SAR image earthquake landslide boundary detection method and system based on deep learning

Abstract

The application discloses a SAR image earthquake landslide boundary detection method and system based on deep learning, and the method comprises the following steps: inputting two SAR images before and after a landslide to generate a normalized mean ratio image, inputting the normalized mean ratio image into a pre-trained landslide detection network model based on deep learning after upsampling, extracting a landslide, extracting a boundary, and obtaining a landslide edge detection result; the landslide detection network model comprises an encoder and a decoder, the encoder comprises a backbone network and a coordinate attention empty space pyramid module, and the decoder comprises an attention adjustment module and a multi-scale feature fusion resampling module.The application can realize rapid, high-precision and automatic detection of large-scale earthquake landslide boundaries, and can realize detection of different earthquake types and different sizes of earthquake landslides.

Description

Technical Field

[0001] This invention relates to the field of image processing and recognition technology, specifically to a method and system for detecting seismic landslide boundaries in SAR images based on deep learning. Background Technology

[0002] To ensure effective post-disaster emergency investigation, disaster assessment, and rescue, rapid, large-scale detection of earthquake-induced landslides is crucial. Currently, emergency investigations of earthquake-induced landslide disasters primarily rely on manual field surveys and visual interpretation of optical remote sensing images. Field surveys suffer from drawbacks such as travel difficulties and extreme time and labor consumption, making it impossible to provide an effective disaster assessment in a short period. Furthermore, using optical remote sensing images for landslide detection is often rendered unusable due to adverse weather conditions such as clouds, overcast skies, and rain before and after the earthquake-induced landslide. Other methods, such as UAVs and ground-based SAR measurements, cannot provide rapid responses to large-scale earthquake-induced landslide areas. Utilizing spaceborne SAR for earthquake-induced landslide detection can solve these problems. Spaceborne SAR systems can penetrate clouds, rain, and fog, enabling all-weather, 24 / 7 operation. They offer wide coverage (up to 250 km per image) and high spatial and temporal resolution, providing reliable SAR image data within a very short time after an earthquake-induced landslide (maximum 6 days; 2-3 days if combined with other spaceborne SAR images). Therefore, by using spaceborne SAR data, large-scale landslides triggered by earthquakes can be quickly located and precise landslide boundaries can be drawn, thus providing first-hand and reliable information support for emergency investigations of earthquake-induced landslides.

[0003] Currently, landslide detection methods based on spaceborne SAR images can be mainly divided into four categories: visual interpretation, feature thresholding, change detection, and machine learning. Machine learning methods require the extraction of various relevant features first, and then the use of machine learning classifiers to identify landslides, resulting in a higher degree of automation than the previous methods. However, when dealing with landslides of different types, backgrounds, or scales, this method requires manual feature selection or complex parameter adjustments, which greatly affects the detection speed and automation level. Currently, there is still relatively little research on landslide detection based on SAR images using deep learning. The main challenges it faces are (1) how to extract effective characterization information and essential features of earthquake landslides from SAR images, especially medium-resolution SAR data; and (2) how to use deep learning to extract high-precision earthquake landslide boundaries. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for detecting earthquake landslide boundaries based on deep learning using SAR images, which addresses the above-mentioned problems in the prior art. The present invention aims to achieve rapid, high-precision, and automated detection of large-scale earthquake landslide boundaries, and can also detect earthquake landslides of different earthquake types and scales.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A deep learning-based SAR image seismic landslide boundary detection method includes:

[0007] S101, Input two SAR images before and after the landslide;

[0008] S102, a normalized mean ratio map is generated based on two SAR images before and after the landslide, and then upsampled;

[0009] S103, the upsampled normalized mean ratio map is input into a pre-trained deep learning-based landslide detection network model for landslide extraction, and then the landslide extraction result is subjected to boundary extraction to obtain the landslide edge detection result; the landslide detection network model includes an encoder and a decoder. The encoder includes a backbone network and a coordinate attention hole spatial pyramid module (CAASP), and the decoder includes an attention adjustment module (AMM) and a multi-scale feature fusion resampling module (MFFRM); the backbone network extracts features at multiple levels from the input normalized mean ratio map. The features at the lowest level are sent to the coordinate attention hole spatial pyramid module (CAASP) to extract high-level features, and the features at the specified intermediate levels are sent to the attention adjustment module (AMM) to extract low-level features. The high-level features and low-level features are then fused by the multi-scale feature fusion resampling module (MFFRM) to obtain the landslide extraction result.

[0010] Optionally, the function expression for generating the normalized mean ratio plot in step S102 is:

[0011]

[0012] In the above formula, D mean (i,j) represents the value of pixel (i,j) in the normalized mean ratio map. , , represent the average intensity of pixel (i,j) in the two SAR images before and after the landslide, respectively, and min is the operation of taking the smaller value.

[0013] Optionally, the Coordinate Attention Hollow Spatial Pyramid (CAASP) module includes a Global Average Pooling (GAP) module, an Attention (CA) module, and multiple dilated convolutions with different sampling rates. The features of the last layer output by the backbone network are passed in parallel through multiple dilated convolutions with different sampling rates to capture the contextual information of the image from different receptive fields, resulting in feature maps X∈R. (H×W×C)The obtained feature maps X are then fed into the attention module CA to obtain attention-enhanced feature maps. The feature maps of the last layer output by the backbone network are also simultaneously input into the global average pooling module GAP for downsampling processing to improve the network's generalization ability and prevent overfitting while preserving the spatial structure information of the input image. The attention-enhanced feature maps output by the attention module CA are superimposed with the output feature maps of the global average pooling module GAP to form the output feature maps of the coordinate attention hole spatial pyramid module CAASP, where H, W, and C represent the height, width, and number of channels of feature map X, respectively.

[0014] Optionally, the attention module CA processes the input feature map X as follows:

[0015] S201, For the input feature map X, global average pooling is used to embed the precise spatial location information into each channel for encoding in both width XA and height YAP, resulting in feature maps in both width XA and height YAP directions.

[0016] S202, stitch together the feature maps in the width XA and height YAP directions;

[0017] S203, the concatenated feature maps are reduced in dimensionality using a 1×1 convolution kernel, normalization, and a ReLU activation function to obtain the feature map f∈R. C / r×1×(H+W) , where r is the factor by which the channel is to be reduced in dimensionality, and H, W, and C represent the height, width, and number of channels of the feature map X, respectively;

[0018] S204, decompose the feature map f into two independent tensors f along the spatial dimension. h ∈R C / r×H and f w ∈R C / r×W ;

[0019] S205, respectively, tensor f h and f w Channel attention weights are obtained through a 1×1 two-dimensional convolution and a sigmoid function. The two channel attention weights are then multiplied with the input feature map X to obtain the attention-enhanced feature map.

[0020] Optionally, the multi-scale feature fusion resampling module (MFFRM) fuses high-level and low-level features to obtain landslide extraction results, including:

[0021] S301, the high-level input features are adjusted in number of channels using a 1×1 convolution to become the high-level input features X. H ; to use high-level input features X H Through 1×1 convolution and upsampling, and then combined with low-level features X from the feature input... LChannel compression is performed using 1×1 convolution to obtain feature maps with the same number of channels, which are then concatenated to obtain the concatenated feature X. F ;

[0022] S302, splicing feature X F The semantic flow field is obtained through 3×3 convolution. Where H a-1 and W a-1 Representing semantic flow field Δ a-1 Height and width;

[0023] S303, high-level input features X H and semantic flow field Δ a-1 Perform feature alignment, and then match the aligned features with the low-level feature X. L Adding them together yields feature X DF ;

[0024] S304, feature X DF Feature refinement is performed using a 3×3 convolutional layer, followed by bilinear interpolation and 4x upsampling to obtain the landslide extraction result. The landslide extraction result is a binary image containing only two types of pixel values: the pixel value of the landslide area is 1, and the pixel value of the background area is 0.

[0025] Optionally, step S103, which involves boundary extraction of the landslide extraction result, includes: scanning the landslide extraction result pixel by pixel to determine the outer boundary and hole boundary of the landslide area, wherein the outer boundary is the boundary of the connected region with a pixel value of 1, and the hole boundary is the boundary of the connected region with a pixel value of 0.

[0026] Optionally, the backbone network is a deep residual network (DRN), and the attention adjustment module (AMM) is composed of a channel attention adjustment module (CAMM) and a spatial attention adjustment module (SAMM) cascaded together.

[0027] Optionally, before step S103, a step of training a landslide detection network model is included, and the landslide dataset used to train the landslide detection network model is generated including:

[0028] S401. First, based on the time of the earthquake-landslide occurrence, select optical remote sensing images with few clouds in the landslide-covered area before and after the landslide. Then, acquire the pre-earthquake and post-earthquake SAR images that are closest in time to the two optical remote sensing images. Visually compare the pre-earthquake and post-earthquake optical remote sensing images with the landslide catalog. If it is found that no landslide has occurred in the optical remote sensing images in the landslide catalog, then delete that part of the landslide from the landslide catalog. If it is found that a landslide has occurred outside the landslide catalog, then add that part to the landslide catalog, thereby completing the correction of the landslide catalog.

[0029] S402: Obtain two SAR image samples before and after the landslide. First, perform noise removal and geocoding. Then, generate a normalized mean ratio map sample based on the two SAR image samples before and after the landslide. Map the corrected landslide catalog onto the post-earthquake optical remote sensing image to obtain the landslide catalog mapped on the optical remote sensing image. Match the landslide catalog mapped on the optical remote sensing image with the normalized mean ratio map sample to obtain the normalized mean ratio map sample of the landslide area.

[0030] S403, Upsample the normalized mean ratio map sample of the landslide area to obtain the upsampled normalized mean ratio map sample; Use a sliding window with a preset step size to generate a slice of a specified size from the upsampled normalized mean ratio map sample, and perform data augmentation by mirroring and rotating the slice to obtain the landslide dataset used to train the landslide detection network model.

[0031] Furthermore, the present invention also provides a deep learning-based SAR image seismic landslide boundary detection system, comprising an interconnected microprocessor and a memory, wherein the microprocessor is programmed or configured to execute the deep learning-based SAR image seismic landslide boundary detection method.

[0032] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program that is programmed or configured by a microprocessor to execute the deep learning-based SAR image seismic landslide boundary detection method.

[0033] Compared with existing technologies, the present invention has the following advantages: The present invention includes generating a normalized mean ratio map from two SAR images before and after a landslide, upsampling the map, and then inputting it into a pre-trained deep learning-based landslide detection network model for landslide extraction and boundary extraction to obtain the landslide edge detection result; the landslide detection network model includes an encoder and a decoder. The encoder includes a backbone network and a coordinate attention void space pyramid module, and the decoder includes an attention adjustment module and a multi-scale feature fusion resampling module. The present invention uses a normalized mean ratio map to effectively utilize the information of earthquake landslides in SAR images, and proposes a new deep learning-based landslide detection network model that can achieve fast, high-precision, and automated detection of large-scale earthquake landslide boundaries, and can detect earthquake landslides of different earthquake types and scales. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the research area in an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram illustrating the testing and training process of the landslide detection network model in an embodiment of the present invention.

[0036] Figure 3 This is a schematic diagram of the landslide detection network model in an embodiment of the present invention.

[0037] Figure 4 This is a schematic diagram of the attention module CA in an embodiment of the present invention.

[0038] Figure 5 This is a schematic diagram illustrating the principle of flow field alignment in an embodiment of the present invention.

[0039] Figure 6 This is a schematic diagram illustrating the generation of the training dataset in an embodiment of the present invention.

[0040] Figure 7 This is a comparison of the landslide boundary detection results of different networks for the Mainling landslide in this embodiment of the invention.

[0041] Figure 8 This is a comparison of the landslide boundary detection results of different networks for the Palu landslide in this embodiment of the invention.

[0042] Figure 9 This is a comparison of landslide boundary detection results for PNG landslides using different networks in this embodiment of the invention. Detailed Implementation

[0043] The following text will be based on Figure 1 Three large EQTL (earthquake-induced landslide) areas were studied: the landslide triggered by the 6.9 magnitude earthquake in Milin County, Tibet on November 18, 2017. Figure 1 (a) and (b) in the text), and the landslide triggered by the 7.5 magnitude earthquake in Palu, Central Sulawesi, Indonesia on September 28, 2018. Figure 1 (c) and (d) in the text, and the landslide triggered by the 7.5 magnitude earthquake in PNG on February 26, 2018. Figure 1 (e) and (f) in the middle). Among them: (1) Earthquake landslide in Mainling, Tibet: Mainling County is located in eastern Tibet, China, in the middle reaches of the Yarlung Tsangpo River, between the Nyainqêntanglha Mountains and the Himalayas. It is located between 28°39′ and 29°50′ north latitude and 93°07′ and 95°12′ east longitude, such as Figure 1 As shown, Milin County is long and narrow from east to west, with higher elevations in the west and lower elevations in the east. It features many wide valleys with relatively small differences in altitude. The average elevation of the county is 3700 meters, characterized by mountainous and valley terrain. Due to severe river erosion, weak rock mass stability, and steep terrain slopes, earthquakes easily trigger landslides. On November 18, 2017, a magnitude 6.5 earthquake struck Milin County, triggering 3130 coseismic landslides in the epicenter area, covering a total area of ​​19.72 km². 2The landslides were mainly small to medium-sized, with very few large ones. Of these, 2,912 landslides were located near the fault but far from the epicenter. The smallest landslide area was 81 m². 2 The maximum landslide was 4.35 km long. 2 Area less than 1000m 2 There were 1,526 landslides, accounting for 49%. (2) Earthquake landslide in Palu, Indonesia: Palu, a city in central Sulawesi, Indonesia, is located in a narrow valley facing a long and narrow bay, making it particularly vulnerable to tsunamis and landslides. On September 28, 2018, a magnitude 7.5 earthquake occurred about 78 kilometers north of Palu, the capital of Central Sulawesi province, Indonesia, at 0.178°S latitude and 119.84°E longitude. The earthquake and tsunami damaged more than 65,000 houses and left 60,000 people homeless. The earthquake occurred at approximately 3,800 km. 2 This triggered approximately 7,063 coseismic landslides within the region, with 7,055 concentrated in mountainous areas and 8 occurring in the Palu Basin. According to the landslide catalog, these landslides were all small in size, with over 95% having an area of ​​less than 102 m². 2 The landslide area is mainly composed of rocks. (3) PNG landslide: Papua New Guinea (PNG) is located at the junction of the Pacific Plate, the Indian Ocean Plate and the Australian Plate, and is prone to earthquakes. On February 26, 2018, a Mw 7.5 earthquake occurred, with the epicenter located 92 kilometers southwest of PNG Porgera, at 6.19 degrees south latitude and 142.77 degrees east longitude, and a focal depth of 35 kilometers. The earthquake caused at least 67 deaths, about 300 injuries, and affected about 143,000 people. This earthquake caused about 11,607 landslides near the Hegigio River Basin (Tanyas et al., 2022), of which the main shock triggered 10,469 landslides (total area of ​​145 square kilometers), while the remaining 1,138 landslides (total area of ​​40 square kilometers) were triggered by aftershocks and subsequent rainfall events between February 26 and March 2019. Figure 1In the study, the landslide catalog and optical remote sensing images (a), (c) and (e) are fusion maps of the Mainling, Palu and PNG (Papua New Guinea) landslide catalog and Landsat-8 optical remote sensing images; (b), (d) and (f) are the Sentinel-1 images (boxes) used in the landslide areas and the locations of the three landslide experimental areas ((a), (c) and (e)) on the map, with the pentagrams in the figures indicating the epicenter locations. Based on the usability of the landslide catalog, the several EQTLs studied in this embodiment have the following characteristics: (1) They are landslides triggered by earthquakes of different types and magnitudes. For example, the Mainling earthquake had a magnitude of Mw 6.5, a focal depth of 10 km, and a maximum intensity of VIII, and was mainly a thrust earthquake; the Palu earthquake had a magnitude of Mw 7.5, a focal depth of 10 km, and was a strike-slip earthquake; the PNG earthquake had a magnitude of Mw 7.5, a focal depth of about 35 km, and a maximum intensity of IX, and was mainly a thrust earthquake. (2) These landslides vary greatly in size (the largest area is 4.35 km²). 2 The smallest one is only 40m² 2 (Approximately). While the landslides triggered by the Palu and PNG earthquakes were relatively large, their areas were small; Mainling experienced several large landslides, but the others were mostly small. This embodiment of the method, through the detection of landslides of different earthquake types and scales, better verifies the effectiveness of the method and demonstrates its significant application value in landslide detection.

[0044] like Figure 2 As shown in this embodiment, the SAR image seismic landslide boundary detection method based on deep learning includes:

[0045] S101, Input two SAR images before and after the landslide;

[0046] S102, a normalized mean ratio map is generated based on two SAR images before and after the landslide, and then upsampled;

[0047] S103, the upsampled normalized mean ratio map is input into a pre-trained deep learning-based landslide detection network model (Multi-level Features Effective Weighting and Fusion, or MFEWF network) for landslide extraction, and then the landslide extraction result is subjected to boundary extraction to obtain the landslide edge detection result; the landslide detection network model includes an encoder and a decoder. The encoder includes a backbone network and a coordinate attention hole spatial pyramid module (CAASP), and the decoder includes an attention adjustment module (AMM) and a multi-scale feature fusion resampling module (MFFRM); the backbone network extracts features at multiple levels from the input normalized mean ratio map. The features at the lowest level are sent to the coordinate attention hole spatial pyramid module (CAASP) to extract high-level features, and the features at the specified intermediate levels are sent to the attention adjustment module (AMM) to extract low-level features. The high-level features and low-level features are then fused by the multi-scale feature fusion resampling module (MFFRM) to obtain the landslide extraction result.

[0048] In this embodiment, the function expression for generating the normalized mean ratio plot in step S102 is:

[0049]

[0050] In the above formula, D mean (i,j) represents the value of pixel (i,j) in the normalized mean ratio map. , , represent the average intensity of pixel (i,j) in the two SAR images before and after the landslide, respectively, and min is the operation of taking the smaller value.

[0051] The Coordinate Attention Hollow Space Pyramid (CAASP) module is used to extract features at different levels and scales in landslide areas. For example... Figure 3 As shown, in this embodiment, the Coordinate Attention Spatial Pyramid (CAASP) module includes a Global Average Pooling (GAP) module, an Attention (CA) module, and multiple dilated convolutions with different sampling rates (which can be set as needed; for example, in this embodiment, four sampling rates are used: 1, 6, 12, and 18, with kernel sizes of 1×1 and 3×3). The features of the last layer output by the backbone network are passed in parallel through multiple dilated convolutions with different sampling rates to capture the contextual information of the image from different receptive fields, resulting in feature maps X∈R. (H×W×C)The obtained feature maps X are then fed into the attention module CA to obtain attention-enhanced feature maps. The feature maps of the last layer output by the backbone network are also simultaneously input into the global average pooling module GAP for downsampling processing to improve the network's generalization ability and prevent overfitting while preserving the spatial structure information of the input image. The attention-enhanced feature maps output by the attention module CA are superimposed with the output feature maps of the global average pooling module GAP to form the output feature maps of the coordinate attention hole spatial pyramid module CAASP, where H, W, and C represent the height, width, and number of channels of feature map X, respectively.

[0052] The attention module (CA) is used to capture the channel relationships and location information of the target, improving the performance of target feature extraction. For example... Figure 4 As shown, the attention module CA processes the input feature map X as follows:

[0053] S201, For the input feature map X, global average pooling is used to embed the precise spatial location information into each channel for encoding in both width XA and height YAP, resulting in feature maps in both width XA and height YAP directions.

[0054] S202, stitch together the feature maps in the width XA and height YAP directions;

[0055] S203, the concatenated feature maps are reduced in dimensionality using a 1×1 convolution kernel, normalization, and a ReLU activation function to obtain the feature map f∈R. C / r×1×(H+W) , where r is the factor by which the channel is to be reduced in dimensionality, and H, W, and C represent the height, width, and number of channels of the feature map X, respectively;

[0056] S204, decompose the feature map f into two independent tensors f along the spatial dimension. h ∈R C / r×H and f w ∈R C / r×W ;

[0057] S205, respectively, tensor f h and f w Channel attention weights are obtained by using a 1×1 two-dimensional convolution and a sigmoid function. The two channel attention weights are then multiplied with the input feature map X to obtain the attention-enhanced feature map, thereby achieving the selection of effective features and suppressing redundant features.

[0058] This embodiment proposes a multi-scale feature fusion resampling module (MFFRM) to process high-level and low-level features differently, enabling better fusion of deep semantic information and shallow detail information, thus achieving full extraction of landslide features. The MFFRM module fully fuses the high-level features output from the encoding network with the low-level features processed by the feature input attention adjustment module (AMM), ultimately producing a landslide edge detection result of the same size as the input. Figure 3 As shown, in this embodiment, the multi-scale feature fusion resampling module (MFFRM) fuses high-level features and low-level features to obtain the landslide extraction result, including:

[0059] S301, the high-level input features are adjusted in number of channels using a 1×1 convolution to become the high-level input features X. H ; to use high-level input features X H Through 1×1 convolution and upsampling, and then combined with low-level features X from the feature input... L Channel compression is performed using 1×1 convolution to obtain feature maps with the same number of channels, which are then concatenated to obtain the concatenated feature X. F ;

[0060] S302, splicing feature X F The semantic flow field is obtained through 3×3 convolution. Where H a-1 and W a-1 Representing semantic flow field Δ a-1 Height and width;

[0061] S303, high-level input features X H and semantic flow field Δ a-1 Perform feature alignment, and then match the aligned features with the low-level feature X. L Adding them together yields feature X DF ;

[0062] S304, feature X DF Feature refinement is performed using a 3×3 convolutional layer, followed by bilinear interpolation and 4x upsampling to obtain the landslide extraction result. The landslide extraction result is a binary image containing only two types of pixel values: the pixel value of the landslide area is 1, and the pixel value of the background area is 0.

[0063] In step S303, the high-level input feature X is... H and semantic flow field Δ a-1 Feature alignment is achieved by performing a warp operation, which is an existing method, such as... Figure 5 As shown, due to the high-level input feature X H and semantic flow field Δa-1 There are resolution differences, according to, for example Figure 5 High-level input features X H Each pixel in Figure 5 The value in the middle (represented as offset) is mapped to the low-level feature X according to the following formula. L Pixels in:

[0064] p a = a-1 +Δ a-1 (p a-1 ) / 2,

[0065] In the above formula, p a The low-level features X obtained by mapping L The pixels in p a-1 For high-level input features X H The pixels in Δ a-1 (p a-1 ) represents the semantic flow field Δ a-1 The corresponding pixels; then, using a differentiable bilinear sampling mechanism, the data mapped to the low-level feature X are processed. L The aligned features are obtained by linear interpolation of the values ​​of the four neighborhoods (top left, top right, bottom left, bottom right) of a pixel. Figure 5 (represented as warp in Chinese).

[0066] In this embodiment, step S103, boundary extraction of the landslide extraction results, includes: scanning the landslide extraction results pixel by pixel to determine the outer boundary and hole boundary of the landslide area. The outer boundary is the boundary of a connected region with a pixel value of 1, and the hole boundary is the boundary of a connected region with a pixel value of 0. A sliding window test is performed on the test image using a trained model to obtain the corresponding binary result image. The binary image contains only two types of pixel values: the target of interest in the study, i.e., the landslide area, has a pixel value of 1, and the background pixel value is 0. To more intuitively analyze the extraction effect of different networks on the landslide area, we perform topological analysis on the binary labels and result image to determine the surrounding relationship of the binary image boundaries, i.e., to determine the outer boundary, hole boundary, and their hierarchical relationship. Since these boundaries have a one-to-one correspondence with the original image (the outer boundary corresponds to a connected region with a pixel value of 1, and the hole boundary corresponds to a connected region with a pixel value of 0), this embodiment extracts the landslide area boundary based on this and simultaneously plots the two boundary images on the difference map of the test image for comparative analysis. Boundary extraction involves scanning each image pixel-by-pixel to determine the outer boundary of the landslide area and the boundary of the borehole. The specific steps are as follows: Assume f(i,j) represents the pixel value in the i-th row and j-th column of the image. When the boundary extraction algorithm scans rows, f(i,j) = 1 and f(i,j-1) = 0 are the starting points of the outer boundary, while f(i,j) ≥ 1 and f(i,j+1) = 0 are the starting points of the borehole boundary. Starting from the starting point, pixels on the boundary are marked, and an identifier "Boundary Sequence Number (NBD)" is defined for each newly discovered boundary. Initially, NBD = 1, and NBD is incremented by 1 for each new boundary discovered. During this process, when encountering f(i,j) = 1 and f(i,j+1) = 0, f(i,j) is set to -NBD, which is the ending point of the rightmost boundary. Based on this, each pixel is scanned cyclically until the entire image is traversed.

[0067] like Figure 3As shown, in this embodiment, the backbone network is a Deep Residual Network (DRN) to extract richer features. DRN is an existing feature extraction backbone network based on ResNet (He et al., 2016). It replaces the max-pooling layers in ResNet with convolutional layers, reducing subsequent gridding effects. Furthermore, DRN improves ResNet's 4th and 5th convolutional layers using dilated convolutions and adds two dilated convolutional blocks with different dilation rates at the back end of the network to further reduce gridding. This maintains the original spatial resolution of the image while ensuring the resolution of the receptive fields of subsequent neurons, thus improving the feature resolution of the model's output. Therefore, DRN maintains better resolution than ResNet, which is beneficial for improving semantic segmentation accuracy. To fully utilize the different layer features of the Deep Residual Network (DRN) backbone, this embodiment uses the output feature map of a specified intermediate convolutional layer (the 3rd convolutional layer) of the DRN as a low-level feature input to the Attention Adjustment Module (AMM), and the output feature map of the last convolutional layer (the 8th convolutional layer) as a high-level feature input to the Coordinate Attention Hole Spatial Pyramid Module (CAASP).

[0068] like Figure 3 As shown, in this embodiment, the attention modulation module (AMM) consists of a cascaded channel attention modulation module (CAMM) and a spatial attention modulation module (SAMM). Both the channel attention modulation module (CAMM) and the spatial attention modulation module (SAMM) are existing network modules, so their detailed implementation will not be described here. The attention modulation module (AMM) is used to enhance secondary features and suppress the most sensitive and least sensitive features. This function emphasizes directly extracting easily overlooked regions through the less important regions, which is crucial for segmentation tasks. Input features first enter the channel attention modulation module (CAMM), where average pooling and convolutional layers establish inter-channel dependencies, reflecting the sensitivity to information features. To establish spatial relationships in the spatial dimension, the spatial attention modulation module (SAMM) is introduced and cascaded after the channel attention modulation module (CAMM).

[0069] For deep learning networks, dataset quality is crucial, directly impacting the network's effectiveness in extracting landslide features. If the dataset quality is poor, even the best network will fail to extract valid landslide features. For example... Figure 6As shown, the constructed high-quality earthquake landslide dataset mainly consists of three parts: landslide catalog correction, SAR image preprocessing, and dataset generation. Before step S103 in this embodiment, there is also a step of training a landslide detection network model, and the generation of the landslide dataset used to train the landslide detection network model includes:

[0070] S401. First, based on the time of the earthquake-landslide occurrence, select lightly cloudy optical remote sensing images (Landset-8 or Sentinel-2) of the landslide-covered area before and after the landslide. Then, acquire pre-earthquake and post-earthquake SAR images (such as SAR images from the Sentinel-1 system) that are closest in time to the two optical remote sensing images. Visually compare the pre-earthquake and post-earthquake optical remote sensing images with the landslide catalog. If no landslide occurs in the optical remote sensing images within the landslide catalog, remove that portion of the landslide from the catalog. If a landslide occurs outside the landslide catalog, add that portion to the catalog, thus completing the correction of the landslide catalog. Although the correction of the landslide catalog is very time-consuming, it is extremely necessary to ensure the accuracy of the landslide area in the corresponding SAR images, allowing the deep learning network to learn the effective features of the landslide more effectively. After the landslide catalog is corrected, it can be mapped onto the post-earthquake optical remote sensing images.

[0071] S402: Obtain two SAR image samples before and after the landslide. First, perform noise removal and geocoding. Then, generate a normalized mean ratio map sample based on the two SAR image samples before and after the landslide. Map the corrected landslide catalog onto the post-earthquake optical remote sensing image to obtain the landslide catalog mapped on the optical remote sensing image. Match the landslide catalog mapped on the optical remote sensing image with the normalized mean ratio map sample to obtain the normalized mean ratio map sample of the landslide area.

[0072] After an earthquake and landslide, the surface covered by fresh soil, landslide material, or debris flow exhibits diffuse reflection characteristics, resulting in a much higher echo power received by sensors compared to before the disaster. Therefore, in the dataset generation section, a normalized mean ratio map is first generated from the pre-earthquake and post-earthquake SAR images.

[0073] S403, Upsample the normalized mean ratio map sample of the landslide area to obtain the upsampled normalized mean ratio map sample; Use a sliding window with a preset step size to generate a slice of a specified size from the upsampled normalized mean ratio map sample, and perform data augmentation by mirroring and rotating the slice to obtain the landslide dataset used to train the landslide detection network model.

[0074] In this embodiment, the mean ratio map of the landslide area is obtained by corresponding the landslide catalog mapped on the optical remote sensing image. Due to the low resolution of Sentinel-1 (5m × 20m), many small landslide areas contain too few pixels (some landslide areas have only a few pixels), which cannot provide sufficient information for landslide feature extraction, thus severely affecting the effective extraction of landslide features. Therefore, this paper uses bilinear interpolation upsampling to supplement information. For landslide areas in Mainling and PNG, which have relatively large scales in China, a 3x upsampling was used; since the landslides in the Palu area are very small, a 6x upsampling was used. The upsampled images are then processed by a sliding window with a step size of 150 pixels to generate 512×512 pixel slices. The sample slices containing landslide areas are then horizontally and vertically mirrored, and rotated ±90° and 180° for data augmentation, generating a landslide dataset for training the landslide detection network model.

[0075] In this embodiment, a normalized mean ratio map is first generated from the pre-earthquake and post-earthquake SAR images, and then segmented. This divides a large SAR image sample into multiple smaller SAR image samples, reducing the overall sample data volume and improving the efficiency of network training. Simultaneously, this embodiment utilizes the landslide catalog mapped on the optical remote sensing image to correspond with the mean ratio map, thus obtaining the mean ratio map of the landslide area. Due to the low resolution of Sentinel-1 (5m × 20m), many small landslide areas contain too few pixels (some landslide areas contain only a few pixels), which cannot provide sufficient information for landslide feature extraction, severely affecting the effective extraction of landslide features. Therefore, this embodiment uses bilinear interpolation upsampling to supplement information. For the relatively larger landslide areas in Mainling and PNG in China, a 3x upsampling is used; since the landslides in the Palu area are very small, a 6x upsampling is used. The upsampled image was processed by a sliding window with a step size of 150 pixels to generate 512×512 pixel slices. The sample slices with landslide areas were horizontally and vertically mirrored, and rotated ±90° and 180° to perform data augmentation, resulting in the final dataset used for training.

[0076] The following section will present an experimental verification of the deep learning-based SAR image seismic landslide boundary detection method.

[0077] The experimental software environment in this embodiment is: PyTorch 1.40, CUDA 10.2, and Python 3.7; the hardware environment is: Intel Xeon Gold 5120 CPU and a single NVIDIA RTX 2080Ti GPU. During network training, the learning rate is set to 0.005, the batch size is 4, and the network training iterations are performed for 100 epochs. The best training result is then saved.

[0078] To better evaluate algorithm performance, this embodiment selects pixel accuracy (PA) and intersection over union (IoU) as evaluation metrics (Cai et al., 2021). PA represents the proportion of correctly classified pixels out of the total pixels, and IoU represents the intersection over union ratio between the segmentation result and the true result, specifically defined as follows:

[0079]

[0080]

[0081] In the above formula, j represents the target category, and there are j+1 categories in a single image (including one background category); P ij This represents the number of pixels that are actually of type i but are predicted as type j; similarly, P... ji This represents the number of pixels that are actually of type j but are predicted as type i; P ii This represents the actual number of pixels in class i.

[0082] To verify the method proposed in this embodiment, landslide boundary detection experiments were conducted in three different types of large-scale earthquake landslide areas: Mainling, PNG, and Palu. The landslide detection network model (MFEWF network) proposed in this embodiment was compared and analyzed with existing DeepLabV3+, MF2AM, and GCAM networks. MF2AM and GCAM are network improvements based on DeepLabV3+, and have achieved good results in SAR image overlay extraction and airport runway detection. This embodiment first uses the mean ratio map generated from two Sentinel-1 images before and after the landslide to create a dataset. Data augmentation of the landslide area was performed using translation, horizontal and vertical mirroring, and vertical rotation (±90° and 180°). The augmented dataset contains 17,947 500×500 samples. When creating the dataset, four regions were reserved for independent testing and accuracy evaluation in these three large-scale landslide areas, while the remaining regions were used for dataset training.

[0083] Figure 7 The results of landslide boundary detection using different networks are presented. Figure 7 (a) is a fusion of the Landsat-8 imagery of the Mainling area and its landslide catalog, where the yellow box represents the independent landslide test area, which includes landslides of different sizes. Figure 7 (b) in the figure shows line graphs of landslide detection indicators for the four networks within independent test areas. Figure 7As can be seen in (b), the PA and IoU values ​​of the three networks DeepLabV3+, MF2AM and GCAM are not much different, with IoU around 0.72; while the PA and IoU values ​​of the MFEWF network reach 88.8% and 81.7% respectively, which is about 9% higher than DeepLabV3+. Figure 7 (c) in Figure 7 shows the fusion of landslide boundary detection results and landslide catalog with mean ratio map of the independent test area in (a) of this embodiment using the MFEWF network. Figure 4 The fusion diagram of landslide boundary detection results and landslide catalog vs. mean ratio map for each network is shown in Appendix I. Figure 7 As shown in (c), the MFEWF network proposed in this embodiment can detect landslide boundaries well in the independent test area, and its consistency with the landslide catalog is good. To gain a clearer understanding of the landslide detection performance and detailed information of each network, Figure 7 The (d)-(g) and (h)-(k) in the figure give the following values ​​respectively. Figure 7 The two small areas (A) and (B) in (c) are shown. Figure 7 As shown in (d)-(g), using the landslide boundary detection method proposed in this embodiment, all four networks can detect these small landslides; however, for the small landslide in the upper right corner, DeepLabV3+ and GCAM both missed some detections, while MF2AM and MFEWF detected them better. Overall, the MFEWF network in this embodiment can detect the landslides in the area relatively completely, and the landslide boundary positioning is relatively accurate, with a good match with the landslide catalog (such as the landslide areas pointed to by the green arrows); the other networks have larger boundary positioning errors when detecting landslides, resulting in some false alarms and missed detections, especially DeepLabV3+ and GCAM networks, which missed more detections. Figure 7 As shown in (h)-(k), the landslides in this area vary considerably in size and have complex boundaries. DeepLabV3+, GCAM, and MF2AM all showed significant missed detections of some smaller landslides in this area, and their detection of other landslide areas exhibited inaccurate edge localization issues, resulting in significant deviations from the landslide catalog. In contrast, the MFEWF network showed almost no missed detections, and its landslide boundary detection had a much higher degree of consistency with the landslide catalog than other networks, especially for the landslides indicated by the green arrows, where the boundary localization was clearly superior. This demonstrates the effectiveness of the AMM module in extracting low-level features and detailed information from landslides.

[0084] like Figure 8(a) is a fusion of Landsat-8 imagery and landslide catalog in the Palu region (this image is an optical image after 6x upsampling). The area within the yellow box is an independent test area. As can be seen from the image, the Palu region is mostly composed of very small-scale landslide areas, which places higher demands on the network's ability to extract detailed information. Figure 8 Figure (b) shows the landslide detection accuracy of the four networks. As can be seen from the figure, DeepLabV3+ has the lowest accuracy, with more false alarms and missed alarms. The accuracy of MF2AM and GCAM networks is not much different, and both have an accuracy improvement of more than 10% compared to DeepLabV3+. The landslide detection accuracy of the network in this embodiment is the highest, which is about 17% higher than DeepLabV3+, and the false alarms and missed alarms are greatly reduced. Figure 8 (c) shows the landslide edge detection results and landslide catalog of the MFEWF network. Figure 8 (a) A fusion plot of the mean ratio plots of the independent test regions. Figure 8 (d1)-(d4) and (e1)-(e4) are four network pairs. Figure 8 Enlarged views of Area D and Area E in (c) (Appendix II shows the four networks for...) Figure 8 (a) A fused image of landslide boundary detection results for an independent test area. Figure 8 As shown in (d1)-(d4) and (e1)-(e4), these are small landslides within Area D and Area E, relatively scattered, and with fairly clear boundaries. The DeepLabV3+ network failed to detect several small landslides in these two areas and exhibited poor detection completeness for several others. MF2AM and GCAM both detected landslides in these areas, but their landslide boundary localization errors were significant, resulting in numerous false alarms and missed detections. The MFEWF network in this embodiment significantly improved the accuracy of landslide boundary localization in these two areas compared to the other three networks, showing excellent agreement with the landslide catalog, especially demonstrating a clear advantage in detecting the landslide areas indicated by the green arrows. For these small-scale landslides in the Palu region, detailed information is crucial, and these results fully demonstrate the network's ability to grasp detailed landslide information and extract effective features.

[0085] Figure 9 (a) in the image is a fusion of Landsat-8 imagery and landslide catalog in the PNG region, which includes two independent test areas (the areas within the yellow boxes). It is clear that the landslide scene in this region is relatively complex, with varying scales and a chaotic distribution, which places high demands on the network's global attention capability and multi-scale extraction capability. Figure 9(b) is a line graph showing the mean values ​​of landslide extraction indices for the four networks in two independent test areas. It can be seen that GCAM has the lowest IoU and PA in this area, while MF2AM and DeepLabV3+ have similar IoU values. The network in this embodiment has the highest IoU and PA values, with an improvement of nearly 10% compared to DeepLabV3+. Figure 9 (c) and (d) in the figure represent the MFWEF network pairs respectively. Figure 9 The image shows a fusion plot of the landslide boundary detection results and mean ratio of independent test areas (A) and (B) in (a), where the purple line represents the landslide detection results and the blue line represents the landslide catalog. Figure 9 (c1)-(c4) and (c5)-(c8) are four network pairs. Figure 9 A fusion diagram of landslide detection results for regions (A1) and (A2) in (c) of the image; Figure 9 (d1)-(d4) and (d5)-(d8) are four network pairs. Figure 9 The image shows a fusion of landslide detection results from two regions (B1) and (B2) in (d). Figure 9 As shown in (c1)-(c4), there is a very large landslide in this area, which was detected by all four networks. The MFEWF network detected the landslide boundary better, but the landslide boundary detection of the other three networks differed significantly from the landslide catalog. In particular, the GCAM network clearly missed a portion of the area of ​​this large landslide. Figure 9 (c2)). Additionally, all four networks exhibited missed detections of several very small landslides, with DeepLabV3+ and GCAM networks showing particularly high numbers of missed detections. Figure 9 In (c5)-(c8), it was found that the landslides in this small area were of similar size and relatively scattered. All four networks detected these landslides (the GCAM network missed a small landslide on the right). Figure 9 (c6))). According to Figure 9 In the analysis of (d1)-(d8), it was found that DeepLabV3+, GCAM, and MF2AM exhibited some missed detections of small-scale landslides and inaccurate landslide boundary positioning, while the MFEWF network showed virtually no missed detections and relatively accurate boundary positioning. Figure 7 (c1)-(c8) and Landslide area Overall, the landslide detection results (d1)-(d8) in this embodiment show that the proposed network MFEWF can detect landslides of different scales and is relatively accurate in locating the boundaries of landslides of different shapes and scales, which has obvious advantages over other networks (such as the boundary of the landslide area pointed to by the green arrow).

[0086] To more intuitively evaluate the landslide boundary detection performance of the network proposed in this embodiment, the paper compares the accuracy of the four networks in three earthquake landslide zones with different earthquake types and scales, as shown in Table 1:

[0087] Table 1. Landslide detection results of different networks in different landslide areas.

[0088]

[0089] Depend on Mainling As can be seen, in the Mainling region, apart from two relatively large landslides, the others are all relatively small landslides. Table 1 shows that the landslide detection accuracy (PA) of all four networks reached 80%, but the IoU values ​​of DeepLabV3+, MF2AM, and GCAM were lower. Experimental results indicate that this is because these networks produced significant deviations in landslide boundary detection, resulting in more false alarms. In contrast, the IoU of the network in this embodiment is nearly 10% higher than the other three networks, indicating that its landslide edge localization is more accurate, its match with the landslide catalog is much better, and it produces fewer false alarms. In the independent test area of ​​PNG, GCAM had the lowest PA and IoU values, indicating that this network performs poorly in detecting landslides in complex scenes with varying scales. In contrast, the network MFEWF proposed in this embodiment, when fusing multi-scale features, assigns greater weight to feature maps of appropriate scales based on the target scale of interest through different attention mechanisms. Its average PA value reached 82.7% and IoU reached 74.5% in both regions, indicating that it has better performance in extracting landslide boundaries in complex scenes. In the detection of small-scale earthquake-induced landslides in the Palu region, DeepLabV3+ performed the worst, which may be related to its simple concatenation and fusion of low-level feature outputs with high-level features. MFEWF, on the other hand, uses AMM to filter low-level features and MFFRM to adaptively assign different weights to low-level and high-level features according to the target information during high-low feature fusion, effectively improving the boundary detection performance of small-scale landslides. For independent tests in three different earthquake-type landslide areas, MFEWF showed a 10% improvement in detection accuracy and a significant 8.7% improvement in IoU compared to DeepLabV3+. It also showed nearly 10% and 4% improvements in accuracy for GCAM and MF2AM, respectively, and improvements in IoU of 8.7% and 7.5%.

[0090] The varying landslide detection performance of different networks is largely related to their internal structure. DeepLabV3+ simply fuses low-level and high-level features, which is not fully adaptable to extracting landslide areas with complex scenes and small scales; for example, Palu's landslide detection performance is poor. GCAM is an improved airport runway extraction network based on DeepLabV3+. Airport boundaries are relatively flat and regular, while landslide area boundaries are complex. This may explain GCAM's poor performance in landslide boundary extraction, especially for landslides with large scale differences, such as PNGs. Although MF2AM fuses intermediate-level features, which is beneficial for effective feature fusion in subsequent decoding processes, it pays less attention to low-level features. Low-level features contain a lot of boundary details and play a crucial role in landslide feature extraction. Therefore, MF2AM cannot achieve satisfactory results in landslide boundary localization. The network proposed in this embodiment exhibits very few missed detections in small-scale landslide detection and demonstrates superior performance in landslide boundary localization. This is attributed to the fact that the network first uses a DRN backbone network to better extract low-level and high-level features at different levels; then it utilizes an AMM module to effectively filter low-level landslide features; furthermore, it uses a constructed CAASP module to better extract high-level landslide features; and finally, it uses the proposed MFFRM module to effectively suppress feature misalignment during the fusion process by adaptively assigning different weights to low-level and high-level features according to landslide information, thereby significantly improving the localization capability of landslide boundaries at different scales. Therefore, the MFEWF network proposed in this embodiment can better extract features from earthquake landslide areas, suppress complex background noise, achieve more accurate landslide boundary localization, have the best consistency with the landslide catalog, and is unaffected by landslides of different earthquake types, thus achieving high-precision and effective detection of earthquake landslide boundaries.

[0091] To better verify the effectiveness of the landslide detection network proposed in this embodiment, an ablation experiment was conducted. The training and independent test data were the same as those in the experimental section of Section 4. The average landslide detection accuracy of each network in the four independent test areas is shown in Table 2.

[0092] Table 2 shows the ablation experiments of the MFEWF network presented in this embodiment.

[0093]

[0094] Table 2 shows that the average accuracy of detecting four independent landslide test areas using the original DeepLabV3+ network reaches 72.9%, and the IoU reaches 65.9%. Although landslides can be detected, there are many false alarms and missed detections, which does not achieve satisfactory results. In the original DeepLabV3+, replacing the backbone network from ResNet101 with DRN improves the accuracy by 3.3%, indicating that the backbone performs better in extracting features at different levels of landslides. In the original DeepLabV3+, replacing only the CAASP module proposed in this embodiment improved the accuracy by 4.2%, indicating that this module can more effectively extract high-level features of landslides. Replacing only the AMM module improved the landslide boundary extraction accuracy by 5.4%, indicating that this module can better extract the detailed features of landslides. Replacing only the MFFRM module proposed in this embodiment also improved the accuracy by 5.4%, indicating that this module can better fuse high-level and low-level features of landslides, thereby improving the performance of landslide boundary extraction. After replacing all four modules (i.e., the MFEWF network proposed in this embodiment), the landslide boundary extraction accuracy improved by 10.1%, and the IoU also improved by 8.7%. Therefore, the MFEWF network proposed in this embodiment is very effective. Using the MFEWF network proposed in this embodiment, landslide areas can be extracted relatively effectively even with existing classic deep learning networks (DeepLabV3+). Furthermore, using the network (MFEWF network) proposed in this embodiment, the extraction and fusion performance of high-level and low-level landslide features is greatly improved, resulting in a significant improvement in the extraction accuracy of seismic landslide boundaries in SAR images and achieving satisfactory results.

[0095] The results of the MFEWF network's ability to detect landslides induced by earthquakes of different sizes, as presented in this embodiment, are shown in Table 3.

[0096] Table 3 shows the detectable landslide area in three earthquake landslide cases.

[0097] Palu PNG DeepLabV3+ GCAM minimum (m 2 )]]> 2,733 706 1,726 max(m 2 )]]> 12,522,671 1,065,553 14,060,991

[0098] As shown in Table 3, the minimum landslide area detected in the Mainling test area was 2733 m². 2 Only two landslides went undetected due to their small size. For the independent test areas of Palu and PNG, the method in this embodiment detected landslide areas of 706 m² each. 2 and 1726m 2 The smallest landslide was detected. This demonstrates that the MFEWF network proposed in this embodiment has excellent ability to detect earthquake-induced landslide boundaries. Furthermore, the Sentinel-1 SAR image used in this embodiment is only of medium resolution (approximately 5m × 20m in distance and azimuth). If a higher resolution SAR image is used, the method in this embodiment can detect even smaller landslides.

[0099] In addition, this embodiment also statistically analyzes the landslide boundary detection speed using different deep learning networks, as shown in Table 4.

[0100] Table 4. Landslide detection speeds for different networks.

[0101] MF2AM MFEWF Mainling Palu PNG 1.75s 1.98s 1.84s 2.06s Figure 1 1.71s 1.93s 1.99s 2.17s Figure 1 4.27s 4.82s 4.46s 4.97s

[0102] For the Mainling (1675×1406 pixels) and Palu regions (2544×2578 pixels), the DeepLabV3+ network requires only about 1.7 seconds, while the other three networks require about 2 seconds. The detection time for earthquake-induced landslides in PNGs with two independent test regions (1778×2752 pixels and 1474×2384 pixels) is also very short (<5 seconds), indicating that these networks can detect earthquake-induced landslides quickly. Although the proposed MFEWF network is slightly slower than DeepLabV3+, it achieves higher accuracy in detecting earthquake-induced landslides.

[0103] In this embodiment, Sentinel-1 SAR images can be used to detect earthquake-induced landslides. However, due to their medium resolution (5m × 20m), small landslides contain very few pixels (some contain only a few pixels), severely affecting the effective extraction of landslide features by deep learning networks. Therefore, we upsample the SAR data to improve its resolution, using different upsampling rates depending on the size of most earthquake-induced landslide areas.

[0104] To understand the impact of upsampling rate on earthquake-induced landslide detection, we evaluated landslide detection performance at different upsampling rates. As shown in Table 5:

[0105] Table 5 shows the accuracy of earthquake-induced landslide boundary detection corresponding to different adoption rates.

[0106]

[0107] As shown in Table 5, for Palu, the detection accuracy of PA and IoU values ​​is extremely low (<2%) without upsampling, and increases to approximately 16% with a 2x upsampling. Detection performance improves with increasing upsampling rate, but the rate of improvement slows down when the upsampling rate exceeds 4. Therefore, this embodiment uses a 6x upsampling for Palu, ensuring approximately 80% detection accuracy. For Mainling and PNG, the IoU values ​​are very low without upsampling (32.6% and 24.9%, respectively), meaning that despite high PA values ​​(94.2% and 69.4%, respectively), many background targets are falsely detected as landslides. False alarm rates are significantly reduced (i.e., IoU values ​​increase significantly) when the upsampling rate is set to 2 and 3; however, detection performance does not change much when the upsampling rate exceeds 3. Therefore, we use a 3x upsampling for Mainling and PNG, both with satisfactory detection accuracy (approximately 80%). In summary, when using Sentinel-1 SAR data based on the area of ​​earthquake-induced landslides, it is recommended to perform 3 to 6 upsampling operations.

[0108] In summary, the deep learning-based SAR image seismic landslide boundary detection method in this embodiment provides a high-quality and reliable landslide dataset for the deep learning network at the source. The proposed MFEWF network then generates high-precision landslide detection results, and finally, the boundary is extracted, achieving high-precision extraction and localization of EQTL boundaries. Experiments involving landslides induced by earthquakes of different types and sizes in Mainling, China, Palu, and PNG, Indonesia, demonstrate that the proposed landslide detection framework achieves good landslide boundary detection. In particular, the proposed MFEWF landslide detection network exhibits significant advantages in landslide boundary localization, achieving a detection accuracy of 83% and an IoU of approximately 75%. This landslide detection framework is the first to combine deep learning with SAR image seismic landslide edge detection, promoting the development of deep learning in the field of SAR image landslide detection and encouraging more scholars to conduct in-depth research on the application of deep learning in SAR image landslide detection, ultimately achieving high-precision, rapid, and automatic detection of landslide boundaries in large-scale SAR images.

[0109] Furthermore, this embodiment also provides a deep learning-based SAR image seismic landslide boundary detection system, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the deep learning-based SAR image seismic landslide boundary detection method.

[0110] Furthermore, this embodiment also provides a computer-readable storage medium storing a computer program, which is used to be programmed or configured by a microprocessor to execute the deep learning-based SAR image seismic landslide boundary detection method.

[0111] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. ​ One or more processes and / or boxes ​ The steps of the function specified in one or more boxes.

[0112] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting seismic landslide boundaries based on SAR images using deep learning, characterized in that, include: S101, Input two SAR images before and after the landslide; S102, a normalized mean ratio map is generated based on two SAR images before and after the landslide, and then upsampled; S103, the upsampled normalized mean ratio map is input into a pre-trained deep learning-based landslide detection network model for landslide extraction, and then the landslide extraction result is subjected to boundary extraction to obtain the landslide edge detection result; the landslide detection network model includes an encoder and a decoder. The encoder includes a backbone network and a coordinate attention hole spatial pyramid module (CAASP), and the decoder includes an attention adjustment module (AMM) and a multi-scale feature fusion resampling module (MFFRM); the backbone network extracts features at multiple levels from the input normalized mean ratio map. The features at the lowest level are sent to the coordinate attention hole spatial pyramid module (CAASP) to extract high-level features, and the features at the specified intermediate levels are sent to the attention adjustment module (AMM) to extract low-level features. The high-level features and low-level features are then fused by the multi-scale feature fusion resampling module (MFFRM) to obtain the landslide extraction result; The function expression for generating the normalized mean ratio plot in step S102 is: ， In the above formula, Pixels of the normalized mean ratio plot The value, , The pixels in the two SAR images before and after the landslide are respectively ( i , j The average strength of ) To select the smaller value; The Coordinate Attention Hollow Spatial Pyramid (CAASP) module includes a Global Average Pooling (GAP) module, an Attention (CA) module, and multiple dilated convolutions with different sampling rates. The features of the final layer output by the backbone network are passed in parallel through multiple dilated convolutions with different sampling rates to capture contextual information of the image from different receptive fields, resulting in feature maps X∈R. (H×W×C) The obtained feature maps X are then fed into the attention module CA to obtain attention-enhanced feature maps. The feature maps of the last layer output by the backbone network are also simultaneously input into the global average pooling module GAP for downsampling processing to improve the network's generalization ability and prevent overfitting while preserving the spatial structure information of the input image. The attention-enhanced feature maps output by the attention module CA are superimposed with the output feature maps of the global average pooling module GAP to form the output feature maps of the coordinate attention hole spatial pyramid module CAASP, where H, W, and C represent the height, width, and number of channels of feature map X, respectively. The multi-scale feature fusion resampling module (MFFRM) fuses high-level and low-level features to obtain landslide extraction results, including: S301 uses 1×1 convolution to adjust the number of channels of the input high-level features as high-level input features. ; High-level input features Through 1×1 convolution and upsampling, and then combined with low-level features from the feature input... Channel compression is performed using 1×1 convolution to obtain feature maps with the same number of channels, which are then concatenated to obtain the concatenated feature map. ; S302, splicing features The semantic flow field is obtained through 3×3 convolution. ,in and Representing semantic flow fields respectively Height and width; S303, advanced input features and semantic flow field Perform feature alignment, and then match the aligned features with the lower-level features. Adding together yields the features ; S304, features Feature refinement is performed using a 3×3 convolutional layer, followed by bilinear interpolation and 4x upsampling to obtain the landslide extraction result. The landslide extraction result is a binary image containing only two types of pixel values: the pixel value of the landslide area is 1, and the pixel value of the background area is 0.

2. The SAR image seismic landslide boundary detection method based on deep learning according to claim 1, characterized in that, The attention module CA processes the input feature map X as follows: S201, For the input feature map X, global average pooling is used to embed the precise spatial location information into each channel for encoding in both width XA and height YAP, resulting in feature maps in both width XA and height YAP directions. S202, stitch together the feature maps in the width XA and height YAP directions; S203, the concatenated feature maps are reduced in dimensionality using a 1×1 convolution kernel, normalization, and a ReLU activation function to obtain the final feature map. f ∈R C / r×1×(H+W) , where r is the factor by which the channel is to be reduced in dimensionality, and H, W, and C represent the height, width, and number of channels of the feature map X, respectively; S204, feature map f Decomposed into two independent tensors along the spatial dimension f h ∈R C / r×H and f w ∈R C / r×W ; S205, respectively tensors f h and f w Channel attention weights are obtained through a 1×1 two-dimensional convolution and a sigmoid function. The two channel attention weights are then multiplied with the input feature map X to obtain the attention-enhanced feature map.

3. The SAR image seismic landslide boundary detection method based on deep learning according to claim 1, characterized in that, Step S103, which involves boundary extraction of the landslide extraction results, includes: scanning the landslide extraction results pixel by pixel to determine the outer boundary and hole boundary of the landslide area. The outer boundary is the boundary of the connected region with a pixel value of 1, and the hole boundary is the boundary of the connected region with a pixel value of 0.

4. The SAR image seismic landslide boundary detection method based on deep learning according to claim 1, characterized in that, The backbone network is a deep residual network (DRN), and the attention adjustment module (AMM) is composed of a channel attention adjustment module (CAMM) and a spatial attention adjustment module (SAMM) connected in cascade.

5. The SAR image seismic landslide boundary detection method based on deep learning according to claim 1, characterized in that, Before step S103, there is also a step of training a landslide detection network model, and the landslide dataset used to train the landslide detection network model includes: S401. First, based on the time of the earthquake-landslide occurrence, select optical remote sensing images with few clouds in the landslide-covered area before and after the landslide. Then, acquire the pre-earthquake and post-earthquake SAR images that are closest in time to the two optical remote sensing images. Visually compare the pre-earthquake and post-earthquake optical remote sensing images with the landslide catalog. If it is found that no landslide has occurred in the optical remote sensing images in the landslide catalog, then delete that part of the landslide from the landslide catalog. If it is found that a landslide has occurred outside the landslide catalog, then add that part to the landslide catalog, thereby completing the correction of the landslide catalog. S402: Obtain two SAR image samples before and after the landslide. First, perform noise removal and geocoding. Then, generate a normalized mean ratio map sample based on the two SAR image samples before and after the landslide. Map the corrected landslide catalog onto the post-earthquake optical remote sensing image to obtain the landslide catalog mapped on the optical remote sensing image. Match the landslide catalog mapped on the optical remote sensing image with the normalized mean ratio map sample to obtain the normalized mean ratio map sample of the landslide area. S403, Upsample the normalized mean ratio map sample of the landslide area to obtain the upsampled normalized mean ratio map sample; Use a sliding window with a preset step size to generate a slice of a specified size from the upsampled normalized mean ratio map sample, and perform data augmentation by mirroring and rotating the slice to obtain the landslide dataset used to train the landslide detection network model.

6. A deep learning-based SAR image seismic landslide boundary detection system, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to execute the deep learning-based SAR image seismic landslide boundary detection method according to any one of claims 1 to 5.

7. A computer-readable storage medium storing a computer program, characterized in that, The computer program is used to be programmed or configured by a microprocessor to execute the deep learning-based SAR image seismic landslide boundary detection method according to any one of claims 1 to 5.

Images