deep integration with terrain data. Regarding discriminator design, existing techniques mainly assess the optical realism of the generated images, optimizing network parameters by calculating the differences in optical distribution between generated and real samples, without fully considering the consistency constraints between the generated images and terrain conditions.

[0085] Therefore, current technologies suffer from the following shortcomings: First, they lack comprehensive utilization of multi-source disaster perception data, relying solely on a single optical modality. Second, the generator structure lacks multi-scale terrain constraints; existing methods fail to adaptively adjust generation strategies based on different regional characteristics of complex terrain. There is a lack of clear division of labor and coordination at levels such as large-scale terrain skeleton modeling, mid-scale cross-modal semantic alignment, and small-scale optical detail refinement, leading to problems such as disjointed terrain structures and blurred landslide boundaries in generated images. Furthermore, existing discriminators only assess optical realism, failing to introduce multi-source constraints based on DEM data. This makes it impossible to jointly evaluate the rationality of the generated image's terrain structure and the realism of its optical texture, hindering the consistency and reliability of the generated data in complex terrain scenarios.

[0086] Based on the above, this application also provides a satellite landslide data generation method based on a multi-scale adversarial neural network, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the satellite landslide data generation method based on multi-scale adversarial neural networks of this application.

[0087] In this embodiment, the satellite landslide data generation method based on multi-scale adversarial neural networks includes steps S10~S80:

[0088] Step S10: Acquire optical remote sensing images, digital elevation model data, and auxiliary disaster perception data for the same geographical area.

[0089] It should be noted that the same geographical region refers to a specific geographical space with a completely consistent spatial extent. The delineation of the boundaries of this region must ensure that all subsequently acquired data can accurately correspond to each other, ensuring that different data are comparable in spatial location. Optical remote sensing images are images generated by capturing visible light, near-infrared, and other band signals reflected from the surface of a geographical region using optical remote sensing equipment. They can clearly present intuitive information such as the optical texture of the surface, vegetation cover, water distribution, and the macroscopic outline of landslide areas. Digital elevation model (DEM) data is a set of data that reflects the elevation undulations of a geographical region based on discrete three-dimensional spatial coordinate points and constructed through interpolation calculations. It is the core data describing terrain slope, elevation, and topographic structure characteristics. Auxiliary disaster perception data is supplementary data used to enhance the identification of landslide disaster characteristics. In conjunction with optical remote sensing images and DEM data, it can more comprehensively capture the relevant characteristics of landslide disasters and improve the accuracy of disaster perception. Auxiliary disaster perception data includes InSAR deformation data and thermal infrared image data. InSAR deformation data is data on minute surface deformations in a geographic area obtained using synthetic aperture radar interferometry. It can accurately monitor the deformation rate and extent of the surface and effectively capture potential deformation characteristics in landslide areas. Thermal infrared image data is data generated by capturing surface thermal radiation signals in a geographic area using thermal infrared remote sensing equipment. It can reflect temperature differences in different areas of the surface. Landslide areas exhibit unique thermal anomaly characteristics due to factors such as soil and rock loosening and changes in groundwater.

[0090] Specifically, the geographical area to be processed is determined and optical remote sensing images of the area are acquired. These images typically contain surface reflectance information in the red, green, and blue bands. Simultaneously, digital elevation model (DEM) data for the corresponding area is acquired, which records surface elevation values ​​in a discrete grid format. Additionally, auxiliary disaster perception data needs to be acquired, including InSAR deformation data, which uses synthetic aperture radar interferometry to obtain information on minute surface deformations, and thermal infrared imagery data, which reflects surface temperature distribution characteristics. All of the above data must have a unified geographic coordinate reference system and spatial resolution to ensure spatial alignment accuracy in subsequent processing.

[0091] Step S20: Calculate the terrain gradient field and terrain curvature field based on the digital elevation model data, and calculate the disaster intensity field and thermal anomaly field based on the auxiliary disaster perception data.

[0092] It should be noted that the topographic gradient field, calculated based on digital elevation model data, reflects the rate and direction of change in surface elevation in a geographical area, quantifying the magnitude and direction of surface slope. The topographic curvature field, also calculated based on digital elevation model data, reflects the degree of surface curvature in a geographical area, accurately depicting the surface's undulations and further distinguishing different terrain types such as steep slopes, gentle slopes, valleys, and ridges, aiding in the identification of landslide-prone areas. The catastrophic intensity field, calculated based on InSAR deformation data from the disaster sensing data, quantifies the intensity of landslide development, reflecting the severity of surface deformation and visually demonstrating the activity level and impact range of landslides. The thermal anomaly field, calculated based on thermal infrared imagery data from the disaster sensing data, reflects the abnormal distribution of surface temperature in a geographical area, capturing temperature differences in landslide areas caused by factors such as soil and rock loosening and groundwater infiltration.

[0093] Further, step S20 includes: performing Gaussian filtering denoising on the digital elevation model data to obtain denoised digital elevation model data; calculating the x-direction and y-direction terrain slope components using the finite difference method based on the denoised digital elevation model data, and obtaining the terrain gradient field through vector synthesis; solving the profile curvature and planar curvature of the denoised digital elevation model data using second-order partial derivatives, and performing weighted fusion according to preset weights to form a terrain curvature field, wherein the preset weights are adaptively adjusted according to the complexity of the regional terrain; performing spatiotemporal baseline correction on the InSAR deformation data of the auxiliary disaster perception data, extracting the deformation rate of potential landslide areas, and obtaining the disaster intensity field based on the deformation rate through normalization; performing atmospheric correction and radiometric calibration on the thermal infrared image data of the auxiliary disaster perception data, converting it into real temperature data; and extracting thermal anomaly areas based on the real temperature data using an adaptive threshold segmentation algorithm to generate a thermal anomaly field.

[0094] It should be noted that Gaussian filtering denoising is a processing method based on using a Gaussian function to construct a filter template and performing convolution operations on digital elevation model data. The finite difference method is a numerical calculation method that discretizes continuous digital elevation model data and approximates the derivative of the continuous function by using the difference of function values ​​at discrete points. The terrain slope component decomposes the terrain gradient into two orthogonal horizontal components: the x-axis terrain slope component and the y-axis terrain slope component. These represent the rate of elevation change of the land surface along the horizontal x-axis and y-axis, respectively, and are the fundamental data for vector synthesis of the terrain gradient field. Vector synthesis is the process of using the x-axis and y-axis terrain slope components as two components of a vector and calculating the magnitude and direction of the resultant vector using vector operations. The characteristics of the resultant vector are the core features of the terrain gradient field, fully representing the actual slope and aspect of the land surface. The second-order partial derivative is the result of solving the elevation values ​​of the digital elevation model data twice consecutively. The first partial derivative reflects the rate of elevation change, and the second-order partial derivative further reflects the trend of the rate of elevation change. Profile curvature is a curvature feature calculated using the second-order partial derivative along the slope direction. It characterizes the degree of curvature of the terrain in the slope direction. Concave slope profile curvature is positive, convex slope profile curvature is negative, and straight slope profile curvature is zero. Planar curvature is a curvature feature calculated using the second-order partial derivative along the tangent direction of the terrain contour lines. It characterizes the degree of curvature of the terrain in the horizontal plane perpendicular to the slope direction, reflecting the planar morphological features of terrain such as valleys and ridges. Preset weights are weighting coefficients pre-set for fusing profile curvature and planar curvature. These coefficients are not fixed values ​​but are dynamically adjusted according to the complexity of the regional terrain to make the terrain curvature field more closely match the actual curvature characteristics of different terrains. Regional terrain complexity is a comprehensive characterization of terrain complexity based on indicators such as terrain undulation and slope variability calculated from digital elevation model data. Regions with intersecting valleys and large slope variations have high terrain complexity, while flat regions have low terrain complexity. Spatiotemporal baseline correction is a correction process for systematic errors caused by differences in the spatiotemporal baseline of radar imaging in InSAR deformation data. Landslide potential areas are geographical regions preliminarily identified as prone to landslides based on topographic features, deformation trends, and other indicators. The deformation rate in these areas typically differs significantly from that of surrounding areas, making them key areas for calculating the catastrophic intensity field. Deformation rate, the amount of surface deformation per unit time, is a core indicator characterizing the severity of surface deformation. The deformation rate in landslide potential areas is much higher than in stable areas, serving as a crucial basis for quantifying the catastrophic intensity field. Radiometric calibration is the process of converting the digitally quantified values ​​of thermal infrared image data into physically meaningful radiance values. It is a crucial preliminary step in converting image data into true temperature data, establishing a quantitative relationship between image pixel values ​​and surface radiance. True temperature data, obtained through atmospheric correction and radiometric calibration, is numerical data that accurately reflects the actual surface temperature of a geographical region.The adaptive threshold segmentation algorithm is an image segmentation algorithm that dynamically calculates the segmentation threshold based on the gray value distribution characteristics of thermal infrared image data. It can adaptively adjust the threshold according to the temperature distribution of different regions and distinguish between thermal anomaly regions and normal temperature regions.

[0095] Specifically, Gaussian filtering is applied to the digital elevation model (DEM) data for noise reduction. A Gaussian kernel function is constructed and convolved with the original DEM to effectively suppress random noise and measurement errors in the elevation data, resulting in denoised DEM data. This is done to eliminate the impact of noise interference on subsequent terrain differential calculations, ensuring the accuracy of gradient and curvature features. Secondly, based on the denoised DEM data, the finite difference method is used to calculate the x-direction and y-direction terrain slope components. Specifically, the elevation change rate of adjacent grid points is solved using the central difference formula, and the terrain gradient field is obtained through vector synthesis, i.e., by performing the square root operation on the sum of squares of the two directional components, thereby quantifying the surface tilt and aspect information. Then, the profile curvature and planar curvature of the denoised digital elevation model data are solved by the second-order partial derivatives. The second derivative of the elevation change rate is calculated along the direction of maximum slope for the profile curvature, and the second derivative of the elevation change rate is calculated along the contour line direction for the planar curvature. The data are then weighted and fused according to preset weights to form a terrain curvature field. The preset weights are adaptively adjusted according to the complexity of the regional terrain. Specifically, the profile curvature weight is increased in areas with severe terrain undulations to highlight longitudinal terrain changes, and the planar curvature weight is increased in flat areas to capture lateral terrain features, thereby comprehensively representing the surface undulations and water flow convergence characteristics. Subsequently, spatiotemporal baseline correction was performed on the InSAR deformation data of the auxiliary disaster perception data. Interferometric pairs were screened by selecting appropriate spatiotemporal baseline thresholds, and an external digital elevation model was used to remove topographic phase. Phase unwrapping and atmospheric delay correction were then performed to extract the deformation rate of potential landslide areas. Based on this deformation rate, normalization was applied to map deformation data of different magnitudes to a unified numerical range, resulting in a disaster intensity field. This was done to quantify the relative severity of surface deformation, facilitating subsequent fusion and comparison with other modal data. Finally, atmospheric correction and radiometric calibration were performed on the thermal infrared image data of the auxiliary disaster perception data. An atmospheric radiative transfer model was used to remove the attenuation effects of atmospheric water vapor and aerosols on thermal radiation, and the digital quantized values ​​recorded by the sensors were converted into physically meaningful real temperature data. Based on this real temperature data, an adaptive threshold segmentation algorithm was used to extract thermal anomaly areas. Specifically, the segmentation threshold was dynamically determined using local statistical characteristics to identify anomaly areas with temperatures significantly higher than the background temperature and generate a thermal anomaly field. This captures temperature anomaly signals caused by internal frictional heating of the landslide body or the thermal response of groundwater, providing a basis for the latent characteristic perception of landslide disasters.

[0096] Step S30: The terrain gradient field, terrain curvature field, disaster intensity field and thermal anomaly field are calculated using an adaptive attention mechanism to generate multi-source disaster perception features by calculating the weight coefficients of each modality data.

[0097] It should be noted that the adaptive attention mechanism is a feature fusion mechanism that can adaptively allocate weights based on the feature importance of multimodal data. It does not require manual pre-setting of fixed weights; instead, it learns autonomously from the feature relevance and significance of each modality, dynamically adjusting the weight ratio of different modalities to achieve accurate fusion of multimodal features. The data modalities are topographic gradient field, curvature field, catastrophic intensity field, and thermal anomaly field. Each data type corresponds to a feature dimension characterizing landslide disasters. These four data types are independent yet interconnected, collectively constituting a multimodal feature set for landslide disasters. The weight coefficient is a numerical coefficient used to measure the importance of each modality. The magnitude of the weight coefficient is positively correlated with the contribution of the corresponding modality to the characterization of landslide disasters; the higher the contribution, the larger the weight coefficient. Weight allocation can highlight key features and suppress interference from irrelevant features. Multi-source disaster perception features are comprehensive features obtained by fusing four types of modal data: topographic gradient field, topographic curvature field, disaster intensity field, and thermal anomaly field. These features integrate four core types of information: topographic undulation, topographic curvature, disaster intensity, and thermal anomaly. They can comprehensively and accurately characterize the overall features of landslide disasters and provide core input for subsequent model training.

[0098] Further, step S30 includes: constructing a dual-branch attention network, which includes a channel attention branch and a spatial attention branch; inputting the terrain gradient field, terrain curvature field, catastrophic intensity field, and thermal anomaly field into the channel attention branch, extracting the channel statistical features of each modality data through global mean pooling and global maximum pooling, and outputting the channel weight coefficients of each modality data through a fully connected layer and a Softmax activation function; multiplying each modality data with the corresponding channel weight coefficient to obtain weighted modality data; inputting the weighted modality data into the spatial attention branch, obtaining a spatial attention map through cross-modal feature interaction calculation, which is used to characterize the importance of catastrophic features at different spatial locations; multiplying the spatial attention map and the weighted modality data pixel by pixel to obtain double-weighted modality features; introducing modality correlation constraint loss to iteratively optimize the double-weighted modality features to obtain optimized modality features; concatenating the optimized modality features along the channel dimension, and performing feature dimension fusion and dimensionality reduction through a 1×1 convolutional layer to generate multi-source catastrophic perception features.

[0099] It should be noted that global mean pooling is a pooling operation that calculates the average value of all pixel values ​​within the entire channel of the feature map. It can extract the global average feature within the channel, reflecting the overall distribution trend of the channel's features. Global maximum pooling is a pooling operation that extracts the maximum value of all pixel values ​​within the entire channel of the feature map. It can capture local salient features within the channel, reflecting the extreme value distribution of the channel's features. Combined with global mean pooling, it can completely extract the channel's statistical features. Channel weight coefficients are numerical coefficients used to measure the importance of each modal data channel, highlighting channel features that are key to representing landslide disasters. The spatial attention map is a two-dimensional feature map that represents the importance of disaster features at different spatial locations, obtained through cross-modal feature interaction calculation. Each pixel value in the map corresponds to a weight coefficient for a spatial location; the larger the pixel value, the more important the disaster feature at that spatial location.

[0100] Specifically, a dual-branch attention network is constructed, comprising a channel attention branch and a spatial attention branch. Parallel design decouples the channel and spatial dimensions to capture differences in importance between modalities and differences in disaster sensitivity based on spatial location. Next, the topographic gradient field, topographic curvature field, catastrophic intensity field, and thermal anomaly field are input into the channel attention branch. Global mean pooling and global maximum pooling are performed on each modal feature map to extract channel statistical features. The pooling results are concatenated and input into a shared fully connected layer. After ReLU activation and dimensionality reduction, the channel weight coefficients for each modality are output through a Softmax activation function. This dynamically evaluates the contribution of different physical fields to landslide characterization and suppresses redundant modal interference. Then, each modal data is broadcast and multiplied with its corresponding channel weight coefficient to obtain weighted modal data, achieving adaptive enhancement of the channel dimension. Subsequently, the weighted modal data are input into the spatial attention branch. A spatial attention map is obtained through cross-modal feature interaction. Specifically, the modal features are first concatenated along the channel dimension to form a comprehensive feature map. Then, the average and maximum values ​​are calculated along the channel dimension to obtain two spatial sub-maps. These sub-maps are concatenated and then processed through a 7×7 convolutional layer and a sigmoid activation function to generate the spatial attention map. This spatial attention map is used to characterize the importance of disaster features at different spatial locations, highlighting key landslide areas and suppressing background noise. Next, the spatial attention map and the weighted modal data are multiplied pixel-by-pixel to obtain double-weighted modal features, realizing a collaborative attention mechanism between the channel and spatial dimensions. Then, a modal correlation constraint loss is introduced to iteratively optimize the double-weighted modal features. This loss calculates the mutual information or cosine similarity between modal features, constraining the weight allocation of strongly correlated modalities to avoid information redundancy, while encouraging the complementary fusion of weakly correlated modalities. Optimized modal features are obtained through multiple iterations. Finally, the optimized modal features are stitched together along the channel dimension, and the feature dimensions are fused and reduced through a 1×1 convolutional layer. The high-dimensional stitched features are mapped to a specified number of channels to generate multi-source disaster perception features. This is done to reduce the feature dimension while retaining key disaster information, so as to facilitate efficient cross-modal fusion with optical remote sensing images in the future.

[0101] Step S40: Multi-source disaster perception features and optical remote sensing images are stitched together across modes to form joint input features that include terrain and disaster information.

[0102] It should be noted that topographic and disaster information refers to all the core information contained in the multi-source disaster perception features. Specifically, it includes the topographic undulation and curvature features represented by the topographic gradient field and topographic curvature field, as well as the disaster development intensity and thermal anomaly distribution features represented by the disaster intensity field and thermal anomaly field. It is the key information for characterizing the topographic basis and disaster features of landslide areas.

[0103] Further, step S40 includes: performing feature enhancement processing on the multi-source disaster perception features to extract deep disaster association features; performing multi-scale feature extraction on the optical remote sensing image to obtain optical texture features at different levels; calculating a semantic similarity matrix based on the deep disaster association features and optical texture features at each level; performing cross-modal fusion based on the semantic similarity matrix using a two-dimensional fusion method to obtain fused features, wherein the two-dimensional fusion method is channel dimension splicing and spatial dimension superposition; introducing modal consistency constraints to optimize the fused features through a contrastive loss function to obtain optimized fused features; and unifying the feature distribution range based on the optimized fused features to form a joint input feature containing terrain and disaster information.

[0104] It should be noted that feature enhancement processing is a deep feature mining and enhancement operation performed on multi-source disaster perception features. It extracts the internal correlation information of features through a specific network structure, amplifies the representational strength of key disaster features, and suppresses the interference of redundant information. Deep disaster correlation features are deep features extracted from multi-source disaster perception features through feature enhancement processing, containing the inherent correlations of various disaster features, rather than a superposition of single surface disaster features. Multi-scale feature extraction is an operation performed on optical remote sensing images, extracting features from different spatial resolution levels. It can capture complete optical information from global macroscopic contours to local fine textures, forming multi-level feature sets to meet the feature requirements of image generation at different scales. Optical texture features are features contained in optical remote sensing images that characterize the surface texture morphology, reflecting intuitive texture information such as surface vegetation cover, soil and rock distribution, and landslide boundaries. The semantic similarity matrix is ​​matrix data obtained by calculating the degree of semantic correlation between deep disaster correlation features and optical texture features at each level. Each value in the matrix corresponds to the semantic similarity between a pair of features; the larger the value, the higher the semantic correlation between the two types of features, providing a quantitative matching basis for cross-modal fusion. The dual-dimensional fusion approach refers to a strategy that simultaneously employs channel-dimensional splicing and spatial-dimensional overlay for cross-modal fusion. These two methods complement each other, achieving comprehensive feature integration across both channel and spatial dimensions. Unifying the feature distribution range involves numerical normalization of the optimized fused features, mapping feature values ​​to a fixed and reasonable range, eliminating dimensional and numerical range differences between different features, and ensuring a uniform feature distribution. The optimized fused features are obtained through modality consistency constraints and a contrastive loss function, resolving issues such as inconsistent modality data distribution and inaccurate semantic matching in fused features.

[0105] Specifically, feature enhancement processing is performed on multi-source disaster perception features. Deep disaster-related features are extracted by stacking multiple residual convolutional blocks. Each residual convolutional block contains two 3×3 convolutional layers and skip connections. This layer-by-layer abstraction captures high-order semantic associations and long-distance spatial dependencies of disaster features, overcoming the limitations of shallow feature representation and uncovering hidden geological disaster evolution patterns. Secondly, multi-scale feature extraction is performed on optical remote sensing images. Optical texture features at different levels are obtained through a convolutional pyramid network. Specifically, multiple parallel convolutional branches with different dilation rates or pooling scales are used to extract detailed textures, local structures, and global contextual information to meet the needs of multi-scale visual representation of landslide areas. Then, a semantic similarity matrix is ​​calculated based on the deep disaster-related features and optical texture features at each level. Specifically, the dot product or cosine similarity of the two types of feature vectors is calculated to measure the degree of matching between disaster semantics and visual features in the embedding space, providing a basis for subsequent adaptive fusion. Subsequently, cross-modal fusion is performed using a two-dimensional fusion approach based on the semantic similarity matrix to obtain fused features. This two-dimensional fusion approach involves concatenating channel-dimensional features and superimposing spatial-dimensional features. Specifically, in the channel dimension, disaster features and optical features are concatenated and concatenated to preserve complete information. Simultaneously, in the spatial dimension, weighted superposition based on the similarity matrix strengthens the response of semantically consistent regions, achieving synergy between information complementarity and feature enhancement. Next, modal consistency constraints are introduced to optimize the fused features through a contrastive loss function. Specifically, InfoNCE loss or triplet loss is used to narrow the distance between positive sample pairs and feature pairs, and to widen the distance between negative sample pairs, resulting in optimized fused features. This is done to constrain the consistency of the distribution of features from different modalities in the joint embedding space, avoiding fusion distortion caused by semantic gaps between modalities. Finally, based on the optimized fused features, the feature distribution range is unified. Instance normalization or layer normalization maps the features to a standard normal distribution, forming joint input features containing terrain and disaster information. This ensures that the subsequent multi-scale generator receives stable and standardized data input, improving the convergence and stability of the generation training.

[0106] Step S50: Input the joint input features into the multi-scale generator, and combine the multi-scale features of the digital elevation model data to generate a three-channel optical remote sensing image step by step.

[0107] It should be noted that the multi-scale generator is a network model with multi-scale feature processing and image generation capabilities. It can adapt to feature inputs of different resolutions and can complete image generation layer by layer from global to local, taking into account both the overall spatial structure and local detailed texture of the image. The multi-scale features of digital elevation model data are different levels of terrain features obtained after multi-scale decomposition of the digital elevation model data. They cover large-scale terrain skeleton features, medium-scale terrain transition features, and small-scale terrain detail features. Different scale features correspond to different levels of terrain undulation and structural information.

[0108] Further, step S50 includes: constructing a multi-scale generator, wherein the multi-scale generator includes a large-scale terrain constraint branch, a mesoscale cross-modal fusion branch, and a small-scale detail completion branch; performing multi-scale pyramid decomposition on the digital elevation model data to obtain large-scale terrain skeleton features, mesoscale terrain transition features, and small-scale terrain detail features in sequence; inputting the joint input features into the large-scale terrain constraint branch, performing gated fusion with the large-scale terrain skeleton features, and modeling the global spatial structure through a terrain-guided generation strategy to output a large-scale base image; upsampling the large-scale base image and inputting it into the mesoscale cross-modal fusion branch. The first branch combines mesoscale terrain transition features with deep semantic information from joint input features to align optical textures with terrain features via a cross-modal attention mechanism, resulting in a mesoscale refined image. The mesoscale refined image is then upsampled and input into the small-scale detail completion branch, which fuses small-scale terrain detail features with high-frequency texture information from the optical remote sensing image. An adaptive edge enhancement mechanism is used to strengthen key details, resulting in a full-scale initial image. The full-scale initial image undergoes channel filtering and feature calibration, and is mapped to a three-channel feature map via a 1×1 convolutional layer. After normalization, a three-channel optical remote sensing image is obtained.

[0109] It should be noted that the large-scale terrain constraint branch is the branch in the multi-scale generator responsible for constructing the global framework of the image, focusing on the overall undulation and macroscopic structure of the terrain. The mesoscale cross-modal fusion branch is the branch in the multi-scale generator responsible for optimizing the mid-level semantics and feature matching of the image, focusing on the synergistic fusion of terrain transitions and optical textures within the mesoscale range. The small-scale detail completion branch is the branch in the multi-scale generator responsible for supplementing the local subtle features of the image, focusing on terrain details and optical high-frequency textures within the small-scale range, making the generated data more consistent with the actual scene. Multi-scale pyramid decomposition is a decomposition operation that performs layer-by-layer downsampling and feature extraction on digital elevation model data, generating feature levels of different scales by continuously reducing the data resolution. Deep semantic information is the core semantic information contained in the joint input features, reflecting the intrinsic relationship between terrain and disaster, rather than a surface feature superposition, which can provide semantic support for mesoscale cross-modal fusion. High-frequency texture information is the information contained in optical remote sensing images that reflects subtle changes in surface texture, corresponding to details such as small-scale soil and rock distribution, vegetation texture, and landslide boundaries. Adaptive edge enhancement is a mechanism that can adaptively adjust the enhancement intensity according to the edge features of an image, focusing on key edge areas such as landslide boundaries and abrupt terrain changes to enhance the clarity and recognizability of edges.

[0110] Specifically, a multi-scale generator is constructed, comprising a large-scale terrain constraint branch, a mesoscale cross-modal fusion branch, and a small-scale detail completion branch. This three-level branch structure enables progressive generation from coarse to fine, matching the needs of multi-level semantic representation of complex terrain. Secondly, multi-scale pyramid decomposition is performed on the digital elevation model data. Large-scale terrain skeleton features, mesoscale terrain transition features, and small-scale terrain detail features are obtained by sequential downsampling through Gaussian or Laplacian pyramids. These features represent the macroscopic undulations, regional transitions, and micro-terrain textures of the terrain, providing corresponding scale terrain constraints for each branch. Then, the joint input features are input into the large-scale terrain constraint branch and fused with the large-scale terrain skeleton features using gated fusion. Specifically, gated cyclic units or gated linear units control the injection ratio of terrain information into the optical generation process. A terrain-guided generation strategy is used to model the global spatial structure, that is, using terrain skeleton features to constrain the overall layout and terrain orientation of the image, outputting a low-resolution large-scale base image. This ensures that the generated image has correct terrain skeleton support from the beginning, avoiding structural deviations in subsequent refinement stages. Subsequently, the large-scale base image is upsampled through bilinear interpolation or transposed convolution and input into the mesoscale cross-modal fusion branch. Combining mesoscale terrain transition features with deep semantic information from the joint input features, a cross-modal attention mechanism is used to calculate the association weights between terrain features and optical semantic features, achieving alignment between optical texture and terrain features, suppressing optical representations that do not match the terrain, and obtaining a mesoscale refined image to bridge the semantic gap between large-scale structures and mesoscale textures. Next, the mesoscale refined image is upsampled and input into the small-scale detail completion branch, fusing small-scale terrain detail features with high-frequency texture information from the optical remote sensing image. Specifically, residual connections are used to introduce texture details from the original optical image, and an adaptive edge enhancement mechanism is used to strengthen key details. That is, edge detection operators are used to locate landslide boundaries and abrupt terrain change areas and perform feature enhancement to obtain the full-scale initial image. Finally, channel filtering and feature calibration are performed on the initial full-scale image. Multi-channel features are mapped to three-channel feature maps through a 1×1 convolutional layer. After instance normalization or batch normalization to stabilize the data distribution, a three-channel optical remote sensing image is obtained. This is done to convert the generator's internal feature representation into a standard optical image format, ensuring that the output meets the specifications of subsequent discriminator input and visual perception.

[0111] Step S60: The three-channel optical remote sensing images are stitched together with the corresponding batch of digital elevation model data to form generated samples, and the optical remote sensing images of the real landslide area are stitched together with the corresponding batch of real digital elevation model data as real samples.

[0112] It should be noted that the generated three-channel optical remote sensing images are selected and stitched together with corresponding batches of digital elevation model (DEM) data to construct generated samples for adversarial training. Conversely, optical remote sensing images of real landslide areas are selected and stitched together with corresponding batches of real DEM data to construct real samples for adversarial training. By stitching optical images and DEM data together according to corresponding batches to construct samples, it is ensured that the spatial location of optical and terrain information within the samples is accurately matched, providing structurally consistent and dimensionally complete sample data for subsequent adversarial discrimination.

[0113] Step S70: Input the generated samples and real samples into the discriminator for adversarial discrimination, and iteratively train the multi-scale generator and the multi-task discriminator alternately until the number of iterations reaches the preset maximum number of iterations, and output the current multi-scale generator as the target multi-scale generator.

[0114] It should be noted that the generated samples and real samples are input into the discriminator together. The discriminator performs multi-dimensional adversarial discrimination and outputs the discrimination result. Based on the discrimination result, the multi-scale generator and the multi-task discriminator are trained alternately and iteratively. During the game process, the two continuously optimize their own parameters and performance. When the number of training iterations reaches the preset maximum number of iterations, the training is terminated and the currently optimized multi-scale generator is output, which is used as the target multi-scale generator for subsequent data generation. Through adversarial discrimination, the generator and discriminator form a closed loop of mutual game and joint optimization, which drives the generated samples to continuously approach the characteristics of real samples, significantly improving the authenticity and rationality of the generated data.

[0115] Step S80: Input the optical remote sensing image of the area to be enhanced, the digital elevation model data, and the auxiliary disaster perception data into the target multi-scale generator to obtain the target satellite landslide data.

[0116] It should be noted that the areas to be enhanced refer to geographical areas where there are problems such as scarce satellite landslide data, insufficient data accuracy, and incomplete data coverage, and where satellite landslide data needs to be supplemented or optimized through data generation methods.

[0117] Further, step S80 includes: constructing a multi-dimensional prior model library based on historical real satellite landslide data, wherein the multi-dimensional prior model library includes data distribution patterns of different terrain complexities and data distribution patterns corresponding to different disaster development stages; fusing the multi-dimensional prior model library with a target multi-scale generator to obtain an optimization generator; preprocessing the optical remote sensing image, digital elevation model data, and auxiliary disaster perception data of the area to be enhanced to obtain standardized data; and inputting the standardized data into the optimization generator to obtain initial satellite landslide data, wherein the initial satellite landslide data includes terrain structure information, disaster feature information, and optical texture. Information: The initial satellite landslide data is compared dimension-by-dimensionally with source data in the real satellite landslide data sample library to obtain terrain structure deviation values, disaster feature deviation values, and optical texture deviation values. Based on the terrain structure deviation values, disaster feature deviation values, and optical texture deviation values, the initial satellite landslide data is optimized using a gradient descent algorithm to obtain optimized satellite landslide data. Based on the optimized satellite landslide data, a multi-dimensional quality assessment is performed to obtain data credibility, feature consistency, and scene adaptability. When the data credibility, feature consistency, and scene adaptability all meet the corresponding preset thresholds, the optimized satellite landslide data is used as the target satellite landslide data.

[0118] It should be noted that topographic complexity is a comprehensive indicator used to describe the complexity of a geographical region's topographic relief, valley distribution, slope changes, etc. Different topographic complexities correspond to different landslide data distribution characteristics. The disaster development stage refers to the different stages in the complete life cycle of a landslide disaster, from its inception, emergence, development to stabilization. Landslide areas at different stages exhibit different characteristics in terms of topography, soil and rock, and thermal radiation, corresponding to different satellite landslide data distribution patterns.

[0119] Specifically, a multi-dimensional prior model library is constructed based on historical real satellite landslide data. Cluster analysis is used to extract data distribution patterns corresponding to different terrain complexities, including characteristic statistics of flat areas, gentle slopes, steep slopes, and complex gully areas. Simultaneously, data distribution patterns corresponding to different disaster development stages are extracted, including evolutionary characteristics of the gestation period, creep period, acceleration period, and instability period. A prior constraint model covering both terrain and disaster dimensions is established. This is done to introduce physically reasonable distribution boundaries and avoid abnormal samples from the generator outputting data that deviate from the actual landslide patterns. Secondly, the multi-dimensional prior model library is integrated with the target multi-scale generator. By adding a prior constraint regularization term to the generator's loss function or using a Bayesian neural network framework, the generation process is controlled by historical data statistical patterns, resulting in an optimized generator. Then, the optical remote sensing images, digital elevation model data, and auxiliary disaster perception data of the area to be enhanced are preprocessed, including geometric registration, radiometric correction, normalization, and spatial resampling, to obtain standardized data, ensuring that the input data meets the receiving specifications of the optimized generator. Subsequently, standardized data is input into the optimization generator, and initial satellite landslide data is obtained through forward inference. This data contains topographic structure information, disaster feature information, and optical texture information, but may contain slight deviations from the actual distribution. Next, the initial satellite landslide data is compared dimension-by-dimensionally with source data in a real satellite landslide data sample library. The topographic structure deviation value is obtained by calculating the elevation correlation coefficient, the disaster feature deviation value is obtained by calculating the deformation field mutual information, and the optical texture deviation value is obtained by calculating the texture feature cosine similarity, quantifying the difference between the generated data and the real data. Then, based on the topographic structure deviation value, disaster feature deviation value, and optical texture deviation value, the initial satellite landslide data is optimized using a gradient descent algorithm. Specifically, the objective function is to minimize the sum of the three-dimensional deviations, and iterative fine-tuning is performed in the data space to obtain optimized satellite landslide data, gradually approximating the center of the actual distribution. Next, a multi-dimensional quality assessment is conducted based on optimized satellite landslide data. Data reliability is obtained by calculating the distribution matching degree with the prior model library; feature consistency is obtained by calculating the similarity between the internal feature covariance matrix and real samples; and scene adaptability is obtained by calculating the degree of conformity with the geographical environmental parameters of the area to be enhanced. Finally, when data reliability, feature consistency, and scene adaptability all meet their respective preset thresholds, the optimized satellite landslide data is used as the target satellite landslide data. This is done to ensure that the final output data meets application standards in terms of statistical characteristics, physical laws, and scene adaptability through a closed-loop control of multiple rounds of generation-comparison-optimization-evaluation.

[0120] This embodiment integrates optical remote sensing images, DEM and InSAR deformation, thermal infrared and other auxiliary disaster perception data to construct an adaptive attention fusion mechanism for topographic gradient field, curvature field, disaster intensity field and thermal anomaly field, generating multi-source disaster perception features. A multi-scale generator is used to combine DEM multi-scale features to generate images step by step, and a multi-task discriminator is used to jointly judge optical realism and topographic plausibility, ultimately obtaining target satellite landslide data. Through precise alignment of multi-source disaster perception data and optical features, satellite landslide data with both topographic structural plausibility and optical texture realism is generated, improving the accuracy and reliability of the generated complex terrain satellite landslide data.

[0121] 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 satellite landslide data generation method based on multi-scale adversarial neural networks, step S70, further includes steps S201 to S210:

[0122] Step S201: Construct a multi-task discriminator.

[0123] It should be noted that the multi-task discriminator includes a feature extraction network, a terrain structure discrimination branch, a disaster feature discrimination branch, and a global authenticity discrimination branch. The terrain structure discrimination branch loads prior parameters of digital elevation model features, and the disaster feature discrimination branch loads a disaster intensity and thermal anomaly feature template library.

[0124] In step S202, the generated samples and real samples are input into the underlying feature extraction network of the multi-task discriminator to obtain a shared feature map.

[0125] Specifically, generated samples and real samples are respectively input into the bottom feature extraction network of the multi-task discriminator. This network consists of multiple convolutional layers, batch normalization layers, and activation function layers stacked together. The local texture and structural features of the samples are extracted through layer-by-layer convolution operations. At the bottom network stage, the types of real and fake samples are not distinguished, and the same weight parameters are used for forward propagation to obtain a shared feature map. This is done to ensure that generated samples and real samples are comparable in feature space, avoid feature distribution shifts caused by differences in network parameters, and provide a unified and aligned feature basis for subsequent multi-branch discrimination.

[0126] Step S203: Input the shared feature map into the terrain structure discrimination branch, and obtain the terrain rationality score by matching and calculating with the prior parameters of the digital elevation model features.

[0127] It should be noted that the channel features of the corresponding digital elevation model in the shared feature map are separated and extracted to obtain the DEM feature sub-map. Then, the DEM feature sub-map is matched with the prior parameters of the digital elevation model features loaded in the terrain structure discrimination branch. Specifically, the similarity between the generated terrain and the real terrain in terms of attributes such as slope distribution, elevation undulation, and terrain roughness is quantified by calculating the Euclidean distance or KL divergence between the statistical moments (mean, variance) of the DEM feature sub-map and the real DEM statistical distribution stored in the prior parameters, or by calculating the correlation coefficient between the DEM feature sub-map and the typical terrain pattern template in the prior parameters. Finally, the matching calculation result is mapped to a probability value between 0 and 1 through a fully connected layer and a Sigmoid activation function, and the terrain rationality score is output. This is done to enable the discriminator to have professional terrain structure identification capabilities, and to identify structural defects such as terrain skeleton distortion and elevation anomalies in the generated samples, so as to provide the generator with targeted terrain constraint gradient signals.

[0128] Step S204: Input the shared feature map into the disaster feature discrimination branch, and calculate the disaster typicality score by comparing it with the disaster intensity and thermal anomaly feature template library.

[0129] It should be noted that the channel features corresponding to the catastrophic intensity field and thermal anomaly field in the shared feature map are separated and extracted to obtain a catastrophic feature sub-map. Then, this catastrophic feature sub-map is input into the catastrophic feature discrimination branch. Through comparison with the catastrophic intensity and thermal anomaly feature template library, the catastrophic feature sub-map is divided into multiple spatial regions. The cosine similarity or Euclidean distance between each region and the prototype vector of each typical catastrophic pattern in the template library is calculated. The nearest neighbor similarity is selected as the matching score of the region. Then, the matching scores of all regions are spatially weighted and averaged. The weights are adaptively adjusted according to the spatial distance between the region and the landslide center. The closer to the center, the higher the weight. Finally, the weighted average matching score is mapped to a probability value between 0 and 1 through a fully connected layer and a Sigmoid activation function, and the catastrophic typicality score is output. This is done so that the discriminator can identify whether the catastrophic features in the generated samples conform to the deformation evolution law and thermal anomaly distribution pattern of the real landslide, avoid generating false catastrophic signals with unreasonable physical characteristics, and provide the generator with refined feedback on the authenticity of catastrophic features, guiding it to generate satellite landslide data that conforms to the catastrophic mechanism.

[0130] Step S205: Input the shared feature map into the global realism discrimination branch, and obtain the overall realism score by mapping through a fully connected layer and activating with Sigmoid.

[0131] It should be noted that the shared feature map is input into the global authenticity discrimination branch, and the spatial dimension is compressed into a statistical vector of channel dimension through global average pooling operation to aggregate global context information; then the statistical vector is input into the fully connected layer for feature mapping and dimension transformation to extract high-level semantic discrimination features; finally, the output value is mapped to a probability value between 0 and 1 by the Sigmoid activation function to obtain the overall authenticity score.

[0132] Step S206: Construct a joint loss function based on terrain rationality score, disaster typicality score and overall sample authenticity score.

[0133] It should be noted that the weight coefficients of the joint loss function include global realism weight coefficients, terrain structure weight coefficients, and disaster feature weight coefficients. The global realism weight coefficients, terrain structure weight coefficients, and disaster feature weight coefficients are adaptively adjusted according to the current training iteration round.

[0134] Specifically, the terrain rationality score, disaster typicality score, and overall realism score are compared with their corresponding labels to calculate the binary cross-entropy loss, resulting in terrain structure loss, disaster feature loss, and global realism loss. Then, a joint loss function is constructed based on these three losses. Specifically, the global realism loss is multiplied by the global realism weight coefficient, the terrain structure loss is multiplied by the terrain structure weight coefficient, and the disaster feature loss is multiplied by the disaster feature weight coefficient. The three weight coefficients are then weighted and summed to obtain the discriminator's total loss. The three weight coefficients are adaptively adjusted according to the current training iteration. Specifically, a higher global realism weight coefficient is set in the early stage of training to ensure the basic visual quality of the generated images. As training progresses, the terrain structure weight coefficient and the disaster feature weight coefficient are gradually increased, so that the discriminator pays more attention to physical rationality and disaster feature realism in the later stages.

[0135] Step S207: Fix the parameters of the multi-scale generator, and update the parameters of the multi-task discriminator by using the backpropagation algorithm with a preset learning rate based on the calculation results of the joint loss function, thereby obtaining the optimized discriminator.

[0136] It should be noted that the parameters of the multi-task discriminator include the feature extraction network parameters, terrain structure discrimination branch parameters, disaster feature discrimination branch parameters, and global authenticity discrimination branch parameters.

[0137] Specifically, all parameters of the multi-scale generator are fixed, including the convolution weights, normalization parameters, and attention mechanism parameters of each scale generation branch, ensuring that the generator remains unchanged in the current training phase, thus isolating the discriminator optimization process. Then, the gradient of the joint loss function is calculated layer by layer through the backpropagation algorithm, tracing back from the joint loss output to the parameters of each layer of the feature extraction network, terrain structure discrimination branch, disaster feature discrimination branch, and global realism discrimination branch. Specifically, the gradient values ​​of the parameters of the feature extraction network, terrain structure discrimination branch, disaster feature discrimination branch, and global realism discrimination branch are calculated separately. Finally, all the above parameters are updated by the Adam or SGD optimizer according to the preset learning rate to obtain the optimized discriminator. This is done so that the discriminator can fully learn the difference distribution between real samples and generated samples under the condition that the generator is fixed, improve the professional discrimination ability of each branch, provide a more accurate and stable gradient signal for the next round of generator training, and form a dynamic balance mechanism of alternating optimization between the discriminator and the generator.

[0138] Step S208: Fix the parameters of the optimized discriminator, input the generated sample into the optimized discriminator to obtain the updated terrain rationality score, the updated disaster typicality score, and the updated overall authenticity score.

[0139] It should be noted that all parameters of the optimized discriminator are fixed, including the parameters of the feature extraction network, terrain structure discrimination branch, disaster feature discrimination branch, and global realism discrimination branch. This ensures that the discriminator remains unchanged during the current training phase, thus isolating the generator's optimization process. Then, the generated samples are input into the optimized discriminator, passing through the feature extraction network to obtain updated shared feature maps, which are then input into the terrain structure discrimination branch, disaster feature discrimination branch, and global realism discrimination branch respectively. The updated terrain rationality score is calculated by matching with the prior parameters of the digital elevation model features, the updated disaster typicality score is calculated by comparing with the disaster intensity and thermal anomaly feature template library, and the updated overall realism score is obtained through fully connected layer mapping and Sigmoid activation. This is done to re-evaluate the quality of the current generated samples under the condition of a fixed discriminator, providing the generator with the latest multi-dimensional feedback signals, enabling the generator to adjust its optimization direction according to the discriminator's current discrimination ability, and gradually improve the terrain rationality, disaster typicality, and visual realism of the generated data.

[0140] Step S209: Calculate the generator loss value based on the updated terrain rationality score, updated disaster typicality score, and updated overall realism score combined with the joint loss function. Update the parameters of the multi-scale generator using the gradient descent algorithm to obtain the optimized generator.

[0141] It should be noted that the updated terrain rationality score, updated disaster typicality score, and updated overall realism score are each compared with the true labels (all 1s) to calculate the binary cross-entropy loss, obtaining the generator's component losses in the three discriminative dimensions. Then, based on these three component losses and the joint loss function, the generator loss value is calculated. Specifically, the loss corresponding to the updated overall realism score is multiplied by the global realism weight coefficient, the loss corresponding to the updated terrain rationality score is multiplied by the terrain structure weight coefficient, and the loss corresponding to the updated disaster typicality score is multiplied by the disaster feature weight coefficient. These three are then weighted and summed to obtain the generator's total loss. This loss aims to deceive the discriminator, i.e., to achieve the desired result. The discriminator is expected to misclassify generated samples as real samples. Finally, the gradient descent algorithm is used to calculate the gradient of the generator loss value with respect to the parameters of each layer of the multi-scale generator, including the convolution weights, normalization parameters, and attention mechanism parameters of the large-scale terrain constraint branch, the medium-scale cross-modal fusion branch, and the small-scale detail completion branch. All the above parameters are updated by the Adam optimizer according to the preset learning rate to obtain the optimized generator. This is done so that the generator can adjust the generation strategy according to the latest feedback from the discriminator, and gradually improve the quality of the generated samples in the three dimensions of visual realism, terrain rationality, and disaster typicality, forming an adversarial training process of dynamic game with the discriminator.

[0142] Step S210: Repeat the alternating iterative process of discriminator training and generator training. When the number of iterations reaches the preset maximum number of iterations, output the current optimized generator as the target multi-scale generator.

[0143] It should be noted that the process involves repeatedly performing alternating iterative training of the discriminator and the generator. This involves fixing the generator parameters to train the discriminator to improve its discrimination ability, then fixing the discriminator parameters again to train the generator to improve its generation quality, creating a dynamic game. The current iteration count is then monitored. When the number of iterations reaches the preset maximum, it is determined that the adversarial training has fully converged. At this point, the generator and discriminator have reached a Nash equilibrium, and the generator has the ability to stably generate high-quality satellite landslide data. Finally, the current optimized generator is output as the target multi-scale generator. This approach ensures that the generator fully learns the distribution characteristics of real data through sufficient iterative training, while avoiding pattern collapse caused by overtraining. This guarantees that the final model can generate satellite landslide data with reasonable terrain structure, realistic disaster characteristics, and lifelike optical textures in complex terrain scenarios.

[0144] This embodiment constructs a multi-task discriminator with a feature extraction network and three discriminative branches, loads prior parameters and a template library, extracts shared feature maps of samples and completes multi-dimensional scoring, constructs a weighted adaptive joint loss function, iteratively trains the discriminator and generator, optimizes the parameters in a fixed loop, and outputs a target multi-scale generator after reaching a preset number of iterations. This achieves accurate multi-dimensional discrimination and model collaborative optimization, improves generator performance, and ensures that the subsequently generated satellite landslide data is both realistic, has reasonable terrain, and is typical of disasters, thus guaranteeing the quality of data generation.

[0145] Based on the first embodiment of this application, this application also provides a satellite landslide data generation device based on a multi-scale adversarial neural network. Please refer to... Figure 3 The device includes:

[0146] The acquisition module 10 is used to acquire optical remote sensing images, digital elevation model data and auxiliary disaster perception data of the same geographical area, wherein the auxiliary disaster perception data includes InSAR deformation data and thermal infrared image data.

[0147] The calculation module 20 is used to calculate the terrain gradient field and terrain curvature field based on the digital elevation model data, and to calculate the disaster intensity field and thermal anomaly field based on the auxiliary disaster perception data.

[0148] The multimodal fusion module 30 is used to calculate the weight coefficients of each modal data through an adaptive attention mechanism, which combines the terrain gradient field, terrain curvature field, disaster intensity field and thermal anomaly field to generate multi-source disaster perception features.

[0149] The cross-modal stitching module 40 is used to stitch together multi-source disaster perception features and optical remote sensing images across modalities to form joint input features that include terrain and disaster information.

[0150] The image generation module 50 is used to input the joint input features into the multi-scale generator, and combine the multi-scale features of the digital elevation model data to generate a three-channel optical remote sensing image step by step.

[0151] The sample construction module 60 is used to stitch together three-channel optical remote sensing images with corresponding batches of digital elevation model data to form generated samples, and to stitch together optical remote sensing images of real landslide areas with corresponding batches of real digital elevation model data as real samples.

[0152] The adversarial training module 70 is used to input the generated samples and the real samples into the discriminator for adversarial discrimination, and to alternately iterate the training of the multi-scale generator and the multi-task discriminator until the number of iterations reaches the preset maximum number of iterations, and output the current multi-scale generator as the target multi-scale generator.

[0153] The results module 80 is used to input the optical remote sensing image of the area to be enhanced, the digital elevation model data, and the auxiliary disaster perception data into the target multi-scale generator to obtain the target satellite landslide data.

[0154] The satellite landslide data generation device based on multi-scale adversarial neural networks provided in this application, employing the satellite landslide data generation method based on multi-scale adversarial neural networks described in the above embodiments, can solve the technical problem of how to improve the accuracy of generated satellite landslide data. Compared with the prior art, the beneficial effects of the satellite landslide data generation device based on multi-scale adversarial neural networks provided in this application are the same as those of the satellite landslide data generation method based on multi-scale adversarial neural networks provided in the above embodiments, and other technical features in the satellite landslide data generation device based on multi-scale adversarial neural networks are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0155] This application provides a satellite landslide data generation device based on a multi-scale adversarial neural network. The satellite landslide data generation device based on a multi-scale adversarial neural network includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the satellite landslide data generation method based on a multi-scale adversarial neural network in the above embodiment 1.

[0156] The following is for reference. Figure 4 The diagram illustrates a structural schematic of a satellite landslide data generation device based on a multi-scale adversarial neural network suitable for implementing embodiments of this application. The satellite landslide data generation device based on a multi-scale adversarial neural network in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), 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 satellite landslide data generation device based on multi-scale adversarial neural networks shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0157] like Figure 4As shown, the satellite landslide data generation device based on a multi-scale adversarial neural network 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 satellite landslide data generation device based on the multi-scale adversarial neural network. 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, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the satellite landslide data generation device based on multi-scale adversarial neural networks to communicate wirelessly or wiredly with other devices to exchange data. Although various satellite landslide data generation devices based on multi-scale adversarial neural networks 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.

[0158] 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.

[0159] The satellite landslide data generation device based on multi-scale adversarial neural networks provided in this application, employing the satellite landslide data generation method based on multi-scale adversarial neural networks described in the above embodiments, can solve the technical problem of how to improve the accuracy of generated satellite landslide data. Compared with the prior art, the beneficial effects of the satellite landslide data generation device based on multi-scale adversarial neural networks provided in this application are the same as those of the satellite landslide data generation method based on multi-scale adversarial neural networks provided in the above embodiments, and other technical features in this satellite landslide data generation device based on multi-scale adversarial neural networks are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0160] 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.

[0161] 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.

[0162] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the satellite landslide data generation method based on a multi-scale adversarial neural network in the above embodiments.

[0163] 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.

[0164] The aforementioned computer-readable medium may be included in a satellite landslide data generation device based on a multi-scale adversarial neural network; or it may exist independently and not assembled into a satellite landslide data generation device based on a multi-scale adversarial neural network.

[0165] The aforementioned computer-readable medium carries one or more programs that, when executed by a satellite landslide data generation device based on a multi-scale adversarial neural network, enable the 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).

[0166] 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.

[0167] 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.

[0168] The readable medium provided in this application is a computer-readable medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-described satellite landslide data generation method based on multi-scale adversarial neural networks, and can solve the technical problem of how to improve the accuracy of the generated satellite landslide data. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as the beneficial effects of the satellite landslide data generation method based on multi-scale adversarial neural networks provided in the above embodiments, and will not be repeated here.

[0169] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the satellite landslide data generation method based on a multi-scale adversarial neural network as described above.

[0170] The computer program product provided in this application solves the technical problem of how to improve the accuracy of generated satellite landslide data. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the satellite landslide data generation method based on multi-scale adversarial neural networks provided in the above embodiments, and will not be repeated here.

[0171] 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 satellite landslide data generation method based on a multi-scale adversarial neural network, characterized by, The method includes: Acquire optical remote sensing images, digital elevation model data, and auxiliary disaster perception data for the same geographical area, wherein the auxiliary disaster perception data includes InSAR deformation data and thermal infrared image data; The terrain gradient field and terrain curvature field are calculated based on the digital elevation model data, and the disaster intensity field and thermal anomaly field are calculated based on the auxiliary disaster perception data. The terrain gradient field, the terrain curvature field, the catastrophic intensity field, and the thermal anomaly field are used to calculate the weight coefficients of each modal data through an adaptive attention mechanism to generate multi-source disaster perception features; The multi-source disaster perception features and the optical remote sensing images are stitched together across modalities to form a joint input feature that includes terrain and disaster information; The joint input features are input into a multi-scale generator, and combined with the multi-scale features of the digital elevation model data, a three-channel optical remote sensing image is generated step by step. The three-channel optical remote sensing images are stitched together with the corresponding batch of digital elevation model data to form generated samples, and the optical remote sensing images of the real landslide area are stitched together with the corresponding batch of real digital elevation model data as real samples. The generated samples and the real samples are input into the discriminator for adversarial discrimination. The multi-scale generator and the multi-task discriminator are trained alternately and iteratively until the number of iterations reaches the preset maximum number of iterations. The current multi-scale generator is then output as the target multi-scale generator. The optical remote sensing image, digital elevation model data and auxiliary disaster perception data of the area to be enhanced are input into the target multi-scale generator to obtain the target satellite landslide data; The step of inputting the joint input features into the multi-scale generator and combining them with the multi-scale features of the digital elevation model data to generate a three-channel optical remote sensing image step by step includes: Construct a multi-scale generator, wherein the multi-scale generator includes a large-scale terrain constraint branch, a medium-scale cross-modal fusion branch, and a small-scale detail completion branch; Multi-scale pyramid decomposition is performed on the digital elevation model data to obtain large-scale terrain skeleton features, meso-scale terrain transition features, and small-scale terrain detail features in sequence. The joint input features are input into the large-scale terrain constraint branch, and gated fusion is performed with the large-scale terrain skeleton features. The global spatial structure is modeled through the terrain-guided generation strategy, and a large-scale base image is output. The large-scale base image is upsampled and input into the mesoscale cross-modal fusion branch. The optical texture and terrain features are aligned through a cross-modal attention mechanism by combining the mesoscale terrain transition features with the deep semantic information in the joint input features to obtain a mesoscale refined image. The mesoscale refined image is upsampled and then input into the small-scale detail completion branch. The small-scale terrain detail features are fused with the high-frequency texture information of the optical remote sensing image. Key details are enhanced through an adaptive edge enhancement mechanism to obtain the full-scale initial image. The initial full-scale image is subjected to channel filtering and feature calibration, and mapped to a three-channel feature map through a 1×1 convolutional layer. After normalization, a three-channel optical remote sensing image is obtained.

2. The method of claim 1, wherein, The steps of calculating the terrain gradient field and terrain curvature field based on the digital elevation model data, and calculating the disaster intensity field and thermal anomaly field based on the auxiliary disaster perception data, include: The digital elevation model data is subjected to Gaussian filtering for noise reduction to obtain the noise-reduced digital elevation model data. Based on the denoised digital elevation model data, the finite difference method is used to calculate the terrain slope components in the x and y directions, and the terrain gradient field is obtained through vector synthesis. The profile curvature and plane curvature of the denoised digital elevation model data are solved by second-order partial derivatives, and then weighted and fused according to preset weights to form a terrain curvature field, wherein the preset weights are adaptively adjusted according to the complexity of the regional terrain. Spatiotemporal baseline correction is performed on the InSAR deformation data of the auxiliary disaster perception data to extract the deformation rate of the potential landslide area, and the disaster intensity field is obtained by normalization based on the deformation rate. The thermal infrared image data of the auxiliary disaster perception data is subjected to atmospheric correction and radiometric calibration to convert it into real temperature data; Based on the real temperature data, thermal anomaly regions are extracted using an adaptive threshold segmentation algorithm to generate a thermal anomaly field.

3. The method of claim 1, wherein, The step of calculating the weight coefficients of each modal data (topography gradient field, topography curvature field, catastrophic intensity field, and thermal anomaly field) through an adaptive attention mechanism to generate multi-source disaster perception features includes: Construct a dual-branch attention network, which includes a channel attention branch and a spatial attention branch; The terrain gradient field, the terrain curvature field, the catastrophic intensity field, and the thermal anomaly field are input into the channel attention branch. The channel statistical features of each modality data are extracted through global mean pooling and global maximum pooling. The channel weight coefficients of each modality data are output through a fully connected layer and a Softmax activation function. After multiplying each modal data by its corresponding channel weight coefficient, the weighted modal data is obtained. The weighted modal data are input into the spatial attention branch, and a spatial attention map is obtained through cross-modal feature interaction calculation. The modal data are the terrain gradient field, curvature field, catastrophic intensity field and thermal anomaly field. The spatial attention map is used to characterize the importance of catastrophic features at different spatial locations. The spatial attention map and the weighted modal data are multiplied pixel by pixel to obtain the double-weighted modal features; Modality correlation constraint loss is introduced to iteratively optimize the dual-weighted modal features to obtain the optimized modal features. The optimized modal features are concatenated along the channel dimension, and feature dimension fusion and dimensionality reduction are performed through a 1×1 convolutional layer to generate multi-source disaster perception features.

4. The method of claim 1, wherein, The step of cross-modal stitching of the multi-source disaster perception features and the optical remote sensing image to form a joint input feature containing terrain and disaster information includes: The multi-source disaster perception features are subjected to feature enhancement processing to extract deep disaster correlation features; Multi-scale feature extraction is performed on the optical remote sensing image to obtain optical texture features at different levels; Based on the deep disaster association features and optical texture features at each level, a semantic similarity matrix is ​​obtained. Based on the semantic similarity matrix, cross-modal fusion is performed using a two-dimensional fusion method to obtain fused features, wherein the two-dimensional fusion method is channel dimension splicing and spatial dimension superposition. By introducing modality consistency constraints and optimizing the fused features through a comparative loss function, the optimized fused features are obtained. Based on the optimized fusion features and unified feature distribution range, a joint input feature containing terrain and disaster information is formed.

5. The method of claim 1, wherein, The steps of inputting the generated samples and the real samples into the discriminator for adversarial discrimination, alternately iteratively training the multi-scale generator and the multi-task discriminator until the number of iterations reaches a preset maximum number of iterations, and outputting the current multi-scale generator as the target multi-scale generator include: A multi-task discriminator is constructed, wherein the multi-task discriminator includes a feature extraction network, a terrain structure discrimination branch, a disaster feature discrimination branch, and a global authenticity discrimination branch. The terrain structure discrimination branch loads prior parameters of digital elevation model features, and the disaster feature discrimination branch loads a disaster intensity and thermal anomaly feature template library. The generated samples and the real samples are input into the low-level feature extraction network of the multi-task discriminator to obtain a shared feature map; The shared feature map is input into the terrain structure discrimination branch, and the terrain rationality score is obtained by matching and calculating with the prior parameters of the digital elevation model features. The shared feature map is input into the disaster feature discrimination branch, and the disaster typicality score is obtained by comparing it with the disaster intensity and thermal anomaly feature template library. The shared feature map is input into the global authenticity discrimination branch, and the overall authenticity score is obtained by mapping through a fully connected layer and activating with Sigmoid. A joint loss function is constructed based on the terrain rationality score, disaster typicality score, and overall sample authenticity score. The weight coefficients of the joint loss function include global authenticity weight coefficients, terrain structure weight coefficients, and disaster feature weight coefficients. The global authenticity weight coefficients, terrain structure weight coefficients, and disaster feature weight coefficients are adaptively adjusted according to the current training iteration round. The parameters of the multi-scale generator are fixed, and the calculation results of the joint loss function are used to update the parameters of the multi-task discriminator according to a preset learning rate through the backpropagation algorithm to obtain the optimized discriminator. The parameters of the multi-task discriminator include the feature extraction network parameters, terrain structure discrimination branch parameters, disaster feature discrimination branch parameters, and global authenticity discrimination branch parameters. By fixing the parameters of the optimized discriminator, the generated sample is input into the optimized discriminator to obtain an updated terrain rationality score, an updated disaster typicality score, and an updated overall authenticity score; Based on the updated terrain rationality score, updated disaster typicality score, and updated overall realism score, combined with the joint loss function, the generator loss value is calculated, and the parameters of the multi-scale generator are updated through the gradient descent algorithm to obtain an optimized generator. Repeatedly execute the alternating iterative process of discriminator training and generator training. When the number of iterations reaches the preset maximum number of iterations, output the current optimized generator as the target multi-scale generator.

6. The method of claim 1, wherein, The step of inputting the optical remote sensing image of the area to be enhanced, digital elevation model data, and auxiliary disaster perception data into the target multi-scale generator to obtain target satellite landslide data includes: A multi-dimensional prior model library is constructed based on historical real satellite landslide data. The multi-dimensional prior model library includes the data distribution patterns of different terrain complexities and the data distribution patterns corresponding to different disaster development stages. The multi-dimensional prior model library is fused with the target multi-scale generator to obtain an optimized generator; The optical remote sensing images, digital elevation model data, and auxiliary disaster perception data of the area to be enhanced are preprocessed to obtain standardized data. The standardized data is input into the optimization generator to obtain initial satellite landslide data, wherein the initial satellite landslide data includes terrain structure information, disaster feature information and optical texture information; The initial satellite landslide data is compared dimension by dimension with the source data in the real satellite landslide data sample library to obtain the terrain structure deviation value, disaster characteristic deviation value and optical texture deviation value. Based on the terrain structure deviation value, disaster feature deviation value, and optical texture deviation value, the initial satellite landslide data is optimized using a gradient descent algorithm to obtain optimized satellite landslide data. Based on the optimized satellite landslide data, a multi-dimensional quality assessment was conducted to obtain data reliability, feature consistency, and scenario adaptability. When the data credibility, feature consistency, and scene adaptability all meet the corresponding preset thresholds, the optimized satellite landslide data will be used as the target satellite landslide data. 7.A satellite landslide data generation device based on a multi-scale adversarial neural network, characterized by, The apparatus is applied to the satellite landslide data generation method based on multi-scale adversarial neural networks as described in any one of claims 1-6, and the apparatus comprises: The acquisition module is used to acquire optical remote sensing images, digital elevation model data and auxiliary disaster perception data of the same geographical area, wherein the auxiliary disaster perception data includes InSAR deformation data and thermal infrared image data. The calculation module is used to calculate the terrain gradient field and terrain curvature field based on the digital elevation model data, and to calculate the disaster intensity field and thermal anomaly field based on the auxiliary disaster perception data. The multimodal fusion module is used to calculate the weight coefficients of each modal data by means of the terrain gradient field, the terrain curvature field, the disaster intensity field and the thermal anomaly field through an adaptive attention mechanism, so as to generate multi-source disaster perception features; The cross-modal stitching module is used to stitch the multi-source disaster perception features and the optical remote sensing image across modalities to form a joint input feature that includes terrain and disaster information; The image generation module is used to input the joint input features into the multi-scale generator, and combine the multi-scale features of the digital elevation model data to generate a three-channel optical remote sensing image step by step. The sample construction module is used to stitch the three-channel optical remote sensing images with the corresponding batch of digital elevation model data to form generated samples, and to stitch the optical remote sensing images of the real landslide area with the corresponding batch of real digital elevation model data as real samples. The adversarial training module is used to input the generated samples and the real samples into the discriminator for adversarial discrimination, and to alternately iterate the training of the multi-scale generator and the multi-task discriminator until the number of iterations reaches the preset maximum number of iterations, and output the current multi-scale generator as the target multi-scale generator. The results module is used to input the optical remote sensing image of the area to be enhanced, digital elevation model data, and auxiliary disaster perception data into the target multi-scale generator to obtain target satellite landslide data.

8. A satellite landslide data generation device based on a multi-scale adversarial neural network, characterized in that, The device includes: a memory, a processor, and a satellite landslide data generation program based on a multi-scale adversarial neural network stored in the memory and running on the processor, the satellite landslide data generation program based on the multi-scale adversarial neural network being configured to implement the steps of the satellite landslide data generation method based on a multi-scale adversarial neural network as described in any one of claims 1-6.

9. A storage medium, characterized in that, The storage medium stores a satellite landslide data generation program based on a multi-scale adversarial neural network. When the satellite landslide data generation program based on the multi-scale adversarial neural network is executed by the processor, it implements the steps of the satellite landslide data generation method based on the multi-scale adversarial neural network as described in any one of claims 1-6.

Images