A SAR image landslide information extraction method based on a deep learning algorithm

Abstract

The present application relates to the technical field of landslide information extraction, and particularly relates to a SAR image landslide information extraction method based on a deep learning algorithm. The method comprises the following steps: acquiring SAR images and optical images of an earthquake area; mapping the landslide boundary labeled in the optical images to the SAR images to label the SAR images to obtain a label image; constructing a data set based on the original image and the label image of the SAR image; training a pre-constructed semantic segmentation model using the data set; and generating a segmentation map based on the trained semantic segmentation model to extract a landslide area. The SAR images have rich polarization characteristics, and the deep learning semantic segmentation network is used for landslide area segmentation, thereby achieving better segmentation accuracy.

Description

Technical Field

[0001] This invention relates to the field of landslide information extraction technology, specifically to a method for extracting landslide information from SAR images based on deep learning algorithms. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active imaging system that acquires target information by emitting microwave pulses and receiving backscattered echoes from the ground. Unlike visible light imaging, SAR records slant range data rather than the spatial distances between ground features. Furthermore, in areas with significant topographic relief, such as mountains, geometric distortions such as perspective contraction, shadows, overlays, forward tilting of mountains, and inverted tops and bottoms can easily occur, making the reflected signals complex and varied. The outlines and textures of ground features in the images cannot be directly mapped to the actual terrain. Simultaneously, SAR echoes are affected by multiple factors, including wavelength, incident angle, polarization, and the dielectric properties of the ground features. Its imaging mechanism differs fundamentally from visible light, requiring specialized interpretation based on microwave scattering theory. These characteristics result in highly complex morphological and textural features of landslides in SAR images, making traditional visual interpretation methods inefficient and reliant on expert experience.

[0003] Traditional SAR image interpretation methods largely rely on expert-designed scattering features and prior models. While these methods offer physical interpretability, feature engineering struggles to fully characterize the heterogeneity of landslide bodies and their surrounding environment. Furthermore, manual feature extraction and model tuning are time-consuming and labor-intensive, heavily reliant on domain knowledge. With the launch of a series of radar remote sensing satellites, including Radarsat-2, COSMO-SkyMed, TerraSAR-X, ALOS, Sentinel-1, Gaofen-3, and Maritime Silk Road-1, a wealth of open-source satellite data and high-resolution imagery resources have become readily available at low cost. Various types of SAR data provide rich data support and assurance for the timely extraction of landslide disaster information. Among these, the Sentinel-1 satellite, with its free and open policy, stable observation capabilities, global coverage of dual-polarized SAR data, and high-frequency revisit characteristics, provides a multi-dimensional information source for dynamic landslide monitoring, becoming a core data support for wide-area disaster monitoring. Against this backdrop, overcoming the limitations of traditional methods and constructing an interpretation model that adapts to complex terrain and scattering mechanisms has become crucial for improving the accuracy of landslide identification. Summary of the Invention

[0004] The purpose of this invention is to provide a method for extracting landslide information from SAR images based on deep learning algorithms. Based on SAR images with rich polarization features, a deep learning semantic segmentation network is used to segment landslide areas, achieving better segmentation accuracy.

[0005] This invention provides a method for extracting landslide information from SAR images based on a deep learning algorithm, comprising:

[0006] Acquire SAR and optical images of the earthquake zone;

[0007] The landslide boundaries marked in the optical image are mapped onto the SAR image to annotate the SAR image and obtain a labeled image.

[0008] A dataset is constructed based on the original and labeled images of the SAR images;

[0009] The pre-built semantic segmentation model is trained using the dataset;

[0010] A segmentation map is generated based on a trained semantic segmentation model to extract the landslide area.

[0011] As a preferred embodiment, constructing training samples of a preset size based on the original image and the labeled image of the SAR image includes:

[0012] The original image is preprocessed by adjusting it to a uniform format and normalizing it.

[0013] The SAR image is labeled pixel by pixel according to the landslide boundary, and the corresponding label is an 8-bit color image.

[0014] A dataset is obtained by constructing training samples of a preset size based on the original image and its corresponding label image, as well as the original image and its corresponding label image that have undergone simultaneous geometric transformation.

[0015] In a preferred embodiment, the proportion of the target in each image of the dataset is greater than a preset threshold, and the pixel block size of the training samples is 128×128.

[0016] As a preferred embodiment, training a pre-built semantic segmentation model using the dataset includes:

[0017] The dual polarization features of the SAR image are stacked into two channels, and a blank channel is superimposed on a third channel to construct a multi-channel input;

[0018] Based on the multi-channel input, the pre-built semantic segmentation model is trained until the loss function converges.

[0019] As a preferred embodiment, the loss function is a composite loss function (BCEDiceLoss) that combines the binary cross entropy (BCE) and the Dice coefficient.

[0020] As a preferred embodiment, the selection method of the semantic segmentation model includes:

[0021] If the required segmentation accuracy is higher than a preset threshold, then UNet++ is determined as the semantic segmentation model.

[0022] If the SAR image contains severe noise, DeepLabV3+ is determined as the semantic segmentation model.

[0023] If the SAR image contains a complex scene, PSPNet is selected as the semantic segmentation model.

[0024] As a preferred embodiment, when PSPNet is determined to be the semantic segmentation model:

[0025] For the input image, a proportionally scaled feature map is extracted through the backbone network;

[0026] The pyramid pooling module (PPM) sequentially applies four different sizes of adaptive average pooling to the feature map. After each pooling branch, the number of channels is compressed to 1 / 4 of the original number through a 1×1 convolution.

[0027] Then, bilinear interpolation is used to upsample the feature maps of each branch back to the same spatial resolution as the input feature map;

[0028] The feature maps after multi-scale pooling and upsampling are concatenated with the original feature maps along the channel dimension, and the concatenated feature maps are then fused by convolution.

[0029] The fused low-resolution feature map is mapped back to the original size of the input image to output the predicted segmentation map.

[0030] As a preferred embodiment, when DeepLabV3+ is determined as the semantic segmentation model, the DeepLabV3+ network includes two parts: an encoder-ASPP module and a lightweight decoder.

[0031] At the encoder end, the input image extracts shallow and deep features through the ResNet34 backbone network, and then the deep features are fed into the ASPP module.

[0032] At the decoder, the deep semantic features output by the ASPP module are first upsampled to one-quarter of the original resolution using four times bilinear interpolation. In parallel, the high-resolution shallow features from the backbone network are compressed into channels by 1×1 convolution and then concatenated and fused with the upsampled deep features in the channel dimension. After one or two 3×3 convolutions for refinement, the images are finally restored to the original image size using four times interpolation to generate a pixel-level segmentation map.

[0033] As a preferred embodiment, when determining UNet++ as the semantic segmentation model:

[0034] For the input image, the encoder extracts feature maps at five scales in sequence through continuous convolution and downsampling operations. The spatial size of these maps decreases progressively from 1 / 2, 1 / 4, 1 / 8, 1 / 16, to 1 / 32, thus constructing a multi-scale feature pyramid.

[0035] The decoder uses normalization in each stage, and the channel configuration of the decoder corresponds to the encoder depth, which are (256, 128, 64, 32, 16) in sequence. The feature map output from each stage is restored to its size (×2) through upsampling operation until it finally matches the input resolution. Each decoding layer receives features from all preceding coding layers through skip connections.

[0036] The decoder uses a 1×1 convolution to map the last-level features to a single-channel output and generates a landslide binary segmentation mask through Sigmoid activation.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. This invention is based on SAR images with rich polarization features and uses a deep learning semantic segmentation network to segment landslide areas, achieving better segmentation accuracy.

[0039] 2. To avoid the network segmentation effect being affected by the small number of images in the dataset, this invention performs data augmentation by rotating, horizontally flipping, and vertically flipping the original images and label images simultaneously. This expands the dataset, increases sample diversity, alleviates the sample scarcity problem, and meets the data volume requirements for training deep learning models.

[0040] 3. This invention assigns each polarization channel to an independent channel, and each channel to a feature dimension. These channels can be cascaded along their directions to form a multi-channel input, allowing deep learning models to extract complementary polarization features. A stacking operation concatenates two channels along the channel axis, forming an input tensor of shape H×W×2 (H / W being the spatial dimension). This operation preserves the physical properties of the polarization scattering matrix, enabling the model to jointly learn both co-polarization and cross-polarization features. By constructing a multi-channel input, the deep network can simultaneously learn complementary features and cross-channel associations under various polarization modes, thereby improving the accuracy and robustness of tasks such as classification, detection, and change monitoring. Furthermore, to utilize pre-trained models, this invention overlays a blank channel in the third channel.

[0041] 4. Based on the constructed dataset and multi-channel input with stacked polarization features, this invention evaluates the prediction performance of UNet++, DeepLabV3+, and PSPNet to determine the advantages of each semantic segmentation model, thereby enabling the selection of semantic segmentation models according to actual scenario requirements. Attached Figure Description

[0042] Figure 1 This is a flowchart of the SAR image deep learning landslide extraction process in this invention;

[0043] Figure 2 This is the PSPNet network model structure in this invention;

[0044] Figure 3 This is the U-Net++ network model structure in this invention;

[0045] Figure 4 This is the DeepLabV3+ network model structure in this invention;

[0046] Figure 5 This invention presents some experimental results using PSPNet, UNet++, and DeepLabV3+ deep learning networks.

[0047] Figure 6 The training and validation loss curves of the PSPNet model in this invention are shown.

[0048] Figure 7 The training and validation loss curves of the Unet++ model in this invention are shown.

[0049] Figure 8 The training and validation loss curves of the DeepLabV3+ model in this invention are shown. Detailed Implementation

[0050] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0051] Driven by deep learning technology, breakthroughs have been achieved in the accuracy of remote sensing image interpretation. To achieve efficient extraction of landslide information from dual-polarization SAR images, this embodiment selects three representative deep learning semantic segmentation networks—PSPNet, UNet++, and DeepLabV3+—for training and validation on the Sentinel-1 dual-polarization (VV and VH) SAR dataset, as shown in the flowchart in Figure 1. Convolutional neural networks (CNNs) and their variant models significantly improve the extraction efficiency of landslide morphology, texture, and polarization characteristic parameters by automatically learning multi-scale feature representations. Among them, the PSPNet pyramid scene parsing network excels in street scene understanding and terrain classification tasks by fusing contextual information at different scales; the UNet++ architecture effectively alleviates the problem of detail loss in medical image segmentation through dense skip connections and deep supervision mechanisms; the DeepLabV3+ network improves the backbone network based on DeepLabV3 and introduces an encoder-decoder architecture. Its encoder uses ResNet34 instead of Xception as the backbone network, and then extracts multi-scale features through the Spatial Pyramid Pooling (ASPP) module. The decoder then fuses deep semantic features with shallow detail features through upsampling, ultimately achieving high-precision segmentation prediction results. The successful application of these models in the field of natural image segmentation provides a transferable technical path for landslide recognition tasks.

[0052] The Pyramid Scene Parsing Network (PSPNet) is a semantic segmentation model based on a fully convolutional network (FCN). Its core innovation lies in the introduction of a pyramid pooling module (PPM), which enhances pixel-level classification accuracy through multi-scale contextual information aggregation. This module divides the input feature map into different sub-regions (e.g., 1×1, 2×2, 3×3, 6×6), performs adaptive average pooling on each sub-region to capture both global context and local details, and then upsamples and concatenates them with the original feature map. This effectively expands the receptive field and addresses the insufficient scale adaptability problem caused by the fixed receptive field in traditional models. Furthermore, PSPNet uses ResNet34 as its backbone network and leverages residual connections to alleviate gradient vanishing, enhancing model convergence and feature representation capabilities.

[0053] PSPNet's technical advantages are reflected in its multi-scale feature fusion and efficient parameter design: 1) By integrating contextual information of different granularities through pyramid pooling, it significantly improves the segmentation performance of targets with significant size differences in complex scenes, especially for small objects; 2) It adopts a cascaded fusion strategy to upsample multi-resolution feature maps to their original size, and combines channel dimensionality reduction to reduce the number of parameters, thereby improving computational efficiency while ensuring accuracy; 3) Its hierarchical pooling mechanism can generate fine-grained semantic representations, and by fusing global statistics and local details, it enhances the robustness of the model to multi-scale targets in scenarios such as remote sensing images and autonomous driving.

[0054] Experiments show that this model, through hierarchical pooling and residual structure optimization, can efficiently handle large-scale scene segmentation tasks, demonstrating excellent performance on public datasets such as PASCAL VOC and Cityscapes. Its lightweight design (few parameters, high computational efficiency) and multi-scale adaptability make it widely applicable in fields such as scene understanding, high-resolution remote sensing image analysis, and autonomous driving environmental perception, providing an important reference for the scale-aware design of subsequent semantic segmentation models.

[0055] PSPNet is a deep learning-based image semantic segmentation network that uses pooling kernels of different sizes (PPM) to capture contextual information at different scales, achieving high-precision image segmentation. PSPNet possesses global contextual information extraction capabilities, high-precision segmentation and boundary detection capabilities, multi-scale processing capabilities, and efficient training and inference performance, enabling it to provide high-quality object detection results in various practical applications.

[0056] In semantic segmentation tasks, PSPNet combines visual feature extraction, global-local context aggregation, and multi-scale fusion to form an end-to-end four-module architecture: feature extractor, pyramid pooling module, feature fusion module, and classifier (e.g., ...). Figure 2 (As shown). Its basic framework uses a pre-trained convolutional neural network (such as ResNet34) as the core to build a multi-level feature encoder, generating high-dimensional semantic feature maps through convolution stacking. It innovatively introduces a spatial pyramid pooling component, employing an adaptive average pooling strategy to capture multi-scale contextual information while preserving local details. In the feature fusion stage, bilinear interpolation is used to achieve spatial alignment of the feature maps of each branch, and channel concatenation and 1×1 convolutional kernels are used for cross-scale feature fusion to generate enhanced composite features. The classifier is a pixel-level semantic decoder that transforms the fused features to the category space through pixel-by-pixel convolution mapping, and finally outputs the semantic probability distribution of each pixel after normalization by the softmax function, completing the scene parsing task. This architecture significantly improves the semantic segmentation accuracy in complex scenes through a cascaded multi-scale feature extraction and fusion mechanism.

[0057] The feature extractor uses ResNet34 as the backbone. Its core residual structure effectively alleviates the gradient vanishing and performance degradation problems in deep network training by introducing shortcut connections. Specifically, each residual module contains two 3×3 convolutional layers and uses 1×1 convolutions with a "dimensionality increase followed by dimensionality reduction" approach to adjust the number of channels, reducing the number of parameters by about 30% while maintaining feature expressiveness. The residual structure forms a residual mapping by directly adding the input features to the convolutional output, allowing the network to focus on learning the difference function between the input and the target. This design allows the network depth to be scaled to tens or even hundreds of layers, while the training complexity only increases sublinearly, and at the same time ensures the effective transfer of low-level features.

[0058] The pyramid pooling module aims to fuse features at different scales to capture contextual information of varying sizes, thereby obtaining more comprehensive global information. This module employs a four-level parallel pooling structure, dividing the input feature map into 1×1, 2×2, 3×3, and 6×6 grid sub-regions, and performing average pooling on each sub-region. After compressing the number of channels through 1×1 convolution, bilinear interpolation upsamples the four sets of pooled features at different scales to their original size, ultimately obtaining four feature maps with one-quarter the number of channels of the original features and consistent spatial resolution. This design enables the network to simultaneously capture global semantic information and local detailed features, effectively solving the problem of contextual information loss caused by single-scale pooling.

[0059] The feature fusion engine achieves complementary fusion of deep semantic features and shallow detailed features by constructing feature pyramids at different levels. First, the fourth-level pooled features output by the PPM are upsampled to match their spatial dimensions with the mid-level feature maps output by the backbone network. Then, concatenation is performed along the channel-last dimension to generate a composite feature tensor containing multi-scale contextual information. To unify the feature dimensions, a 1×1 convolution is used to adjust the number of channels to a fixed value, and bilinear interpolation is used to upsample high-level semantic features to the resolution of the low-level feature maps. Finally, channel concatenation achieves semantic alignment between high and low-level features, enabling the network to simultaneously utilize deep abstract features and shallow spatial details, significantly improving semantic segmentation accuracy in complex scenes.

[0060] The classifier, as the core component of the decoding stage, employs a lightweight design to achieve end-to-end pixel-level prediction. First, a 1×1 convolution operation is performed on the fused composite feature map to generate a prediction heatmap corresponding to the number of target classes. To restore spatial resolution, bilinear interpolation is used to upsample the low-resolution prediction map to the original size of the input image. This decoding process eliminates the need for complex operations such as deconvolution, effectively mitigating boundary blurring caused by interpolation while maintaining inference efficiency and preserving spatial details through skip connections. The final pixel-level classification result is then supervised and optimized using a cross-entropy loss function, realizing a complete inference process from feature encoding to semantic decoding.

[0061] To address the challenges of complex landslide morphology, significant scale differences, and strong background interference in SAR imagery, the Pyramid Scene Parsing Network (PSPNet) first employs a ResNet34 backbone to stably transfer features. Then, through a pyramid pooling module, it aggregates global and local contextual information at multiple scales, enhancing the spatial semantic understanding of landslide areas. Subsequently, it combines a cross-level feature fusion strategy to concatenate semantic features from different scales and unify the resolution through convolution and upsampling operations, further improving boundary delineation accuracy and maintaining consistency across heterogeneous regions. Finally, a lightweight decoder achieves efficient pixel-level classification, balancing computational efficiency and segmentation accuracy. This method demonstrates excellent resistance to noise and complex terrain interference in high-resolution SAR imagery landslide extraction, providing reliable technical support for earthquake and landslide disaster emergency response.

[0062] The UNet++ model is an improved semantic segmentation model based on the classic U-Net architecture. Its core innovation lies in the introduction of nested convolutional blocks and a dense skip connection mechanism. Compared to U-Net's single skip connection, UNet++ forms a nested topology by densely connecting the encoder and decoder subnetworks at each level, effectively reducing the expression difference between shallow detail features and deep semantic features. This model borrows the dense connection concept from DenseNet, enabling direct interaction between features at different levels during decoding, enhancing the fusion capability of multi-scale contextual information, and significantly improving the accuracy of capturing complex scene boundaries and minute structures.

[0063] In terms of structural design, UNet++ constructs a four-layer nested subnetwork by filling the "hollow" jump paths of the U-Net: each subnetwork consists of an encoder-decoder unit and is densely connected to adjacent subnetworks through upsampling paths. This hierarchical design allows features to be progressively refined during cross-scale propagation, while integrating semantic information from different resolutions using a feature stacking strategy. Furthermore, UNet++ supports dynamic pruning, which can quickly simplify the model structure by removing deep subnetworks, reducing computational complexity while maintaining segmentation accuracy, thus achieving a flexible trade-off between accuracy and efficiency.

[0064] During training, UNet++ employs a multi-level joint optimization strategy, updating model parameters by weighted fusion of the loss functions of each sub-network. This design not only enhances the consistency of feature representation but also provides configurability for model deployment: users can choose to retain sub-network depth or remove redundant layers through pruning, thus adapting to application scenarios under different computational resource constraints. Experiments show that UNet++ demonstrates superior boundary preservation and robustness compared to traditional U-Net in image segmentation tasks, becoming one of the benchmark models in this field.

[0065] Building upon the classic U-Net, UNet++ significantly enhances the fusion and detail representation capabilities of multi-scale features through nested subnetworks and dense skip connections. Unlike traditional U-Net, which uses only a single skip connection to pass features from the encoder to the decoder at the same layer, UNet++ introduces multiple consecutive convolutional blocks between each encoder and decoder layer, and strings these blocks together with dense connections to form a nested U-Net structure. Figure 3 As shown in the figure. This design not only expands the network's effective receptive field, enabling low-level features to interact with high-level semantics multiple times, but also narrows the semantic gap between shallow and deep features through dense pathways, thereby improving the ability to capture image details (such as edges and small objects) and multi-scale contextual information.

[0066] During the decoding phase, UNet++ employs a multi-decoder (multi-sub-network) joint training strategy: each decoding sub-network receives features from encoders of different depths and has an independent supervision branch at its end, collectively optimizing network performance. This multi-supervision not only accelerates gradient flow and alleviates shallow gradient vanishing, but also allows for flexible layer-by-layer pruning of the network to reduce computational and memory overhead without significantly sacrificing accuracy. Users can adjust the number and depth of sub-networks according to task requirements to achieve different trade-offs between accuracy and efficiency.

[0067] In terms of network structure, both the encoder and decoder of UNet++ employ 3×3 convolutions with a stride of 1. During encoding, 2×2 max pooling with a stride of 2 is used for downsampling. During decoding, interpolation with an upsampling factor of 2 and a stride of 2 is used to restore spatial resolution. Through the synergistic effect of short connections, upsampling, and dense skip paths, UNet++ can more comprehensively capture contextual information and detailed features at different scales in high-resolution scenes such as remote sensing images and medical images, becoming an effective solution for high-precision semantic segmentation.

[0068] Based on a comparative analysis of the classic U-Net architecture, this model introduces multi-level residual dense blocks in the lateral connection pathways. The design motivation is to address the gradient decay problem in deep feature propagation. The specific concatenation process can be performed according to the following formula:

[0069]

[0070] In this context, network nodes are identified by their (i, j) coordinates (corresponding to...). Figure 3 (See the top right corner annotation). The feature fusion process involves the following core operations: U(x) is defined as an upsampling operator based on bilinear interpolation, [x, y] represents the concatenation of feature tensors along the channel dimension, and H(x) represents a composite operation containing two consecutive convolutional layers, each followed by an activation function. For example, we can see... Figure 3 This is the topmost convolutional block in the U-net++ architecture. Where U(X) 1,0 ) is for X 1,0 The result is upsampled once, X 0,1 It is U(X) 1,0 ) and X 0,0 The result of convolution after concatenation. Similarly, X 0,4 It is U(X) 1 ,3 ) and X 0,0 X 0,2 X 0,3 The result of convolution after stitching. The number of convolution kernels was adjusted to ensure that the images from the five convolutional blocks in the first layer were of completely identical size.

[0071] The UNet++ model consists of an encoder and a decoder. The encoder uses a series of convolutional kernels and max-pooling layers to extract multi-scale features through progressive downsampling; the decoder gradually restores the feature resolution through upsampling layers and deconvolutional layers. Unlike the direct skip connections of traditional U-Net, UNet++ constructs a multi-level connection network between the encoder and decoder through nested dense convolutional blocks: a dense skip connection mechanism is introduced in the skip path to fuse encoder features and decoder features at different depths layer by layer, effectively reducing the information gap between the deep semantic features of the encoder and the shallow detailed features of the decoder, thereby achieving pixel-level fine segmentation.

[0072] Addressing the diverse landslide morphologies (significant differences in shape, size, and texture) in SAR imagery, UNet++ significantly enhances model adaptability and generalization capabilities through its densely nested structure and multi-level feature fusion. Its densely connected design not only improves gradient propagation efficiency but also enhances multi-scale perception by integrating feature maps from different receptive fields, enabling accurate extraction of landslide area boundaries and texture details even under complex background interference. UNet++ achieves cross-level feature complementarity through nested topological networks, overcoming the limitations of traditional methods such as sensitivity to landslide scale and blurred boundaries while ensuring the integrity of feature representation. This provides a solution that combines accuracy and efficiency for high-resolution remote sensing image segmentation tasks.

[0073] Semantic segmentation faces two main challenges: insufficient adaptability to multi-scale targets and boundary blurring due to feature resolution loss. The former manifests as a single-scale convolutional kernel's inability to simultaneously capture effective features of both large targets and small objects, while the latter stems from the reduced feature map resolution caused by multiple downsampling operations, leading to decreased prediction accuracy and loss of detailed information (such as edges). Although DeepLabV1 and V2 have expanded the receptive field and supported multi-scale feature fusion by introducing dilated convolutions and spatial pyramid pooling (ASPP) modules, limitations remain—dilated convolutions with a single dilation rate struggle to cover a wide range of target scales, the fusion efficiency of global and local contextual information is low, and deep networks lack sufficient ability to resolve edge details.

[0074] DeepLab V 3+ is an improved end-to-end semantic segmentation model. Its core innovation lies in the fusion of an encoder-decoder architecture and a multi-scale feature extraction mechanism. The encoder uses ResNet34 as the backbone network, extracting high semantic features through deep convolutions and combining ASPP (which includes dilated convolutions with different dilation rates and global pooling branches) to achieve parallel capture of multi-scale contextual information. The decoder, on the other hand, fuses deep semantic features with shallow high-resolution features output by the encoder through bilinear upsampling across layers, and gradually restores the original image resolution by optimizing the feature representation using 3×3 convolutions. This model significantly reduces computational complexity while maintaining high segmentation accuracy through lightweight design (such as depthwise separable convolutions), and is particularly adept at handling complex scenes with significant differences in target scale and blurred boundaries. It demonstrates strong robustness and real-time performance advantages on datasets such as PASCAL VOC.

[0075] DeepLab V The encoder of 3+ consists of an improved ResNet34 backbone network and an ASPP module: the input image is extracted by ResNet34 and outputs shallow and deep features. The shallow features are directly passed to the decoder and compressed through 1×1 convolution to reduce the proportion of redundant information. The deep features are input to the ASPP module for multi-scale context enhancement. This module performs one 1×1 convolution, three 3×3 dilated convolutions with different dilation rates (6 / 12 / 18) and global average pooling in parallel. Then, it integrates the multi-branch features through 1×1 convolution and finally outputs high-weight deep features that fuse global semantics and local details, providing the decoder with a core input that combines large-scale contextual information and dense feature expression.

[0076] The backbone network module uses the ResNet34 model as the basic feature extractor, a design that fully leverages ResNet34's advantages in image feature representation. Specifically, ResNet34 effectively alleviates the vanishing gradient problem in deep neural networks through its unique residual learning framework, enabling the network to construct deeper hierarchical structures to capture multi-scale contextual information. (In DeepLab) V In ResNet34 3+, the convolutional layers are reconstructed as dilated separable convolutions—replacing standard convolutions with a combination of depthwise separable and dilated convolutions. This design reduces the number of parameters through depthwise convolutions while expanding the receptive field using the dilation mechanism, capturing multi-scale contextual information while maintaining high-resolution feature maps. Furthermore, the deeper stages of ResNet34 use dilated convolutions instead of traditional downsampling operations to avoid excessive loss of feature map resolution, thereby improving sensitivity to object boundaries. The improved ResNet34 backbone achieves a balance between computational efficiency and feature representation capability, adapting to the needs of dense prediction tasks while providing a high-resolution, multi-granular feature foundation for the global semantic fusion of the subsequent ASPP module, ultimately significantly enhancing the model's adaptability to complex scenes.

[0077] The ASPP module achieves efficient fusion of multi-scale contextual features through a multi-branch parallel structure. Its core design comprises four key parts: First, it employs parallel dilated convolutions with multiple dilation rates, enabling different branches to cover receptive fields from local to wide-area, simultaneously capturing target details and large-scale contextual associations. Second, it introduces a global average pooling branch, compressing feature maps to a global context vector and restoring resolution through 1×1 convolution and bilinear upsampling, supplementing image-level semantic information to enhance background understanding. Furthermore, it includes a standard 1×1 convolution branch to directly extract local basic features and reduces computational redundancy through channel dimensionality reduction. The feature maps output from all branches are concatenated along the channel dimension, then integrated with multi-scale information through 1×1 convolution, followed by batch normalization (BN) and ReLU activation for non-linear mapping and feature calibration. Through an adaptive weight allocation mechanism, this module enables the model to accurately locate small target details and effectively associate large-scale contexts in complex scenes, significantly improving robustness to targets with diverse scales.

[0078] The decoder module is designed to address the issue of reduced feature map resolution and loss of boundary details caused by multiple downsampling in the encoder. It achieves refined segmentation through multi-level feature fusion and progressive upsampling. Specifically, the low-resolution deep features output by the DeepLabv3+ encoder (enhanced with multi-scale semantic information by the ASPP module) are first upsampled by 4x using bilinear interpolation to recover some spatial details. Then, they are concatenated across layers with the high-resolution low-level features extracted from the shallow layers of the encoder. To avoid the influence of shallow feature channel redundancy on the fusion effect, the decoder uses 1×1 convolutions to reduce its dimensionality (e.g., compressing the number of channels from 256 to 48), and optimizes the fused features using 3×3 convolutions combined with BN and ReLU activation functions. Finally, it is upsampled by 4x to output a segmentation result with the same resolution as the original image.

[0079] DeepLab V The core innovation of ASPP 3+ lies in its organic integration of ASPP with a lightweight encoder-decoder architecture. At the encoder, ASPP employs dilated convolutions with varying dilation rates and global average pooling branches in parallel, effectively expanding the receptive field and enhancing multi-scale contextual modeling. At the decoder, shallow high-resolution feature channels are compressed using 1×1 convolutions, then skip-connected to deep semantic features processed by the ASPP module, and spatial details are progressively restored through bilinear upsampling, significantly improving boundary segmentation accuracy. Simultaneously, depthwise separable convolutions are widely used in the optional ResNet34 backbone network, ASPP, and decoder modules, maintaining high expressive power while significantly reducing the number of parameters and computational cost, achieving real-time inference performance. This design retains the advantage of dilated convolutions in acquiring contextual information without reducing resolution, while restoring image details through a concise and efficient decoder, thus demonstrating outstanding performance in high-precision semantic segmentation tasks, especially in scenarios requiring fine boundary recognition.

[0080] To address the diverse morphological characteristics, complex background noise, and multi-scale features of landslide targets in SAR imagery, DeepLab... V The 3+ model achieves an organic fusion of multi-scale feature extraction and noise suppression through parallel acquisition of contextual information at different scales via the ASPP module, reduced computational load via depthwise separable convolutions, and cross-layer skip connections and channel compression strategies in a lightweight encoder-decoder architecture. This allows the model to flexibly adapt to the diverse forms of landslide targets while accurately characterizing landslide boundaries. Its end-to-end design maintains high-precision segmentation while also considering computational efficiency, providing an accurate, practical, and real-time solution for landslide disaster monitoring in complex geological environments.

[0081] This embodiment first constructs a sample set of landslides from the Luding earthquake based on VV and VH dual-polarization SAR images acquired by the Sentinel-1A satellite. A deep learning semantic segmentation method is then used to extract and analyze the post-earthquake landslide areas. Specifically, the Sentinel-1A image of the experimental area taken on September 14, 2022, after the earthquake, with a spatial resolution of 5m × 20m, is selected. Considering the lack of intuitive ground feature information in polarimetric SAR data, sample zoning presents a significant challenge. Therefore, this study utilizes post-earthquake optical images from the Sentinel-2 satellite for auxiliary annotation. By registering the images with Sentinel-2 optical images from the same time period, the manually annotated landslide boundaries in the optical images are mapped to the SAR images, ensuring the accuracy of the training sample areas.

[0082] During sample construction, systematic tests were conducted for different pixel block sizes (e.g., 64×64, 128×128, 256×256, etc.). Results showed that excessively small pixel blocks could lead to insufficient information, affecting classification accuracy; excessively large pixel blocks introduced significant redundancy, increasing the computational burden on the model. After comprehensively comparing classification performance and training efficiency, the 128×128 pixel block performed best in both landslide and non-landslide segmentation tasks, effectively capturing local structural features while maintaining high computational efficiency, providing a solid foundation for subsequent model training.

[0083] In the semantic segmentation model training process, the dataset is divided into two parts. The first part is the original images, which are uniformly formatted as .tif for ease of training and normalized at the input. The second part is the labeled images, which are annotated pixel by pixel according to the actual landslide shape, and the corresponding labels are 8-bit color images.

[0084] In dual-polarization SAR data, surface scattering characteristics are captured through different polarization combinations (such as horizontal transmit-horizontal receive (HH) and horizontal transmit-vertical receive (HV)), providing rich feature information. Each polarization channel corresponds to an independent channel, and each channel corresponds to a feature dimension. These channels can be cascaded along the channel direction to form a multi-channel input, which is used by deep learning models to extract complementary polarization features. The stacking operation concatenates two channels along the depth dimension (channel axis) to form an input tensor of shape H×W×2 (H / W being the spatial dimension). This operation preserves the physical properties of the polarization scattering matrix, allowing the model to jointly learn co-polarization and cross-polarization features. By constructing a multi-channel input, the deep network can simultaneously learn complementary features and cross-channel correlations under various polarization modes, thereby improving the accuracy and robustness of tasks such as classification, detection, and change monitoring.

[0085] In landslide extraction and segmentation, this embodiment stacks dual-polarization features into two channels. To maintain dataset balance, data is randomly sampled to ensure the target's proportion in each image is greater than 0.3, preventing model overfitting. To utilize the pre-trained model, a blank channel is overlaid in the third channel. Operations were performed using PSPNet, UNet++, and DeepLabV3+ deep learning networks, respectively. Some experimental results are shown below. Figure 5 As shown, the model's accuracy is gradually improved by adjusting the learning rate and the number of training iterations.

[0086] The core architectural innovation of PSPNet lies in the PSP module, which enhances the network's ability to perceive multi-scale targets through a multi-level context aggregation mechanism. The PSPNet model structure can be divided into two parts: first, the backbone feature extraction network, which extracts features from the image by performing downsampling operations; and second, the enhanced feature extraction structure, which upsamples the extracted features at multiple scales to restore the image size and output the corresponding predicted image.

[0087] First, the input image is uniformly scaled to 128×128 pixels, and a proportionally reduced feature map is extracted through the backbone network. This feature map is then divided into two parts, one of which enters the pyramid pooling module (PPM). In this module, four different sizes of adaptive average pooling are applied sequentially to the feature map—dividing the image into 1×1, 2×2, 3×3, and 6×6 blocks and pooling them separately—to capture global and local contextual information. After each pooling branch, a 1×1 convolution is used to compress the number of channels to 1 / 4 of the original, and then bilinear interpolation is used to upsample the feature maps of each branch back to the same spatial resolution as the input feature map. Finally, these multi-scale pooled and upsampled features are concatenated with the retained original feature parts along the channel dimension, and further fused through subsequent convolutions. This allows the model to possess both sensitivity to detailed textures and good perception of the overall structure, thus achieving high-quality feature extraction.

[0088] In this experiment, PSPnet uses a ResNet34 backbone network with a 3-level encoder depth. A multi-scale feature pyramid is extracted using a convolutional network. Successive 3×3 convolutions and downsampling operations compress the feature map layer by layer, outputting feature maps with three spatial resolutions, sized to 1 / 2, 1 / 4, and 1 / 8 of the input. To accelerate convergence and improve feature representation quality, pre-trained weights on ImageNet are loaded during encoder initialization.

[0089] Following the Pyramid Pooling (PPM) module, the decoder first performs batch normalization on the number of channels (512) of the multi-scale feature maps to improve training stability and suppress overfitting. Simultaneously, a dropout rate of 0.2 is set in the spatial pyramid branch to enhance robustness. The PPM contains four parallel adaptive average pooling branches (1×1, 2×2, 3×3, and 6×6 grids). The outputs of each branch are dimensionality-reduced by 1×1 convolution, then upsampled to the original feature map size via bilinear interpolation, and concatenated with the encoder's third-level features along the channel direction. Finally, a 1×1 convolution compresses the channels, achieving multi-scale fusion of global and local contexts. The decoder performs a series of convolution and upsampling operations. The training and validation loss curves are shown below. Figure 6 As shown.

[0090] Building upon the core ideas of UNet's encoder-decoder symmetric structure and skip connections, UNet++ innovatively introduces DenseNet's dense connection mechanism to construct a nested, multi-level feature fusion architecture, achieving an exponential expansion of feature representation. The encoder consists of multiple downsampling modules that progressively extract low-resolution semantic features from the image. The decoder, through a nested branch design, receives features from all preceding encoder layers via skip connections and progressively upsamples to restore resolution. In the landslide extraction and semantic segmentation task, its nested connections enable the model to simultaneously perceive local details (such as landslide boundary textures in SAR images) and global context (such as regional geological structure background), better fusing feature information from different levels and significantly improving semantic segmentation accuracy and feature representation capabilities in complex scenes.

[0091] In this experiment, UNet++ uses ResNet34 as the backbone network, leveraging its residual structure to effectively alleviate gradient vanishing and enhance feature representation capabilities. The encoder extracts feature maps at five scales sequentially through continuous convolution and downsampling operations, with their spatial dimensions decreasing progressively from 1 / 2, 1 / 4, 1 / 8, 1 / 16, to 1 / 32, constructing a multi-scale feature pyramid to meet the multi-scale fusion requirements of the decoder. To accelerate convergence and improve feature quality, pre-trained weights on ImageNet are loaded during encoder initialization.

[0092] The decoder employs normalization at each stage to improve training stability, suppress overfitting, and accelerate convergence. The decoder's channel configuration corresponds to the encoder depth, at (256, 128, 64, 32, 16). The feature map output at each stage is upsampled to restore its size (×2) until it matches the input resolution. In each decoding unit, the input of the current layer not only fuses skip connections from the same layer and deeper encoders but also forms dense connections with the output of the previous decoding layer; subsequently, cross-layer feature fusion is achieved through convolution and upsampling. At the end of the decoder, a 1×1 convolution maps the final stage features to a single-channel output, and a sigmoid activation is used to generate a sliding binary segmentation mask to meet the requirements of binary classification segmentation tasks. The training and validation loss curves are shown below. Figure 7 As shown.

[0093] UNet++, as an improved semantic segmentation model, demonstrates significant advantages in remote sensing image land cover classification and target extraction tasks due to its dense skip connection structure and multi-scale feature fusion mechanism. This model enhances feature representation through nested decoder paths, effectively improving segmentation accuracy in complex scenes. However, its dense connection mechanism leads to a high number of parameters and computational complexity, placing stringent requirements on hardware resources and the completeness of labeled data, and still carries the risk of detail loss in the segmentation of small targets. Addressing the challenges of significant multi-scale characteristics, complex spatial structure, and high resolution in SAR remote sensing images, UNet++'s hierarchical feature aggregation capability can adaptively capture land cover information of different granularities, thereby optimizing segmentation efficiency and accuracy. In practical applications, a trade-off between model performance and computational cost is necessary, adjusting network depth and parameter configuration according to task requirements.

[0094] UNet++'s innovative architecture reduces the complexity of the optimization process by reconstructing the jump path between the encoder and decoder and introducing dense convolutional blocks to bridge the feature-semantic gap. Simultaneously, a deep supervision mechanism embeds auxiliary loss functions at each stage of the decoder, forcing the model to learn multi-level discriminative features and significantly improving the segmentation consistency of landslide bodies across multiple scales. To verify its effectiveness, this embodiment conducts comparative experiments based on the Luding earthquake landslide SAR image dataset. UNet++ performs better in terms of landslide area boundary detail and small-area landslide detection. This demonstrates the effectiveness of nested jump paths for remote sensing landslide segmentation and provides a reference for subsequent appropriate pruning or lightweight modification in resource-constrained environments.

[0095] DeepLab VThe 3+ network consists of two parts: an encoder – the ASPP module and a lightweight decoder. At the encoder, the input image is first processed by the ResNet34 backbone network to extract shallow and deep features. Then, the deep features are fed into the ASPP module, which contains multi-branch dilated convolutions (commonly with dilation rates of 6, 12, and 18) and global average pooling. After concatenation, the features are reduced in dimensionality by 1×1 convolutions and fused to ensure that multi-scale contextual information is obtained while maintaining resolution.

[0096] At the decoder, the deep semantic features output by ASPP are first upsampled to one-quarter of the original resolution using four times bilinear interpolation, significantly restoring spatial details. Simultaneously, the shallow high-resolution features from the backbone network are compressed into channels via 1×1 convolutions and then concatenated with the upsampled deep features along the channel dimension. This is followed by one or two 3×3 convolutions for refinement, and finally, the image is restored to its original size using four times interpolation, generating a pixel-level segmentation map. This design balances wide-area context capture with detailed boundary restoration, demonstrating high accuracy and good real-time performance in scenarios requiring fine boundaries, such as remote sensing landslide extraction.

[0097] In experiments based on SAR imagery landslide datasets, DeepLab V The 3+ model significantly improves landslide extraction performance through multi-level structural optimization: the backbone network adopts ResNet34, whose residual connection design enhances feature reuse capabilities through skip gradient propagation, effectively alleviating the gradient degradation problem in deep network training, especially in complex terrain areas, where it can retain high-resolution features and support fine segmentation of landslide boundaries; the ASPP module integrates multi-scale dilated convolutions (dilation rates 6 / 12 / 18) with global average pooling branches, simultaneously capturing local scattering anomalies (such as deformation at the trailing edge of the landslide body) and wide-area terrain correlations, and combines depthwise separable convolutions to reduce computational complexity, suppress SAR image speckle noise interference, and improve the robustness of multi-scale landslide detection; the decoder, through a cross-level feature fusion mechanism, concatenates deep semantic features with shallow high-resolution features after bilinear upsampling, uses 3×3 convolutions to optimize boundary responses, and combines progressive upsampling strategies to restore details, achieving sub-pixel-level segmentation accuracy even under weak annotation conditions, providing an efficient and robust end-to-end solution for landslide monitoring in complex terrain scenarios. The training and validation loss curves are shown below. Figure 8 As shown.

[0098] DeepLabV3+ employs ResNet34 as its encoder, with a hierarchical structure comprising four stages (Stage-1 to Stage-4), outputting [64, 128, 256, 512] channels respectively. The encoder loads ImageNet pre-trained weights, enhancing feature reuse and gradient propagation efficiency through residual connections, significantly accelerating model convergence and improving the representation of SAR image texture features. To adapt to the multi-scale analysis requirements of high-resolution remote sensing images, the encoder introduces an extended dilated spatial pyramid pooling (ASPP) module after Stage-4: this module deploys three sets of dilated convolutional kernels with different dilation rates (Rate = 6, 12, 18) in parallel, combined with a global average pooling branch to capture global context, constructing a multi-scale feature pyramid. The ASPP output channel number is compressed to 256 through 1×1 convolution, reducing computational complexity while maintaining semantic richness.

[0099] The decoder employs a dual-path feature fusion strategy: low-level features (Stage-2 output, 128 channels) are projected to 256 dimensions via 1×1 convolution and then skipped to the ASPP output, achieving complementary spatial details and semantic information through cascaded weighted fusion. The fused features are then refined through 3×3 convolutions to refine boundary information, and bilinear interpolation is used to restore spatial resolution by four times (upsampling=4), ultimately generating a feature map of the same size as the input image. The segmentation head consists of a single-layer 1×1 convolution, mapping the 256-dimensional features to a single channel (number of classes = 1), and generating a landslide binary segmentation mask using a sigmoid activation function. This configuration balances the requirements of accuracy and efficiency in remote sensing image segmentation tasks.

[0100] Because deep learning model training requires a large amount of data, to avoid the impact of a small number of images in the dataset on the network's segmentation performance, data augmentation was performed on the original and labeled images by rotating, horizontally flipping, and vertically flipping them. This resulted in a dataset of 5000 sets containing dual-polarized original SAR images and labeled images. The dataset was then randomly divided into a 70% training set, a 15% test set, and a 15% validation set. The training set was used to train the model parameters, the validation set was used to evaluate the model's generalization ability after each training batch and hyperparameter tuning, and the test set was used to evaluate the network's final segmentation results. By increasing the amount of sample data through data augmentation methods, the training effect and generalization ability of the network model were improved.

[0101] The three segmentation models proposed in this embodiment are all implemented based on the PyTorch deep learning framework. Model training uses the stochastic gradient descent Adam optimizer to optimize the classification network. It combines momentum and adaptive learning rate methods, adjusting the learning rate of each parameter by calculating the first moment (mean) and second moment (variance) of the gradient, thereby accelerating training and improving convergence. The initial learning rate is lr = 0.0001, dynamically adjusted using a poly strategy, weight_decay = 0.01, and the batch size is set to 4. In each epoch, 3500 samples (128×128 pixels) are randomly selected from all samples for training. The hyperparameters are shown in Table 1.

[0102] Table 1 Hyperparameter Settings

[0103]

[0104] The loss function is a key metric that measures the difference between the model's predictions and the true labels. During training, minimizing this function guides model parameter updates, thereby achieving good generalization to new data. This embodiment employs a composite loss function (BCEDiceLoss) combining binary cross entropy (BCE) and the Dice coefficient. This combines the advantages of pixel-level classification accuracy and overall region overlap, particularly beneficial for landslide extraction in binary SAR images with foreground-background imbalance. In such cases, the foreground (landslide area) often occupies a few pixels, while background noise is strong and easily confused. BCEDiceLoss significantly improves the model's detection accuracy for landslide boundaries and small landslide bodies, reducing false positives and false negatives. The mathematical definition of the composite loss function is as follows:

[0105] BCEDiceLoss=α*BCE+β*DiceLoss (2)

[0106] Here, α and β are weighting coefficients (usually equal weighted, i.e., α = β = 0.5). BCE focuses on the classification accuracy of each pixel, providing a stable gradient signal for fine-grained errors. Dice Loss focuses more on overall shape consistency and region-level matching, and is particularly effective for situations with sparse foreground pixels or class imbalance, effectively mitigating the problem of small objects or a few classes being ignored. Typically, α = β = 0.5 is chosen to balance pixel-level and region-level errors. BCEDiceLoss considers both local classification accuracy and global region overlap, helping the model maintain robustness when the foreground and background are imbalanced.

[0107] This embodiment uses five evaluation metrics to comprehensively assess the model's performance on the test set: mean Intersection over Union (mIoU), accuracy, precision, recall, and F1 score. mIoU is the average of the IoU calculated for each class, measuring the average overlap between the predicted and actual segmented regions. Its value ranges from 0 to 1; a higher value indicates more accurate segmentation. Accuracy reflects the proportion of pixels correctly classified by the model (all pixels). Precision focuses on how many of the results the model classifies as "positive" are actually slippery pixels, reflecting the reliability of the positive predictions; higher precision means a lower false positive rate. Recall is how many of all real slippery pixels are successfully detected by the model, reflecting the model's detection capability; high recall indicates fewer missed detections, meaning the model can identify most real slippery regions. The F1 score is a harmonic average of precision and recall, comprehensively considering the model's positive class detection performance.

[0108] In the SAR image landslide extraction task, PSPNet, UNet++, and DeepLabV3+ models all demonstrated their advantages across multiple evaluation metrics (as shown in Table 2). Overall, UNet++ performed best in mIoU and Precision, PSPNet slightly outperformed in Recall and F1, while DeepLab... V Due to its high dependence on global context, SAR imagery with noise and small-scale targets presents certain challenges. Specifically:

[0109] In terms of overall segmentation accuracy and region coverage: UNet++ (86.05%) slightly outperforms PSPNet (84.05%) in mIoU, while DeepLab... V The score for 3+ was relatively low (78.75%), indicating that UNet++ performed best in predicting the spatial overlap between the mask and the actual landslide area; PSPNet was second, followed by DeepLab. V The 3+ version, due to its emphasis on the trade-off between global semantics and detail restoration, has slightly inferior multi-scale feature fusion performance.

[0110] In terms of boundary detail capture and error control: UNet++ (93.31%) and PSPNet (93.26%) are almost equal in precision, both outperforming DeepLab. V 3+ (88.36%) indicates that the first two methods are more effective at avoiding false alarms; DeepLab V Due to its dilated convolution and decoder design, 3+ focuses more on capturing large-scale context, resulting in more misclassifications of small targets or edge details.

[0111] In terms of missed detection risk and recall capability: PSPNet's recall (91.36%) is higher than UNet++ (89.31%) and DeepLab. V 3+ (87.30%), indicating that its pyramid pooling module has stronger false negative suppression capability in detecting landslide edges and fragmented areas; in contrast, DeepLab V Although 3+ aggregates richer global information, some small-scale landslides are still missed due to the speckle noise and complex background of SAR images.

[0112] Overall performance and F1 score: In terms of F1 score, PSPNet (92.20%) and UNet++ (90.96%) both significantly outperform DeepLab. V 3+ (87.72%), comprehensively reflecting their balance between precision and recall. DeepLab V While the 3+ model has advantages in multi-scale context modeling and lightweight design, its end-to-end decoder needs further optimization in terms of robustness to noise and scarce annotations on this SAR landslide extraction dataset.

[0113] Table 2 Comparison of landslide extraction results from deep learning network models

[0114]

[0115]

[0116] By comparing and analyzing the training and validation loss curves of three models, PSPNet, UNet++, and DeepLabV3+, Figures 6-8 Regarding training loss, UNet++ converged the fastest, with the loss value rapidly decreasing from 0.62 to 0.09; PSPNet showed the most stable decrease, with no significant oscillations throughout; while DeepLabV3+ exhibited an exponential decay trend. Validation loss performance differed significantly: DeepLabV3+ performed best, with a minimum loss of 0.18 and fluctuations of only ±0.04; UNet++ experienced a sudden increase in validation loss to 0.57 at the 40th epoch, indicating the highest risk of overfitting; PSPNet showed abnormal fluctuations in the mid-term but eventually stabilized at 0.24.

[0117] In terms of model characteristics, DeepLabV3+'s dilated convolutions and encoder-decoder structure effectively balance multi-scale features, demonstrating the best generalization ability. While UNet++'s dense skip connections enhance feature reuse, its regularization strategy needs optimization. PSPNet's pyramid pooling module requires improved feature fusion weight allocation. Optimization suggestion: DeepLabV3+ could try data augmentation;

[0118] UNet++ should introduce Dropout or Early Stopping; PSPNet can extend the training cycle and optimize the pooling layer weights. In practical applications, DeepLabV3+ is suitable for high-precision scenarios, UNet++ needs to focus on overfitting, and PSPNet needs to optimize the feature fusion mechanism.

[0119] In summary, DeepLabV3+ exhibits the most stable validation loss and strong generalization ability, but its mIOU and accuracy are slightly lower. PSPNet shows a steady decrease in training loss, but its validation loss fluctuates significantly in the mid-term, with all evaluation metrics falling in the middle. UNet++ shows the fastest decrease in training loss, but its validation loss rises significantly in the later stages, posing a risk of overfitting. However, it boasts the highest mIOU, accuracy, precision, recall, and F1 scores, demonstrating the strongest segmentation performance. Therefore, if model stability and generalization ability are prioritized, DeepLabV3+ is the best choice; if higher segmentation accuracy is required, UNet++ performs optimally after controlling for overfitting; and PSPNet is suitable for tasks requiring a higher level of global context understanding.

[0120] Experimental results using landslide data from dual-polarization SAR images show that UNet++ performs best in mIoU and Precision, while PSPNet slightly outperforms in Recall and F1, and DeepLab... V Due to its high dependence on global context, SAR imagery presents certain challenges in handling noise and small-scale targets.

[0121] PSPNet's training loss curve is the most stable with few fluctuations, showing that the pyramid pooling module has a natural ability to suppress noise in multi-scale feature extraction, making it suitable for tasks that require a higher level of global context understanding. Its stable training process and good robustness give it certain advantages when dealing with complex scenes.

[0122] UNet++ exhibits the fastest decrease in training loss, reflecting that its dense skip connections and deep supervision effectively accelerate feature learning, making it suitable for tasks with extremely high segmentation accuracy requirements. Although there is a risk of overfitting, its performance can be further improved by optimizing the regularization strategy.

[0123] DeepLabV3+ exhibits a significant exponential decay in training loss, decreasing rapidly in the initial stage and then leveling off, reflecting the progressive learning characteristics of dilated convolution and lightweight decoder. It is particularly suitable for SAR landslide extraction scenarios with weak annotations and complex noise. Its strongest generalization ability and stability make it perform exceptionally well in handling uncertainty and noise.

[0124] Future work could explore a hybrid architecture that integrates local and global feature interactions to further optimize the detection sensitivity of small landslides.

[0125] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for extracting landslide information from SAR images based on deep learning algorithms, characterized in that, include: Acquire SAR and optical images of the earthquake zone; The landslide boundaries marked in the optical image are mapped onto the SAR image to annotate the SAR image and obtain a labeled image. A dataset is constructed based on the original and labeled images of the SAR images; The pre-built semantic segmentation model is trained using the dataset; A segmentation map is generated based on a trained semantic segmentation model to extract the landslide area; Based on the original and labeled images of the SAR images, the dataset is constructed as follows: The original image is preprocessed by adjusting it to a uniform format and performing normalization. The SAR image is labeled pixel by pixel according to the landslide boundary, and the corresponding label is an 8-bit color image. Based on the original image and its corresponding label image, as well as the original image and its corresponding label image that have undergone simultaneous geometric transformation, training samples of a preset size are constructed to obtain the dataset; Training a pre-built semantic segmentation model using the dataset includes: The dual polarization features of the SAR image are stacked into two channels, and a blank channel is superimposed on a third channel to construct a multi-channel input; Based on the multi-channel input, the pre-built semantic segmentation model is trained until the loss function converges; The selection methods for the semantic segmentation model include: If the required segmentation accuracy is higher than a preset threshold, then UNet++ is determined as the semantic segmentation model. If the SAR image contains severe noise, DeepLabV3+ is determined as the semantic segmentation model. If the SAR image contains a complex scene, PSPNet is selected as the semantic segmentation model.

2. The method for extracting landslide information from SAR images based on deep learning algorithms according to claim 1, characterized in that, The target proportion in each image of the dataset is greater than a preset threshold, and the pixel block size of the training samples is 128×128.

3. The method for extracting landslide information from SAR images based on deep learning algorithms according to claim 1, characterized in that, The loss function is a composite loss function (BCEDiceLoss) that combines the binary cross entropy (BCE) and the Dice coefficient.

4. The method for extracting landslide information from SAR images based on deep learning algorithms according to claim 1, characterized in that, When PSPNet is determined to be the semantic segmentation model: For the input image, a proportionally scaled feature map is extracted through the backbone network; The pyramid pooling module (PPM) sequentially applies four different sizes of adaptive average pooling to the feature map. After each pooling branch, the number of channels is compressed to 1 / 4 of the original number through a 1×1 convolution. Then, bilinear interpolation is used to upsample the feature maps of each branch back to the same spatial resolution as the input feature map; The feature maps after multi-scale pooling and upsampling are concatenated with the original feature maps along the channel dimension, and the concatenated feature maps are then fused by convolution. The fused low-resolution feature map is mapped back to the original size of the input image to output the predicted segmentation map.

5. The method for extracting landslide information from SAR images based on deep learning algorithms according to claim 1, characterized in that, When DeepLabV3+ is determined as the semantic segmentation model, the DeepLabV3+ network consists of two parts: an encoder-ASPP module and a lightweight decoder. At the encoder end, the input image extracts shallow and deep features through the ResNet34 backbone network, and then the deep features are fed into the ASPP module. At the decoder, the deep semantic features output by the ASPP module are first upsampled to one-quarter of the original resolution using four times bilinear interpolation. In parallel, the high-resolution shallow features from the backbone network are compressed into channels by 1×1 convolution and then concatenated and fused with the upsampled deep features in the channel dimension. After one or two 3×3 convolutions for refinement, the images are finally restored to the original image size using four times interpolation to generate a pixel-level segmentation map.

6. The method for extracting landslide information from SAR images based on deep learning algorithms according to claim 1, characterized in that, When determining UNet++ as the semantic segmentation model: For the input image, the encoder extracts feature maps at five scales in sequence through continuous convolution and downsampling operations. The spatial size of these maps decreases progressively from 1 / 2, 1 / 4, 1 / 8, 1 / 16, to 1 / 32, thus constructing a multi-scale feature pyramid. The decoder uses normalization in each stage, and the channel configuration of the decoder corresponds to the encoder depth, which is (256, 128, 64, 32, 16) in sequence. The feature map output from each stage is restored to its size through upsampling until it finally matches the input resolution. Each decoding layer receives features from all preceding coding layers through skip connections. The decoder uses a 1×1 convolution to map the last-level features to a single-channel output and generates a landslide binary segmentation mask through Sigmoid activation.

Images