Method and system for identifying hot melt collapse based on comprehensive deformation, structure and form

Abstract

The invention discloses a hot melt collapse identification method and system based on comprehensive deformation, structure and form. Optical images, SAR images and terrain and geological data are obtained; performing wave band synthesis on the optical image, the SAR image and the topographic and geological data to obtain synthesized data; and inputting the synthesized data into a hot melt slump identification model for training, and gradually adjusting the weight and deviation by the hot melt slump identification model through multiple iterations and back propagation, thereby effectively capturing and learning the form and characteristics of hot melt slump and outputting a hot melt slump identification result. According to the method, a multi-source data collaborative analysis framework is established by fusing remote sensing images, earth surface deformation data and other multi-source data such as topography and geology, and slump characteristics are comprehensively described. And an improved deep learning model is adopted, and a multi-feature fusion recognition model is established in combination with the morphological features of the slump boundary, so that automatic recognition and accurate monitoring of the hot melt slump are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster identification, and in particular to a method and system for identifying thermal melt collapses based on comprehensive deformation, structure and morphology. Background Art

[0002] Thermal thaw landslides are sudden geological disasters caused by rising temperatures in permafrost regions, leading to the melting of underground ice and the destabilization of surface structures. In recent years, with the intensification of global warming, ground temperatures in typical permafrost regions such as Northeast my country and the Qinghai-Tibet Plateau have continued to rise, accelerating the degradation of permafrost and causing frequent thermal thaw landslides. These disasters are characterized by their sudden and widespread destructive nature, threatening the safety of infrastructure such as railways, highways, oil pipelines, and power transmission corridors. They also accelerate the release of permafrost carbon, alter regional hydrological conditions, and have profound impacts on the ecological environment. For example, thermal thaw landslides in Northeast China have led to the degradation of surface vegetation, increased soil erosion, and significantly affected the stability of roadbeds along power transmission lines. Therefore, the development of efficient and accurate thermal thaw landslide monitoring and identification technologies is of great practical significance for disaster early warning, engineering protection, and ecological conservation. Summary of the Invention

[0003] The present invention realizes automatic identification and accurate monitoring of thermal melt collapse by providing a thermal melt collapse identification method and system based on comprehensive deformation, structure and morphology.

[0004] The present invention provides a method for identifying thermal melt collapse based on comprehensive deformation, structure and morphology, comprising:

[0005] Obtain optical images, SAR images and topographic and geological data;

[0006] Performing band synthesis on the optical image, SAR image and topographic and geological data to obtain synthesized data;

[0007] The synthesized data is input into the thermal melt landslide recognition model for training. The thermal melt landslide recognition model gradually adjusts the weights and biases through multiple iterations and back propagation, thereby effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting thermal melt landslide recognition results. The overall architecture of the thermal melt landslide recognition model consists of an encoder, a decoder and a jump connection. In the encoder part, the first two layers of encoding modules perform local feature extraction based on ResNet50, which can effectively capture the local texture and edge information in the optical image, SAR image and topographic geological data. The third and fourth layers of encoding modules use Swin Transformer Block to learn cross-scale global features through sliding window self-attention, thereby improving the model's perception of complex landform areas and long time series information. In the decoder part, a U-Net decoding structure is adopted, and a combination of shallow ResNet features + deep Swin Transformer features are used for multi-scale information fusion; in the jump connection part, CBAM is used to enhance the feature selection capability in the encoding-decoding connection channel, so that the shallow detail features and deep global information are more fully integrated. At the same time, the key spatial areas are selected through the adaptive attention mechanism to enhance the segmentation accuracy of the target area; finally, the thermal melt landslide mask map is generated through the Softmax or Sigmoid activation function to realize the recognition of thermal melt landslide.

[0008] Specifically, after obtaining the optical image, SAR image and topographic and geological data, the method further includes:

[0009] performing atmospheric correction, radiation correction, registration, panchromatic and multispectral fusion, and cropping processing on the optical image to obtain a processed optical image;

[0010] The SAR image is subjected to radiation correction, geometric correction, and orbital precision correction to obtain a corrected SAR image; a small baseline set strategy is used to select a corrected SAR image pair with a small temporal and spatial baseline for interferometric processing; a phase difference of the interferometric SAR image pair is calculated to generate an interferogram; the interferogram is subjected to superposition and difference processing to obtain a processed interferogram; and a time series analysis is performed on the processed interferogram to obtain a cumulative deformation amount.

[0011] performing clipping, resampling, and reprojection processing on the topographic and geological data to obtain processed topographic and geological data;

[0012] The performing band synthesis on the optical image, SAR image and topographic and geological data to obtain synthesized data includes:

[0013] The processed optical image, the accumulated deformation amount and the processed topographic and geological data are subjected to band synthesis to obtain the synthesized data.

[0014] Specifically, performing a time series analysis on the processed interference graph to obtain a cumulative deformation amount includes:

[0015] performing singular value decomposition on the processed interference pattern to extract linear and nonlinear components of the surface deformation;

[0016] The linear and nonlinear components of the surface deformation are reconstructed to obtain a complete deformation field.

[0017] Specifically, after obtaining the synthesized data, the method further includes:

[0018] outlining the landslide boundary from the synthesized data based on the optical image, and performing binarization processing on the landslide boundary and the background area to generate a raster image;

[0019] Overlapping and cropping the raster image to obtain an overlapping and cropped image block;

[0020] The step of inputting the synthesized data into a thermal melt collapse identification model for training comprises:

[0021] The overlapped and cropped image blocks are input into the thermal melt and landslide recognition model for training.

[0022] Specifically, during the training of the thermal melt collapse identification model, the Adam optimizer is selected, the initial learning rate is set to 1e-4, and the cosine annealing strategy is used to dynamically adjust the weights and biases.

[0023] The present invention also provides a thermal melt collapse identification system that integrates deformation, structure, and morphology, including:

[0024] Data acquisition module, used to obtain optical images, SAR images and topographic and geological data;

[0025] A data synthesis module is used to perform band synthesis on the optical image, SAR image and topographic and geological data to obtain synthesized data;

[0026] The thermal melt landslide recognition module is used to input the synthesized data into the thermal melt landslide recognition model for training. The thermal melt landslide recognition model gradually adjusts the weights and biases through multiple iterations and back propagation, thereby effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting thermal melt landslide recognition results. The overall architecture of the thermal melt landslide recognition model consists of an encoder, a decoder and a jump connection. In the encoder part, the first two layers of encoding modules perform local feature extraction based on ResNet50, which can effectively capture the local texture and edge information in the optical image, SAR image and topographic geological data. The third and fourth layers of encoding modules use Swin Transformer Block to learn cross-scale global features through sliding window self-attention, thereby improving the model's perception of complex landform areas and long time series information. In the decoder part, a U-Net decoding structure is adopted, and a combination of shallow ResNet features + deep Swin Transformer features are used for multi-scale information fusion; in the jump connection part, CBAM is used to enhance the feature selection capability in the encoding-decoding connection channel, so that the shallow detail features and deep global information are more fully integrated. At the same time, the key spatial areas are selected through the adaptive attention mechanism to enhance the segmentation accuracy of the target area; finally, the thermal melt landslide mask map is generated through the Softmax or Sigmoid activation function to realize the recognition of thermal melt landslide.

[0027] Specifically, it also includes:

[0028] An optical image processing module, configured to perform atmospheric correction, radiation correction, registration, panchromatic and multispectral fusion, and cropping processing on the optical image to obtain a processed optical image;

[0029] The SAR image processing module is configured to perform radiation correction, geometric correction, and orbital precision correction on the SAR image to obtain a corrected SAR image; select a corrected SAR image pair with a small temporal and spatial baseline using a small baseline set strategy for interferometric processing; calculate the phase difference of the interferometric SAR image pair to generate an interferogram; perform superposition and difference processing on the interferogram to obtain a processed interferogram; and perform time series analysis on the processed interferogram to obtain a cumulative deformation;

[0030] A topographic and geological data processing module is used to perform clipping, resampling, and reprojection processing on the topographic and geological data to obtain processed topographic and geological data;

[0031] The data synthesis module is specifically used to perform band synthesis on the processed optical image, the accumulated deformation and the processed topographic and geological data to obtain the synthesized data.

[0032] Specifically, the SAR image processing module includes:

[0033] A SAR image correction unit, configured to perform radiation correction, geometric correction, and orbital precision correction on the SAR image to obtain a corrected SAR image;

[0034] SAR image interferometry unit, used to select the corrected SAR image pairs with small temporal and spatial baselines for interferometry processing using a small baseline set strategy;

[0035] an interference pattern generating unit, configured to perform superposition and differential processing on the interference pattern to obtain a processed interference pattern;

[0036] a singular value decomposition unit, configured to perform singular value decomposition on the processed interference pattern to extract linear and nonlinear components of the surface deformation;

[0037] The deformation field obtaining unit is used to reconstruct the linear component and the nonlinear component of the surface deformation to obtain a complete deformation field.

[0038] Specifically, it also includes:

[0039] a raster image generation module for outlining the landslide boundary from the synthesized data based on the optical image after obtaining the synthesized data, and performing binarization processing on the landslide boundary and the background area to generate a raster image;

[0040] A raster image overlapping and cropping module is used to perform overlapping and cropping on the raster image to obtain an overlapping and cropped image block;

[0041] The thermal melt landslide recognition module is specifically used to input the overlapping cropped image blocks into the thermal melt landslide recognition model for training. The thermal melt landslide recognition model gradually adjusts the weights and biases through multiple iterations and back propagation, thereby effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting thermal melt landslide recognition results. The overall architecture of the thermal melt landslide recognition model consists of an encoder, a decoder and a jump connection. In the encoder part, the first two layers of encoding modules perform local feature extraction based on ResNet50, which can effectively capture the local texture and edge information in the optical image, SAR image and topographic and geological data; the third and fourth layers of encoding modules use Swin Transformer Block, through sliding window self-attention learning cross-scale global features, improves the model's perception of complex landform areas and long time series information; in the decoder part, the U-Net decoding structure is adopted, and the shallow ResNet features + deep SwinTransformer features are combined for multi-scale information fusion; in the jump connection part, CBAM is used to enhance the feature selection ability in the encoding-decoding connection channel, so that the shallow detail features and the deep global information are more fully integrated, and at the same time, the key spatial areas are selected through the adaptive attention mechanism to enhance the segmentation accuracy of the target area; finally, the thermal melt landslide mask map is generated through the Softmax or Sigmoid activation function to realize the recognition of thermal melt landslide.

[0042] Specifically, during the training of the thermal melt collapse identification model, the Adam optimizer is selected, the initial learning rate is set to 1e-4, and the cosine annealing strategy is used to dynamically adjust the weights and biases.

[0043] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0044] Obtain optical images, SAR images, and topographic and geological data; synthesize the optical images, SAR images, and topographic and geological data bands to obtain synthesized data; input the synthesized data into a thermal melt landslide identification model for training. The thermal melt landslide identification model gradually adjusts weights and biases through multiple iterations and backpropagation, effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting thermal melt landslide identification results. This invention establishes a multi-source data collaborative analysis framework by fusing remote sensing imagery, surface deformation data, and other multi-source data such as topography and geology to comprehensively characterize landslide characteristics. Furthermore, an improved deep learning model is used, combined with the morphological characteristics of landslide boundaries, to establish a multi-feature fusion recognition model, enabling automatic identification and precise monitoring of thermal melt landslides. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1A flow chart of a method for identifying thermal melt collapse based on comprehensive deformation, structure, and morphology provided in an embodiment of the present invention;

[0046] Figure 2 Schematic diagram of a method for identifying thermal melt collapse based on comprehensive deformation, structure, and morphology provided by an embodiment of the present invention;

[0047] Figure 3 A schematic diagram of the structure of a thermal melt collapse identification model in a thermal melt collapse identification method based on comprehensive deformation, structure, and morphology provided in an embodiment of the present invention;

[0048] Figure 4 Module diagram of the thermal melt collapse identification system that integrates deformation, structure and morphology provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The embodiment of the present invention realizes automatic identification and accurate monitoring of thermal melt collapse by providing a method and system for identifying thermal melt collapse based on comprehensive deformation, structure and morphology.

[0050] The technical solution in the embodiment of the present invention is to achieve the above technical effects, and the overall idea is as follows:

[0051] S1: Collect different types of remote sensing data, including optical images, SAR images, and other multi-source data (such as DEM, precipitation, etc.) to provide comprehensive information on surface deformation, landform structure, and environment.

[0052] S2: Perform different preprocessing on the collected remote sensing data.

[0053] S3: After preprocessing and spatial and temporal unification, remote sensing data are synthesized into images with multiple bands.

[0054] S4: Based on high-resolution images and deformation data, manually or semi-automatically outline the boundaries of the thermal melt landslide area, construct a sample dataset for model training, and divide the sample dataset into a training set and a validation set.

[0055] S5: The dataset is fed into a deep learning model for training to learn the morphology and characteristics of thermal melt collapses. The model's recognition accuracy and generalization ability are evaluated by calculating metrics such as mean intersection over union (MIoU), intersection over union (IoU), and recall.

[0056] S6: Use different spectral band combinations for model training, analyze the impact of different bands on the ability to identify thermal melt landslides, and optimize the best input combination.

[0057] S7: Apply the trained model to the target study area to automatically extract the boundary information of the thermal melt landslide and realize the automated detection of the thermal melt landslide area.

[0058] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0059] like Figure 1 、 Figure 2 and Figure 3 As shown, the embodiment of the present invention provides a method for identifying thermal melt collapse based on comprehensive deformation, structure and morphology, including:

[0060] Step S110: Obtain optical images, SAR images and topographic and geological data (such as DEM, precipitation data, vegetation index, geological type and land cover, etc.) to provide comprehensive surface deformation, landform structure and environmental information.

[0061] Step S120: performing band synthesis on the optical image, SAR image and topographic and geological data to obtain synthesized data;

[0062] The method for identifying thermal melt collapse based on comprehensive deformation, structure, and morphology provided by an embodiment of the present invention is specifically described. After obtaining optical images, SAR images, and topographic and geological data, the method further includes:

[0063] Perform atmospheric correction, radiation correction, registration, panchromatic and multispectral fusion, and cropping on the optical image to obtain the processed optical image;

[0064] Specifically, atmospheric correction is performed on optical images using methods based on radiative transfer models (such as the 6S model) or the fast-path radiative transfer model (Flaash) to eliminate or reduce image distortion caused by atmospheric scattering and absorption, thereby restoring the true reflectivity of the ground objects. Radiometric correction is then performed to calibrate the image's radiometric values ​​to eliminate radiometric inconsistencies caused by factors such as sensor characteristics, imaging conditions, and observation geometry. In this embodiment, radiometric correction includes calibration coefficient correction, dark current correction, and solar altitude correction to ensure radiometric consistency between images, facilitating subsequent multi-temporal and multi-source image analysis. Registration involves geometrically aligning the images so that they precisely overlap within the same geographic coordinate system. Panchromatic and multispectral fusion leverages the high spatial resolution of panchromatic images and the rich spectral information of multispectral images to generate fused images with high spatial and spectral resolution. In this embodiment, panchromatic and multispectral fusion is achieved through IHS transformation, Brovey transformation, PCA transformation, or deep learning-based fusion methods. The fused image can take into account both spatial and spectral characteristics to improve image interpretation accuracy. Finally, irrelevant parts of the image are removed based on the boundaries of the study area, retaining the valid data within the study area. Cropping not only reduces data storage and computational complexity, but also highlights the characteristics of the study area, improving the efficiency of model training and inference.

[0065] The SAR images are subjected to radiation correction (correction of the amplitude of the radar signal to eliminate the influence of sensor characteristics and imaging conditions), geometric correction (mapping the SAR images to a unified geographic coordinate system based on satellite orbit parameters and terrain models), and orbital precision correction (correcting the geometric position error of the image based on accurate orbital data) to obtain the corrected SAR images; using the small baseline set strategy, corrected SAR image pairs with smaller temporal and spatial baselines are selected for interferometric processing; the core of the small baseline set strategy is to reduce the spatial and temporal decoherence problems caused by excessive spatial and temporal baselines by selecting SAR image pairs with shorter baselines. The phase difference of the interferometric SAR image pairs is calculated to generate an interferogram to reflect the surface deformation information in the study area; the interferograms are superimposed and differentially processed to obtain the processed interferograms to eliminate or reduce the errors caused by interference factors such as atmospheric delay, orbit error, and terrain error; the processed interferograms are subjected to time series analysis to obtain the cumulative deformation;

[0066] Specifically, the processed interference pattern is subjected to time series analysis to obtain the cumulative deformation, including:

[0067] Perform singular value decomposition on the processed interferogram to extract the linear and nonlinear components of the surface deformation;

[0068] The linear and nonlinear components of the surface deformation are reconstructed to obtain the complete deformation field.

[0069] The topographic and geological data are clipped, resampled, and reprojected to obtain processed topographic and geological data; specifically, a vector file is used to define the study area, and GIS tools (such as ArcGIS, QGIS) or Python libraries (such as GDAL, Rasterio) are used to clip the raster data according to the vector boundary. The data is converted from one geographic coordinate system to another to ensure that all data are superimposed and analyzed under the same projection system. The spatial resolution of the data is adjusted so that different data sets have the same resolution or zoom level. The nearest neighbor interpolation method is selected to adjust the spatial resolution of the data so that different data sets have the same resolution. In this embodiment, the specific method of the nearest neighbor interpolation method is: in the set of known data points, the known point closest to the target position is found, and then the attribute value of the nearest neighbor point is directly assigned to the target position.

[0070] In this case, the optical image, SAR image and topographic and geological data are synthesized to obtain the synthesized data, including:

[0071] Band synthesis is performed on the processed optical imagery, accumulated deformation, and processed topographic and geological data to generate synthesized data. Specifically, data from different sources and types are combined into a multi-band data set, generating a data structure of the form (channels, height, width). Channels represents the number of bands formed after synthesis, including original optical bands (such as RGB, NIR), derived indices (such as NDVI), topographic factors (such as slope, aspect, DEM), deformation data, and geological data (such as lithology).

[0072] Step S130: The synthesized data is input into the thermal melt landslide recognition model for training. The thermal melt landslide recognition model gradually adjusts the weights and biases through multiple iterations and back propagation, thereby effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting the thermal melt landslide recognition results. During the training process, the model automatically extracts key features related to thermal melt landslides from the input data, such as the boundary morphology of the landslide, the sliding direction, the texture characteristics of the landslide body, the terrain undulation, the vegetation coverage, and possible inducing factors (such as precipitation, temperature changes, etc.). The overall architecture of the thermal melt-slide recognition model consists of an encoder, a decoder, and skip connections. In the encoder, the first two encoding modules use ResNet50 for local feature extraction, effectively capturing local texture and edge information from optical images, SAR images, and topographic and geological data. The third and fourth encoding modules utilize the Swin Transformer Block, which learns cross-scale global features through sliding window self-attention, improving the model's perception of complex terrain areas and long-term temporal information. In the decoder, a U-Net decoding structure is employed, combining shallow ResNet features with deep Swin Transformer features for multi-scale information fusion. In the skip connection, CBAM is employed to enhance feature selection in the encoder-decoder pipeline, enabling a more comprehensive fusion of shallow detail features with deep global information. An adaptive attention mechanism is also employed to select key spatial regions, enhancing the segmentation accuracy of the target region. Finally, a Softmax or Sigmoid activation function is used to generate a thermal melt-slide mask, enabling thermal melt-slide recognition. The detailed workflow of the thermal melt-slide recognition model is as follows: First, in the local feature extraction stage, ResNet50 is used as the first two encoder layers. The input data undergoes convolution, batch normalization, and reinforced linear unit (ReLU) processing to extract preliminary features. Subsequently, high-level local features are further extracted through the convolutional layers and identity fast of ResNet50. Next, in the global feature modeling stage, a two-layer SwinTransformer Block is used for feature learning, leveraging sliding window self-attention to extract global information within a wider receptive field. Patch merging is also used for dimensionality reduction, reducing computational complexity and improving cross-layer information fusion. The decoding stage uses a U-Net decoding structure to gradually restore spatial resolution and combines shallow-layer ResNet features with deep-layer SwinTransformer features for multi-scale information fusion. The skip connection introduces a CBAM module to further integrate shallow-layer detail features with deep-layer global information. An adaptive attention mechanism selects key spatial regions to enhance segmentation accuracy in the target region. Finally, a melt-slump mask is generated using a Softmax or Sigmoid activation function for accurate recognition.

[0073] The method for identifying thermal melt collapse based on comprehensive deformation, structure, and morphology provided in an embodiment of the present invention is further described. After obtaining the synthesized data, the method further includes:

[0074] Using optical imagery, the landslide boundary is delineated from the synthesized data. High-resolution remote sensing data, such as Google Earth, can be used to improve recognition accuracy. The landslide boundary and background areas are binarized to generate a raster image. During binarization, the landslide area is labeled as positive (value 1) and the background area as negative (value 0). NoData values ​​or outliers (such as the value 32767 in the DEM data) are handled to avoid anomalies during model training.

[0075] Perform overlapping cropping on the raster image to obtain overlapping cropped image blocks. Specifically, the overlap between adjacent image blocks is 20% to 30%. Before cropping, ensure that the image and label data have the same size and spatial position alignment.

[0076] In this case, the synthesized data is fed into the thermal melt-slump identification model for training, including:

[0077] The overlapping cropped image blocks are input into the thermal melt and landslide recognition model for training.

[0078] In order to improve the robustness and generalization ability of the model, data enhancement is performed on the cropped sub-image blocks through rotation, flipping, scaling, brightness adjustment, etc. When performing data enhancement, the label data and image data are kept synchronized to ensure that the enhanced label information is consistent with the enhanced image. After completing the data enhancement, the dataset is divided into training set, validation set, and test set in a ratio of 8:1:1. During the division process, the representativeness and balance of the data are maintained to ensure that the training set contains samples of landslide areas under different terrains, different lighting conditions, and different background environments. At the same time, the samples in the validation set and test set are kept independent of the training set to prevent model overfitting problems caused by data leakage.

[0079] The training of the thermal melt-slump identification model is described in detail. During training, the Adam optimizer is used, the initial learning rate is set to 1e-4, and a cosine annealing strategy is employed to dynamically adjust the weights and biases. Specifically, the learning rate is gradually reduced during training in the form of a cosine function, reaching a minimum of 1% of the initial value. This strategy maintains a large learning step size in the early stages of training, accelerating convergence. Meanwhile, the learning rate is gradually reduced in the later stages of training to improve model stability and generalization.

[0080] In this embodiment, the model framework is built using PyTorch, and experiments are conducted on an NVIDIA GeForce RTX 4070 12G GPU to fully utilize the powerful computing power of the GPU and accelerate the model training and inference process.

[0081] In order to comprehensively evaluate the recognition accuracy and generalization ability of the model, a variety of performance indicators are used to quantitatively analyze the detection results of the model. Specifically, by calculating key indicators such as MIoU, IoU, Recall, Precision, Accuracy, etc., the performance of the model is comprehensively measured. Among them, MioU is an important indicator to measure the overall performance of the model in multi-category segmentation tasks. It represents the mean IoU of the model across all categories, reflecting the model's comprehensive recognition ability for targets of different categories. IoU is used to evaluate the regional overlap of the model in a specific category. The calculation formula is the ratio of the intersection area of ​​the predicted result and the true label to the union area. The recall rate (Recall) is used to measure the model's ability to recognize positive samples. The higher the value, the higher the detection rate of the model for positive samples. The formula is as follows:

[0082]

[0083] Among them, TP refers to true positive, which means the number of pixels that are accurately identified as positive; FP refers to false positive, which means the number of pixels that are mistakenly classified as positive when the true label is negative; FN refers to false negative, which means the number of pixels that are mistakenly classified as negative when the true label is positive; k represents the total number of categories.

[0084] To further analyze the role and contribution of various spectral bands in identifying thermal melt landslides, different spectral band combinations were selected as input during model training. The model was then trained and validated, comparing the performance of these different combinations (such as precision, recall, and F1 score) to identify the bands that contribute most significantly to thermal melt landslide identification. Through comparative analysis and feature importance assessment, the optimal band input combination was optimized, thereby improving the model's recognition accuracy and robustness in complex terrain and environmental conditions.

[0085] The method for identifying thermal melt collapse based on comprehensive deformation, structure, and morphology provided by an embodiment of the present invention is further described, and further includes:

[0086] The trained model was applied to typical thermal melt landslide areas, such as the Hoh Xil area. Multi-source remote sensing data for the study area was acquired and preprocessed. This preprocessed data was then fed into the trained model to automatically extract thermal melt landslide boundary information. Post-processing further enhanced the integrity and accuracy of the boundary information, generating a landslide distribution map. Comparison and verification with historical landslide distribution data further evaluated the model's extraction accuracy and applicability.

[0087] In this example, after obtaining and extracting the spatial distribution results of thermal thaw landslides, their spatial characteristics were further analyzed. First, spatial analysis methods were used to clarify the geographical distribution patterns of thermal thaw landslides and identify their spatial differences under different terrain, slope, aspect, landform type, and groundwater conditions. Simultaneously, remote sensing imagery and geographic information system (GIS) technology were combined to construct a spatial distribution model to identify high-risk areas and their potential expansion trends. Second, based on the spatial characteristic analysis, time series data was further introduced to systematically study the dynamic evolution of thermal thaw landslides. By comparing and analyzing the location, area, morphology, and activity frequency of landslides in different periods, the occurrence, development, and disappearance patterns of thermal thaw landslides were revealed, and their temporal variation trends and evolution patterns were clarified. Furthermore, by combining external environmental factors such as meteorological data (such as temperature and precipitation), freeze-thaw cycle intensity, groundwater level changes, and geological conditions (such as soil type and lithology), the main controlling factors affecting the dynamic evolution of thermal thaw landslides and their interaction mechanisms were explored.

[0088] like Figure 4 As shown, the embodiment of the present invention provides a comprehensive deformation, structure and morphology thermal melt collapse identification system, including:

[0089] The data acquisition module 100 is used to obtain optical images, SAR images and topographic and geological data (such as DEM, precipitation data, vegetation index, geological type and land cover, etc.) to provide comprehensive surface deformation, landform structure and environmental information.

[0090] The data synthesis module 200 is used to synthesize the optical image, SAR image and topographic and geological data to obtain synthesized data;

[0091] The thermal melt collapse identification system based on comprehensive deformation, structure and morphology provided by the embodiment of the present invention is specifically described, and further includes:

[0092] The optical image processing module is used to perform atmospheric correction, radiation correction, registration, panchromatic and multispectral fusion, and cropping on the optical image to obtain the processed optical image;

[0093] Specifically, the optical image processing module is used to perform atmospheric correction on optical images using methods based on radiation transfer models (such as the 6S model) or the fast-path radiation transfer model (Flaash) to eliminate or reduce image distortion caused by atmospheric scattering and absorption, thereby restoring the true reflectivity of the ground objects. Next, radiometric correction is performed to calibrate the image's radiometric values ​​to eliminate radiometric inconsistencies caused by factors such as sensor characteristics, imaging conditions, and observation geometry. In this embodiment, radiometric correction includes calibration coefficient correction, dark current correction, and solar altitude correction to ensure radiometric consistency between images, facilitating subsequent multi-temporal and multi-source image analysis. Registration involves geometrically aligning the images so that they precisely overlap within the same geographic coordinate system. Panchromatic and multispectral fusion leverages the high spatial resolution of panchromatic images and the rich spectral information of multispectral images to generate fused images with high spatial and spectral resolution. In this embodiment, panchromatic and multispectral fusion is achieved through IHS transformation, Brovey transformation, PCA transformation, or deep learning-based fusion methods. The fused image balances spatial and spectral characteristics to improve image interpretation accuracy. Finally, irrelevant parts of the image are removed based on the boundaries of the study area, retaining the valid data within the study area. Cropping not only reduces data storage and computational complexity, but also highlights the characteristics of the study area, improving the efficiency of model training and inference.

[0094] The SAR image processing module is used to perform radiation correction (correct the amplitude of the radar signal to eliminate the influence of sensor characteristics and imaging conditions), geometric correction (map the SAR image to a unified geographic coordinate system based on satellite orbit parameters and terrain models), and orbital precision correction (correct the geometric position error of the image based on accurate orbital data) on the SAR image to obtain the corrected SAR image; using the small baseline set strategy, select the corrected SAR image pairs with smaller time and space baselines for interference processing; the core of the small baseline set strategy is to reduce the spatial and temporal decoherence problems caused by excessive spatial and temporal baselines by selecting SAR image pairs with shorter baselines. Calculate the phase difference of the interferometric SAR image pair to generate an interference pattern to reflect the surface deformation information in the study area; perform superposition and differential processing on the interference pattern to obtain the processed interference pattern to eliminate or reduce the errors caused by interference factors such as atmospheric delay, orbit error and terrain error; perform time series analysis on the processed interference pattern to obtain the cumulative deformation;

[0095] Specifically, the SAR image processing module includes:

[0096] SAR image correction unit, used to perform radiation correction, geometric correction and orbit correction on the SAR image to obtain the corrected SAR image;

[0097] SAR image interferometry unit, used to select the corrected SAR image pairs with small temporal and spatial baselines for interferometry processing using a small baseline set strategy;

[0098] An interference pattern generating unit, used for performing superposition and differential processing on the interference pattern to obtain a processed interference pattern;

[0099] A singular value decomposition unit is used to perform singular value decomposition on the processed interferogram to extract the linear component and nonlinear component of the surface deformation;

[0100] The deformation field acquisition unit is used to reconstruct the linear and nonlinear components of the surface deformation to obtain a complete deformation field.

[0101] The topographic and geological data processing module is used to perform cropping, resampling, and reprojection on the topographic and geological data to obtain processed topographic and geological data;

[0102] Specifically, the topographic and geological data processing module is specifically used to define the study area using a vector file, and use GIS tools (such as ArcGIS, QGIS) or Python libraries (such as GDAL, Rasterio) to clip raster data according to the vector boundary. Convert the data from one geographic coordinate system to another to ensure that all data are superimposed and analyzed under the same projection system. Adjust the spatial resolution of the data so that different data sets have the same resolution or zoom level. And select the nearest neighbor interpolation method to adjust the spatial resolution of the data so that different data sets have the same resolution. In this embodiment, the specific method of the nearest neighbor interpolation method is: in the set of known data points, find the known point closest to the target position, and then directly assign the attribute value of the nearest neighbor point to the target position.

[0103] In this case, the data synthesis module 200 is specifically used to perform band synthesis on the processed optical image, accumulated deformation, and processed topographic and geological data to obtain synthesized data. Specifically, data from different sources and types are merged into a multi-band data set, generating data in the form of (channels, height, width). Channels represents the number of bands formed after synthesis, including original optical bands (such as RGB, NIR), derived indices (such as NDVI), topographic factors (such as slope, aspect, DEM), deformation data, geological data (such as lithology), etc.

[0104] The thermal melt landslide identification module 300 is used to input the synthesized data into the thermal melt landslide identification model for training. The thermal melt landslide identification model gradually adjusts the weights and biases through multiple iterations and back propagation, thereby effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting thermal melt landslide identification results. During the training process, the model will automatically extract key features related to thermal melt landslides from the input data, such as the boundary morphology of the landslide, the sliding direction, the texture characteristics of the landslide body, the terrain undulation, the vegetation coverage, and possible inducing factors (such as precipitation, temperature changes, etc.). The overall architecture of the thermal melt-slide recognition model consists of an encoder, a decoder, and skip connections. In the encoder, the first two encoding modules use ResNet50 for local feature extraction, effectively capturing local texture and edge information from optical images, SAR images, and topographic and geological data. The third and fourth encoding modules utilize the Swin Transformer Block, which learns cross-scale global features through sliding window self-attention, improving the model's perception of complex terrain areas and long-term temporal information. In the decoder, a U-Net decoding structure is employed, combining shallow ResNet features with deep Swin Transformer features for multi-scale information fusion. In the skip connection, CBAM is employed to enhance feature selection in the encoder-decoder pipeline, enabling a more comprehensive fusion of shallow detail features with deep global information. An adaptive attention mechanism is also employed to select key spatial regions, enhancing the segmentation accuracy of the target region. Finally, a Softmax or Sigmoid activation function is used to generate a thermal melt-slide mask, enabling thermal melt-slide recognition. The detailed workflow of the thermal melt-slide recognition model is as follows: First, in the local feature extraction stage, ResNet50 is used as the first two encoder layers. The input data undergoes convolution, batch normalization, and reinforced linear unit (ReLU) processing to extract preliminary features. Subsequently, high-level local features are further extracted through the convolutional layers and identity fast layers of ResNet50. Next, in the global feature modeling stage, a two-layer Swin Transformer Block performs feature learning, leveraging sliding window self-attention to extract global information within a wider receptive field. Patch merging is also used for dimensionality reduction, reducing computational complexity and improving cross-layer information fusion. The decoding stage uses a U-Net decoding structure to gradually restore spatial resolution and combines shallow ResNet features with deep Swin Transformer features for multi-scale information fusion. The skip connection introduces a CBAM module to further integrate shallow-layer detail features with deep-layer global information. An adaptive attention mechanism selects key spatial regions to enhance segmentation accuracy in the target region. Finally, a melt-slump mask is generated using a Softmax or Sigmoid activation function for accurate recognition.

[0105] The thermal melt collapse identification system based on comprehensive deformation, structure and morphology provided by the embodiment of the present invention is further described, and further includes:

[0106] The raster image generation module, after obtaining the synthesized data, delineates the landslide boundary from the synthesized data using optical imagery. This delineation process can be combined with high-resolution remote sensing data such as Google Earth to improve recognition accuracy. The landslide boundary and background area are binarized to generate a raster image. During the binarization process, the landslide area is labeled as a positive class (value 1) and the background area is labeled as a negative class (value 0). NoData values ​​or outliers (such as the value 32767 in the DEM data) are handled to avoid anomalies during model training.

[0107] The raster image overlapping cropping module is used to perform overlapping cropping on raster images, obtaining overlapping cropped image blocks. Specifically, the overlap between adjacent image blocks is 20% to 30%. Before cropping, ensure that the image and label data have the same size and spatial alignment.

[0108] In this case, the thermal melt landslide recognition module 300 is specifically used to input the overlapping cropped image blocks into the thermal melt landslide recognition model for training. The thermal melt landslide recognition model gradually adjusts the weights and biases through multiple iterations and back propagation, thereby effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting thermal melt landslide recognition results. The overall architecture of the thermal melt landslide recognition model consists of an encoder, a decoder and a jump connection. In the encoder part, the first two layers of encoding modules perform local feature extraction based on ResNet50, which can effectively capture local texture and edge information in optical images, SAR images and topographic and geological data. The third and fourth layers of encoding modules use Swin Transformer Block to learn cross-scale global features through sliding window self-attention, thereby improving the model's perception of complex landform areas and long time series information. In the decoder part, a U-Net decoding structure is adopted, and a combination of shallow ResNet features + deep Swin Transformer features are used for multi-scale information fusion; in the jump connection part, CBAM is used to enhance the feature selection capability in the encoding-decoding connection channel, so that the shallow detail features and deep global information are more fully integrated. At the same time, the key spatial areas are selected through the adaptive attention mechanism to enhance the segmentation accuracy of the target area; finally, the thermal melt landslide mask map is generated through the Softmax or Sigmoid activation function to realize the recognition of thermal melt landslide.

[0109] In order to improve the robustness and generalization ability of the model, it also includes: a data post-processing module, which is used to perform data enhancement on the cropped sub-image blocks through rotation, flipping, scaling, brightness adjustment, etc. When performing data enhancement, the label data and image data are kept synchronously to ensure that the enhanced label information is consistent with the enhanced image. After completing the data enhancement, the data set is divided into training set, validation set and test set in a ratio of 8:1:1. During the division process, the representativeness and balance of the data are maintained to ensure that the training set contains samples of landslide areas under different terrains, different lighting conditions and different background environments. At the same time, the samples in the validation set and test set are kept independent of the training set to prevent model overfitting problems caused by data leakage.

[0110] The training of the thermal melt-slump identification model is described in detail. During training, the Adam optimizer is used, the initial learning rate is set to 1e-4, and a cosine annealing strategy is employed to dynamically adjust the weights and biases. Specifically, the learning rate is gradually reduced during training in the form of a cosine function, reaching a minimum of 1% of the initial value. This strategy maintains a large learning step size in the early stages of training, accelerating convergence. Meanwhile, the learning rate is gradually reduced in the later stages of training to improve model stability and generalization.

[0111] In this embodiment, the model framework is built using PyTorch, and experiments are conducted on an NVIDIA GeForce RTX 4070 12G GPU to fully utilize the powerful computing power of the GPU and accelerate the model training and inference process.

[0112] The thermal melt collapse identification system based on comprehensive deformation, structure, and morphology provided by the embodiment of the present invention is further described, and further includes:

[0113] The thaw-land collapse analysis module applies the trained model to typical thaw-land collapse areas, such as the Hoh Xil Basin. Multi-source remote sensing data for the study area is acquired and preprocessed. This preprocessed data is then fed into the trained model to automatically extract boundary information for thaw-land collapses. Post-processing further enhances the integrity and accuracy of the boundaries, generating a landslide distribution map. Comparison and validation with historical landslide distribution data further assess the model's extraction accuracy and applicability. Specifically, after obtaining and extracting the spatial distribution of thaw-land collapses, an in-depth analysis of their spatial characteristics is conducted. First, spatial analysis methods are used to clarify the geographic distribution patterns of thaw-land collapses and identify their spatial variability across different terrain, slope, aspect, landform type, and groundwater conditions. Furthermore, combining remote sensing imagery with geographic information system (GIS) technology, a spatial distribution model is constructed to identify high-risk areas and their potential expansion trends. Second, building on this spatial feature analysis, time series data is incorporated to systematically study the dynamic evolution of thaw-land collapses. By comparing and analyzing the location, area, morphology, and frequency of landslides over different periods, the authors reveal the patterns of occurrence, development, and regression of thermal thaw landslides, clarifying their temporal trends and evolutionary patterns. Furthermore, by combining external environmental factors such as meteorological data (such as temperature and precipitation), freeze-thaw cycle intensity, groundwater level fluctuations, and geological conditions (such as soil type and lithology), they explore the main controlling factors and their interaction mechanisms influencing the dynamic evolution of thermal thaw landslides.

[0114] In summary, the embodiments of the present invention combine remote sensing imagery, InSAR deformation monitoring data, and topographic parameters to leverage deep learning to extract geomorphological features of thermal thaw deformation zones within transmission corridors. This, combined with historical landslide samples, enables automatic identification, comprehensively capturing the dynamic characteristics of thermal thaw landslides and improving identification accuracy. These embodiments are suitable for thermal thaw landslide monitoring in permafrost areas, providing a scientific basis for geological disaster prevention and control and ecological and environmental protection.

[0115] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0117] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0119] Any details not described in the embodiments of the present invention are well-known to those skilled in the art. Finally, it should be noted that the above embodiments are only intended to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present invention may be modified or replaced with equivalents without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or equivalents should be included in the scope of the claims of the present invention.

Claims

1. A method for identifying thermal melt collapse based on comprehensive deformation, structure and morphology, characterized in that: include: Obtain optical images, SAR images and topographic and geological data; Performing band synthesis on the optical image, SAR image and topographic and geological data to obtain synthesized data; The synthesized data is input into the thermal melt landslide recognition model for training. The thermal melt landslide recognition model gradually adjusts the weights and biases through multiple iterations and back propagation, thereby effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting thermal melt landslide recognition results. The overall architecture of the thermal melt landslide recognition model consists of an encoder, a decoder and a jump connection. In the encoder part, the first two layers of encoding modules perform local feature extraction based on ResNet50, which can effectively capture the local texture and edge information in the optical image, SAR image and topographic geological data. The third and fourth layers of encoding modules use Swin Transformer Block to learn cross-scale global features through sliding window self-attention, thereby improving the model's perception of complex landform areas and long time series information. In the decoder part, a U-Net decoding structure is adopted, and a combination of shallow ResNet features + deep Swin Transformer features are used for multi-scale information fusion; in the jump connection part, CBAM is used to enhance the feature selection capability in the encoding-decoding connection channel, so that the shallow detail features and deep global information are more fully integrated. At the same time, the key spatial areas are selected through the adaptive attention mechanism to enhance the segmentation accuracy of the target area; finally, the thermal melt landslide mask map is generated through the Softmax or Sigmoid activation function to realize the recognition of thermal melt landslide.

2. The thermal melt collapse identification method based on comprehensive deformation, structure and morphology according to claim 1, characterized in that: After obtaining the optical image, SAR image and topographic and geological data, the method further includes: performing atmospheric correction, radiation correction, registration, panchromatic and multispectral fusion, and cropping processing on the optical image to obtain a processed optical image; The SAR image is subjected to radiation correction, geometric correction, and orbital precision correction to obtain a corrected SAR image; a small baseline set strategy is used to select a corrected SAR image pair with a small temporal and spatial baseline for interferometric processing; a phase difference of the interferometric SAR image pair is calculated to generate an interferogram; the interferogram is subjected to superposition and difference processing to obtain a processed interferogram; and a time series analysis is performed on the processed interferogram to obtain a cumulative deformation amount. performing clipping, resampling, and reprojection processing on the topographic and geological data to obtain processed topographic and geological data; The performing band synthesis on the optical image, SAR image and topographic and geological data to obtain synthesized data includes: The processed optical image, the accumulated deformation amount and the processed topographic and geological data are subjected to band synthesis to obtain the synthesized data.

3. The thermal melt collapse identification method based on comprehensive deformation, structure and morphology according to claim 2, characterized in that: The performing time series analysis on the processed interference graph to obtain a cumulative deformation amount includes: performing singular value decomposition on the processed interference pattern to extract linear and nonlinear components of the surface deformation; The linear and nonlinear components of the surface deformation are reconstructed to obtain a complete deformation field.

4. The thermal melt collapse identification method based on comprehensive deformation, structure and morphology according to claim 1, characterized in that: After obtaining the synthesized data, the method further includes: outlining the landslide boundary from the synthesized data based on the optical image, and performing binarization processing on the landslide boundary and the background area to generate a raster image; Overlapping and cropping the raster image to obtain an overlapping and cropped image block; The step of inputting the synthesized data into a thermal melt collapse identification model for training comprises: The overlapped and cropped image blocks are input into the thermal melt and landslide recognition model for training.

5. The method for identifying thermal melt collapse based on comprehensive deformation, structure and morphology according to any one of claims 1 to 4, characterized in that: During the training process of the thermal melt collapse identification model, the Adam optimizer was selected, the initial learning rate was set to 1e-4, and the cosine annealing strategy was used to dynamically adjust the weights and biases.

6. A thermal melt collapse identification system integrating deformation, structure and morphology, characterized in that: include: Data acquisition module, used to obtain optical images, SAR images and topographic and geological data; A data synthesis module is used to perform band synthesis on the optical image, SAR image and topographic and geological data to obtain synthesized data; The thermal melt landslide recognition module is used to input the synthesized data into the thermal melt landslide recognition model for training. The thermal melt landslide recognition model gradually adjusts the weights and biases through multiple iterations and back propagation, thereby effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting thermal melt landslide recognition results. The overall architecture of the thermal melt landslide recognition model consists of an encoder, a decoder and a jump connection. In the encoder part, the first two layers of encoding modules perform local feature extraction based on ResNet50, which can effectively capture the local texture and edge information in the optical image, SAR image and topographic geological data. The third and fourth layers of encoding modules use Swin Transformer Block to learn cross-scale global features through sliding window self-attention, thereby improving the model's perception of complex landform areas and long time series information. In the decoder part, a U-Net decoding structure is adopted, and a combination of shallow ResNet features + deep Swin Transformer features are used for multi-scale information fusion; in the jump connection part, CBAM is used to enhance the feature selection capability in the encoding-decoding connection channel, so that the shallow detail features and deep global information are more fully integrated. At the same time, the key spatial areas are selected through the adaptive attention mechanism to enhance the segmentation accuracy of the target area; finally, the thermal melt landslide mask map is generated through the Softmax or Sigmoid activation function to realize the recognition of thermal melt landslide.

7. The thermal melt collapse identification system based on comprehensive deformation, structure and morphology according to claim 6, characterized in that: Also includes: An optical image processing module, configured to perform atmospheric correction, radiation correction, registration, panchromatic and multispectral fusion, and cropping processing on the optical image to obtain a processed optical image; The SAR image processing module is used to perform radiation correction, geometric correction and orbital precision correction on the SAR image to obtain a corrected SAR image; using a small baseline set strategy, the corrected SAR image with a small temporal and spatial baseline is selected for interferometric processing; Calculating the phase difference of the interferometric SAR image pair to generate an interferogram; performing superposition and differential processing on the interferogram to obtain a processed interferogram; performing time series analysis on the processed interferogram to obtain a cumulative deformation; A topographic and geological data processing module is used to perform clipping, resampling, and reprojection processing on the topographic and geological data to obtain processed topographic and geological data; The data synthesis module is specifically used to perform band synthesis on the processed optical image, the accumulated deformation and the processed topographic and geological data to obtain the synthesized data.

8. The thermal melt collapse identification system based on comprehensive deformation, structure and morphology according to claim 7, characterized in that: The SAR image processing module includes: A SAR image correction unit, configured to perform radiation correction, geometric correction, and orbital precision correction on the SAR image to obtain a corrected SAR image; SAR image interferometry unit, used to select the corrected SAR image pairs with small temporal and spatial baselines for interferometry processing using a small baseline set strategy; an interference pattern generating unit, configured to perform superposition and differential processing on the interference pattern to obtain a processed interference pattern; a singular value decomposition unit, configured to perform singular value decomposition on the processed interference pattern to extract linear and nonlinear components of the surface deformation; The deformation field obtaining unit is used to reconstruct the linear component and the nonlinear component of the surface deformation to obtain a complete deformation field.

9. The thermal melt collapse identification system integrating deformation, structure and morphology according to claim 6, characterized in that: Also includes: a raster image generation module for outlining the landslide boundary from the synthesized data based on the optical image after obtaining the synthesized data, and performing binarization processing on the landslide boundary and the background area to generate a raster image; A raster image overlapping and cropping module is used to perform overlapping and cropping on the raster image to obtain an overlapping and cropped image block; The thermal melt landslide recognition module is specifically used to input the overlapping cropped image blocks into the thermal melt landslide recognition model for training. The thermal melt landslide recognition model gradually adjusts the weights and biases through multiple iterations and back propagation, thereby effectively capturing and learning the morphology and characteristics of thermal melt landslides and outputting thermal melt landslide recognition results. The overall architecture of the thermal melt landslide recognition model consists of an encoder, a decoder and a jump connection. In the encoder part, the first two layers of encoding modules perform local feature extraction based on ResNet50, which can effectively capture the local texture and edge information in the optical image, SAR image and topographic and geological data; the third and fourth layers of encoding modules use Swin Transformer Block, through sliding window self-attention learning cross-scale global features, improves the model's perception of complex landform areas and long time series information; in the decoder part, the U-Net decoding structure is adopted, and the shallow ResNet features + deep SwinTransformer features are combined for multi-scale information fusion; in the jump connection part, CBAM is used to enhance the feature selection ability in the encoding-decoding connection channel, so that the shallow detail features and the deep global information are more fully integrated, and at the same time, the key spatial areas are selected through the adaptive attention mechanism to enhance the segmentation accuracy of the target area; finally, the thermal melt landslide mask map is generated through the Softmax or Sigmoid activation function to realize the recognition of thermal melt landslide.

10. The thermal melt collapse identification system according to any one of claims 6 to 9, characterized in that: During the training process of the thermal melt collapse identification model, the Adam optimizer was selected, the initial learning rate was set to 1e-4, and the cosine annealing strategy was used to dynamically adjust the weights and biases.

Images