SAR image landslide information extraction method based on deep learning algorithm

Abstract

The invention relates to the technical field of landslide information extraction, in particular to an SAR image landslide information extraction method based on a deep learning algorithm. The method comprises the following steps: acquiring an SAR image and an optical image of a seismic region; mapping the landslide boundary marked in the optical image to the SAR image so as to mark the SAR image to obtain a label image; constructing a data set based on the original image and the tag image of the SAR image; training a pre-constructed semantic segmentation model by using the data set; and generating a segmentation image based on the trained semantic segmentation model to extract a landslide region. On the basis of the SAR image with rich polarization characteristics, the deep learning semantic segmentation network is used for landslide region segmentation, and the better segmentation precision is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide information extraction, and in particular to a method for extracting landslide information from SAR images based on a deep learning algorithm. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active imaging system that acquires target information by emitting microwave pulses and receiving backscattered echoes from the surface. Unlike visible light imaging, SAR records slant range data rather than the spatial distance between objects. In areas with significant terrain undulations, such as mountains, geometric distortions such as perspective contraction, shadows, overlaps, forward tilt of mountains, and top-bottom inversions are prone to occur. This makes the reflected signal complex and variable, making it difficult to directly reflect the contours and textures of objects in the image. Furthermore, SAR echoes are affected by multiple factors, including wavelength, angle of incidence, polarization, and the dielectric properties of the objects. Their imaging mechanism differs fundamentally from that of visible light, requiring specialized interpretation in conjunction with microwave scattering theory. These characteristics make the morphological and textural features of landslides in SAR images highly complex, making traditional visual interpretation methods inefficient and reliant on expert experience.

[0003] Traditional SAR image interpretation methods rely heavily on expert-designed scattering features and prior models. While these methods offer physical interpretability, feature engineering struggles to fully characterize the heterogeneity of landslides and their background environments. Furthermore, manual feature extraction and model tuning are time-consuming and labor-intensive, relying heavily 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 Haisi-1, a wealth of open-source satellite data and high-resolution imagery resources have become readily available at low cost. These diverse SAR data types provide a rich source of information for the timely extraction of landslide hazard information. Sentinel-1, with its free and open access policy and stable observation capabilities, combined with its global coverage and high-frequency revisit characteristics, offers a multidimensional 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 interpretation models that are adaptive to complex terrain and scattering mechanisms is crucial for improving landslide identification accuracy. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for extracting landslide information from SAR images based on a deep learning algorithm. 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] The present invention provides a method for extracting landslide information from SAR images based on a deep learning algorithm, which comprises:

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

[0007] Mapping the landslide boundary marked in the optical image to the SAR image to mark the SAR image to obtain a label image;

[0008] Constructing a data set based on the original image and the labeled image of the SAR image;

[0009] Using the dataset to train a pre-built semantic segmentation model;

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

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

[0012] Preprocessing the original image, adjusting the image to a unified format and performing normalization processing;

[0013] Annotate the SAR image pixel by pixel according to the landslide boundary, and use 8-bit color map for the corresponding label;

[0014] Based on the original image and the corresponding label image and the original image and the corresponding label image that are synchronously geometrically transformed, training samples of preset sizes are constructed to obtain a data set.

[0015] As a preferred embodiment, the target ratio in each image of the data set is greater than a preset threshold, and the pixel block size of the training sample 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 the third channel to construct a multi-channel input;

[0018] Based on the multi-channel input, a 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 binary cross entropy (Binary Cross Entropy, BCE) and Dice coefficient.

[0020] As a preferred embodiment, the semantic segmentation model is selected by:

[0021] If the required segmentation accuracy is higher than a preset threshold, 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 complex scenes, PSPNet is determined as the semantic segmentation model.

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

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

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

[0027] Then use bilinear interpolation to upsample each branch feature map back to the same spatial resolution as the input feature map;

[0028] The multi-scale pooled and up-sampled feature map is concatenated with the original feature map in the channel dimension, and convolution fusion is performed on the concatenated feature map;

[0029] The fused low-resolution feature map is mapped 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 an encoder-ASPP module and a lightweight decoder;

[0031] On the encoder side, the input image is passed through the ResNet34 backbone network to extract shallow features and deep features, and then the deep features are sent to the ASPP module;

[0032] On the decoder side, the deep semantic features output by the ASPP module are first upsampled to one-quarter of the original resolution using fourfold bilinear interpolation. In parallel, the high-resolution shallow features taken from the backbone network are compressed through a 1×1 convolution channel and then concatenated with the upsampled deep features in the channel dimension. The features are then refined through one or two 3×3 convolutions and finally restored to the original image size using fourfold interpolation to generate a pixel-level segmentation map.

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

[0034] For the input image, the encoder extracts feature maps of five scales in sequence through continuous convolution and downsampling operations. The spatial size of the feature maps is reduced step by step by 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 to construct a multi-scale feature pyramid.

[0035] The decoder uses normalization at each level. The decoder channel configuration corresponds to the encoder depth, which is (256, 128, 64, 32, 16). The feature map output by each level is restored to size (×2) through upsampling until it is consistent with the input resolution. Each decoding layer receives the features of all previous encoding layers through skip connections.

[0036] At the end of the decoder, 1×1 convolution is used to map the last level features to a single channel output, and Sigmoid activation is used to generate a landslide binary segmentation mask.

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

[0038] 1. Based on SAR images with rich polarization features, the present invention uses a deep learning semantic segmentation network to segment the landslide area, achieving better segmentation accuracy.

[0039] 2. In order to avoid the network segmentation effect being affected by the small number of images in the data set, the present invention performs data enhancement by performing synchronous geometric transformations such as rotation, horizontal flipping, and vertical flipping on the original images and label images, and expands the data set to increase sample diversity, alleviate the problem of sample scarcity, and meet the data volume requirements for deep learning model training.

[0040] 3. The present invention corresponds each polarization channel to an independent channel, each channel to a feature dimension, and can be cascaded in the channel direction to form a multi-channel input for the deep learning model to extract polarization complementary features. The stacking operation splices the two channels in the depth dimension (Channel Axis) to form an input tensor with a shape of H×W×2 (H / W is the spatial dimension). This operation retains the physical properties of the polarization scattering matrix, allowing the model to jointly learn co-polarization and cross-polarization features. By constructing multi-channel input, the deep network can simultaneously learn the complementary features and cross-channel associations under each polarization mode, thereby improving the accuracy and robustness of tasks such as classification, detection, and change monitoring. In addition, in order to utilize the pre-trained model, the present invention superimposes a blank channel on the third channel.

[0041] 4. Based on the constructed dataset and the multi-channel input of polarization feature stacking, the present invention evaluates the prediction performance of UNet++, DeepLabV3+ and PSPNet to determine the advantages of each semantic segmentation model, so that the semantic segmentation model can be selected according to the actual scenario requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the flow chart of landslide extraction by deep learning of SAR images in the present invention;

[0043] Figure 2 It is the PSPNet network model structure in the present invention;

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

[0045] Figure 4 It is the DeepLabV3+ network model structure in the present invention;

[0046] Figure 5 In this paper, some experimental results of PSPNet, UNet++, and DeepLabV3+ deep learning networks are presented;

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

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

[0049] Figure 8 The training and validation loss curves of the DeepLabV3+ model in this invention; DETAILED DESCRIPTION

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

[0051] Driven by deep learning technology, breakthroughs have been made in the accuracy of remote sensing image interpretation. To efficiently extract landslide information from dual-polarization SAR imagery, this example selected three representative deep learning semantic segmentation networks: PSPNet, UNet++, and DeepLabV3+. These networks were trained and validated on the Sentinel-1 dual-polarization (VV and VH) SAR dataset. The flowchart is shown in Figure 1. Convolutional neural networks (CNNs) and their variants significantly improve the efficiency of extracting landslide morphology, texture, and polarization 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 integrating contextual information at different scales. The UNet++ architecture, with its dense skip connections and deep supervision mechanism, effectively alleviates the problem of detail loss in medical image segmentation. 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 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 identification tasks.

[0052] The Pyramid Scene Parsing Network (PSPNet) is an improved semantic segmentation model based on the fully convolutional network (FCN). Its core innovation lies in the introduction of the Pyramid Pooling Module (PPM), which improves pixel-level classification accuracy by aggregating multi-scale contextual information. This module divides the input feature map into different sub-regions (such as 1×1, 2×2, 3×3, 6×6), and performs adaptive average pooling on each sub-region to capture global context and local details. It is then spliced ​​and fused with the original feature map through upsampling, effectively expanding the receptive field and solving the problem of insufficient scale adaptability caused by the fixed receptive field of traditional models. In addition, PSPNet uses ResNet34 as the backbone network and uses residual connections to alleviate gradient vanishing, thereby enhancing model convergence and feature expression capabilities.

[0053] The technical advantages of PSPNet are reflected in multi-scale feature fusion and efficient parameter design: 1) It integrates contextual information of different granularities through pyramid pooling operations, significantly improving the segmentation performance of targets with significant size differences in complex scenes, especially the segmentation of small objects; 2) It adopts a cascade fusion strategy to upsample multi-resolution feature maps to the 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 with local details, it enhances the robustness of the model to multi-scale targets in scenarios such as remote sensing imagery and autonomous driving.

[0054] Experiments demonstrate that this model, through layered 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 have made it widely applicable in areas such as scene understanding, high-resolution remote sensing image analysis, and autonomous driving environment 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. By introducing PPM and utilizing pooling kernels of varying sizes to capture contextual information at different scales, it achieves high-precision image segmentation. PSPNet's ability to extract global contextual information, provide high-precision segmentation and boundary detection, multi-scale processing capabilities, and efficient training and inference performance enable it to provide high-quality object detection results in a variety of practical applications.

[0056] In the semantic segmentation task, 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 fuser and classifier (such as Figure 2 ). 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 and uses 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 each branch feature map, and channel splicing and 1×1 convolution kernels are used for cross-scale feature fusion to generate enhanced composite features. The classifier is a pixel-level semantic decoder that converts the fused features into the category space through pixel-by-pixel convolution mapping. Finally, it outputs the semantic probability distribution of each pixel through softmax normalization to complete the scene parsing task. Through the cascaded multi-scale feature extraction and fusion mechanism, this architecture significantly improves the semantic segmentation accuracy in complex scenes.

[0057] The feature extractor uses ResNet34 as the backbone. Its core residual architecture effectively mitigates the vanishing gradient and performance degradation issues associated with deep network training by introducing skip connections. Specifically, each residual module consists of two 3×3 convolutional layers, and employs a 1×1 convolutional approach that "upscales first, then reduces" the number of channels. This reduces the number of parameters by approximately 30% while maintaining feature expressiveness. By directly adding the input features to the convolution output to form a residual mapping, the residual architecture allows the network to focus on learning the difference function between the input and the target. This design enables network depth to be expanded to tens or even hundreds of layers, while only increasing training complexity sublinearly, while ensuring efficient transmission of underlying features.

[0058] The purpose of the pyramid pooling module is to fuse features of different scales to capture contextual information of different sizes, thereby obtaining more comprehensive global information. This module adopts 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 an average pooling operation on each sub-region. After compressing the number of channels through 1×1 convolution, the four sets of pooled features of different scales are upsampled to the original size through bilinear interpolation, ultimately obtaining four sets of feature maps with 1 / 4 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 detail features, effectively solving the problem of contextual information loss caused by single-scale pooling.

[0059] The feature fuser achieves the complementary fusion of deep semantic features and shallow detail features by constructing feature pyramids at different levels. First, the four-level pooled features output by PPM are upsampled so that their spatial size matches the intermediate feature map output by the backbone network. Subsequently, a splicing operation is performed along the channel dimension (channel-last) to generate a composite feature tensor containing multi-scale contextual information. To unify the feature dimension, 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 low-level feature maps. Finally, the semantic alignment of high-level and low-level features is achieved through channel splicing, enabling the network to simultaneously utilize deep abstract features and shallow spatial details, significantly improving the semantic segmentation accuracy in complex scenes.

[0060] As the core component of the decoding stage, the classifier uses 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 heat map corresponding to the number of target categories. 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 does not require complex operations such as deconvolution. While maintaining inference efficiency, the spatial detail information retained through jump connections effectively alleviates the boundary blurring caused by interpolation. The final pixel-level classification result is 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 morphological features, significant scale differences, and strong background interference in SAR imagery, the Pyramid Scene Parsing Network (PSPNet) first uses a ResNet34 backbone to stably transmit features. A pyramid pooling module aggregates global and local contextual information at multiple scales to enhance the spatial semantic understanding of landslide areas. Subsequently, a cross-level feature fusion strategy is combined to concatenate semantic features at 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 is used to achieve efficient pixel-level classification, achieving a balance between computational efficiency and segmentation accuracy. This method excels in extracting landslides from high-resolution SAR imagery, combating noise and complex terrain interference, and provides reliable technical support for emergency response to earthquake landslide disasters.

[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 dense skip connections. Compared to U-Net's single skip connection, UNet++ densely connects the encoder and decoder sub-networks at each level to form a nested topology, effectively narrowing the expression gap between shallow detail features and deep semantic features. Drawing on the dense connection concept of DenseNet, this model enables direct interaction between features at different levels during the decoding process, enhancing the ability to integrate multi-scale contextual information and significantly improving the accuracy of capturing complex scene boundaries and subtle structures.

[0063] In terms of structural design, UNet++ constructs a four-layer nested subnetwork by filling the "hollow" skip paths of U-Net. Each subnetwork consists of encoding-decoding units and is densely connected to adjacent subnetworks via upsampling paths. This hierarchical design allows for gradual refinement of features during cross-scale transfer, while integrating semantic information at different resolutions through a feature stacking strategy. Furthermore, UNet++ supports dynamic pruning, which rapidly simplifies the model structure by removing deep subnetworks. This reduces computational complexity while maintaining segmentation accuracy, achieving a flexible trade-off between accuracy and efficiency.

[0064] During training, UNet++ employs a multi-level joint optimization strategy, jointly updating model parameters by weightedly fusing the loss functions of each sub-network. This design not only enhances the consistency of feature representation but also allows for configurability during model deployment: users can choose to retain sub-network depth or remove redundant layers through pruning, adapting to different application scenarios with varying computational resource constraints. Experiments demonstrate that UNet++ demonstrates superior boundary preservation and robustness in image segmentation tasks, surpassing traditional U-Net, making it a benchmark model in this field.

[0065] Based on the classic U-Net, UNet++ significantly enhances the fusion of multi-scale features and the ability to express details through nested sub-networks and dense skip connections. Unlike the traditional U-Net, which only has a single skip connection to pass the encoder features of the same layer to the decoder, UNet++ introduces multiple continuous convolution blocks between each layer of encoder and decoder, and connects these convolution blocks in series with dense connections to form a nested U-net structure. Figure 3 This design not only expands the effective receptive field of the network, enabling multiple interactions between low-level features and high-level semantics, 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-subnetwork) joint training strategy: each decoding subnetwork receives features from encoders at different depths, and an independent supervision branch is set up at its end to jointly optimize network performance. This multiple 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 subnetworks based on task requirements, achieving different trade-offs between accuracy and efficiency.

[0067] In terms of network structure, both the encoder and decoder of UNet++ use 3×3 convolutions with a stride of 1. The encoding process is supplemented by 2×2 max pooling with a stride of 2 for downsampling. The decoding process restores spatial resolution through interpolation with an upsampling factor of 2 and a stride of 2. Through the synergy of short connections, upsampling and downsampling, 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 and medical images, making it 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 path. Its design motivation is to solve the gradient attenuation problem in deep feature transfer. The specific splicing process can be performed according to the following formula:

[0069]

[0070] Among them, the network nodes use (i, j) coordinates for hierarchical identification (corresponding to Figure 3 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 consisting of two consecutive convolutional layers, each followed by an activation function. For example, we can see Figure 3 It is the top convolution block in the U-net++ structure. 1,0 ) is X 1,0 The result is upsampled once, X 0,1 Is U(X 1,0 ) and X 0,0 The result of convolution after concatenation. Similarly, X 0,4 Is U(X 1 ,3 ) and X 0,0 、X 0,2 、X 0,3 The convolution result after concatenation. The number of convolution kernels is adjusted to make the image sizes of the five convolution blocks in the first layer completely consistent.

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

[0072] In view of the diverse morphology of landslides in SAR images (significant differences in shape, size, and texture), UNet++ significantly improves the model's adaptability and generalization capabilities by leveraging its densely nested structure and multi-level feature fusion. Its dense connection design not only improves the efficiency of gradient propagation, but also enhances multi-scale perception by integrating feature maps with different receptive fields. It can still accurately extract landslide area boundaries and texture details under complex background interference. UNet++ achieves cross-level feature complementarity through nested topological networks. While ensuring the integrity of feature expression, it overcomes the problems of traditional methods such as sensitivity to landslide scale and blurred boundaries, providing a solution that combines both accuracy and efficiency for high-resolution remote sensing image segmentation tasks.

[0073] Semantic segmentation faces two main core challenges: insufficient adaptability to multi-scale targets and blurred boundaries caused by loss of feature resolution. The former manifests itself as a single-scale convolution kernel having difficulty capturing the effective features of both large and small objects simultaneously, while the latter stems from the fact that multiple downsampling operations cause the resolution of feature maps to decrease, resulting in reduced prediction accuracy and loss of detailed information (such as edges). Although DeepLabV1 and V2 expand the receptive field and support multi-scale feature fusion by introducing dilated convolution and spatial pyramid pooling (ASPP) modules, limitations still exist. Dilated convolution with a single dilation rate has difficulty covering a wide range of target scales, the fusion of global and local contextual information is inefficient, and deep networks lack the ability to parse edge details.

[0074] DeepLab V 3+ is an end-to-end improved semantic segmentation model, whose core innovation lies in the integration of encoder-decoder architecture and multi-scale feature extraction mechanism. The encoder uses ResNet34 as the backbone network, extracts high-semantic features through deep convolution, and combines ASPP (including dilated convolutions with different expansion rates and global pooling branches) to achieve parallel capture of multi-scale contextual information; the decoder uses bilinear upsampling to fuse deep semantic features with shallow high-resolution features output by the encoder across levels, and uses 3×3 convolution to optimize feature expression and gradually restore it to the original image resolution. This model significantly reduces computational complexity while maintaining high segmentation accuracy through lightweight design (such as depthwise separable convolution). It is particularly good at handling complex scenes with significant differences in target scales and blurred boundaries, and has demonstrated strong robustness and real-time 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 then shallow and deep features are output. The shallow features are directly passed to the decoder and compressed through a 1×1 convolution channel to reduce the proportion of redundant information; the deep features are input into the ASPP module for multi-scale context enhancement - this module performs one 1×1 convolution, three 3×3 void convolutions with different expansion rates (6 / 12 / 18) and a global average pooling operation in parallel, and then integrates multi-branch features through a 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 context information and dense feature expression.

[0076] The backbone network module uses the ResNet34 model as the basic feature extractor. This design fully utilizes the advantages of ResNet34 in image feature representation. Specifically, ResNet34 effectively alleviates the gradient vanishing problem in deep neural networks through its unique residual learning framework, enabling the network to build a deeper hierarchical structure to capture multi-scale contextual information. V In 3+, the convolutional layers of ResNet34 are reconstructed as dilated separable convolutions—replacing standard convolutions with a combination of depthwise separable convolutions and dilated convolutions. This design reduces the number of parameters through depthwise convolutions, while utilizing the dilation mechanism to expand the receptive field, capturing multi-scale contextual information while maintaining high-resolution feature maps. In addition, dilated convolutions are used in the deep layers of ResNet34 to replace traditional downsampling operations, avoiding excessive loss of feature map resolution and thus improving sensitivity to object boundaries. The improved ResNet34 backbone network strikes a balance between computational efficiency and feature expression capabilities, not only adapting to the needs of intensive prediction tasks, but also providing a high-resolution, multi-granularity feature foundation for the global semantic fusion of the subsequent ASPP module, ultimately significantly enhancing the model's adaptability to complex scenarios.

[0077] The ASPP module achieves efficient fusion of multi-scale contextual features through a multi-branch parallel architecture. Its core design comprises four key components: First, it employs parallel atrous convolutions with varying dilation rates, enabling different branches to cover receptive fields ranging from local to wide areas, simultaneously capturing object details and large-scale context associations. Second, it introduces a global average pooling branch that compresses feature maps to a global context vector and restores 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 essential features and reduce computational redundancy through channel dimensionality reduction. The feature maps output by all branches are concatenated along the channel dimension, integrated through 1×1 convolutions, and then subjected to nonlinear mapping and feature calibration using batch normalization (BN) and ReLU activation functions. Through an adaptive weight allocation mechanism, this module enables the model to accurately localize small object details while effectively associating large-scale context in complex scenes, significantly improving its robustness to scale-varying objects.

[0078] The decoder module is designed to solve the problem of feature map resolution degradation and boundary detail loss caused by multiple downsampling of the encoder, and to achieve refined segmentation through multi-level feature fusion and progressive upsampling. Specifically, the low-resolution deep features output by the DeepLabv3+ encoder (with multi-scale semantic information enhanced by the ASPP module) are first upsampled 4 times by bilinear interpolation to restore some spatial details; then they are cross-level spliced ​​with the high-resolution low-level features extracted by the shallow layer of the encoder. In order to avoid the influence of shallow feature channel redundancy on the fusion effect, the decoder uses 1×1 convolution to reduce its dimensionality (such as compressing the number of channels from 256 to 48), and optimizes the fused features through 3×3 convolution combined with BN and ReLU activation functions, and finally outputs a segmentation result with the same resolution as the original image after 4 times upsampling.

[0079] DeepLab V The core innovation of 3+ lies in its organic integration of ASPP with a lightweight encoder-decoder architecture. On the encoder side, ASPP employs dilated convolutions with different dilation rates and global average pooling branches in parallel to effectively expand the receptive field and enhance multi-scale context modeling. On the decoder side, 1×1 convolutions compress shallow high-resolution feature channels, which are then skip-connected to the deep semantic features processed by the ASPP module. Bilinear upsampling is then used to gradually restore spatial details, significantly improving boundary segmentation accuracy. Simultaneously, depthwise separable convolutions are extensively used in the optional ResNet34 backbone network, ASPP, and decoder modules, maintaining high expressive power while significantly reducing the number of parameters and computation, achieving real-time inference performance. This design retains the advantage of dilated convolutions in capturing contextual information without reducing resolution, while restoring image details through a simple and efficient decoder. This results in excellent performance in high-precision semantic segmentation tasks, particularly in scenarios requiring fine boundary recognition.

[0080] DeepLab V 3+ utilizes the ASPP module to concurrently collect contextual information at different scales, depthwise separable convolution to reduce computational overhead, and a lightweight encoder-decoder architecture with cross-layer skip connections and channel compression. This seamlessly integrates multi-scale feature extraction and noise suppression, enabling the model to flexibly adapt to the diverse morphologies of landslide targets while accurately delineating landslide boundaries. Its end-to-end design maintains high segmentation accuracy while also balancing computational efficiency, providing an accurate, practical, and real-time solution for landslide hazard monitoring in complex geological environments.

[0081] This embodiment first constructs a Luding earthquake landslide sample set based on the VV and VH dual-polarization SAR images acquired by the Sentinel-1A satellite, and uses a deep learning semantic segmentation method to extract and analyze the post-earthquake landslide area. Specifically, the Sentinel-1A image of the experimental area on September 14, 2022, after the earthquake, was selected, with a spatial resolution of 5m×20m. Considering that polarimetric SAR data lacks visually intuitive ground object information, sample zoning presents great challenges. In this study, the post-earthquake optical images of the Sentinel-2 satellite were used for auxiliary annotation. By aligning with the Sentinel-2 optical images of the same period, the manually annotated landslide boundaries in the optical image were mapped to the SAR image to ensure the accuracy of the training sample area.

[0082] When constructing the sample, systematic tests were conducted with different pixel block sizes (e.g., 64×64, 128×128, and 256×256). The results showed that too small a pixel block can lead to insufficient information, affecting classification accuracy, while too large a pixel block introduces significant redundancy, increasing the computational burden of the model. A comprehensive comparison of classification performance and training efficiency revealed that a 128×128 pixel block performed best in the landslide and non-landslide segmentation task, fully capturing local structural features while maintaining high computational efficiency, providing a solid foundation for subsequent model training.

[0083] During the semantic segmentation model training process, the dataset is divided into two parts. The first part is the original image. To facilitate training, the image format is uniformly adjusted to .tif and normalized at the input. The second part is the labeled image. The pixel-by-pixel labels are based on 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 transmission-horizontal reception (HH) and horizontal transmission-vertical reception (HV)), providing rich feature information. Each polarization channel corresponds to an independent channel, and each channel corresponds to a feature dimension. They can be cascaded in the channel direction to form a multi-channel input for the deep learning model to extract polarization complementary features. The stacking operation splices the two channels in the depth dimension (Channel Axis) to form an input tensor with a shape of H×W×2 (H / W is 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 multi-channel input, the deep network can simultaneously learn complementary features and cross-channel associations under each polarization mode, thereby improving the accuracy and robustness of tasks such as classification, detection, and change monitoring.

[0085] When performing landslide extraction and segmentation, this embodiment stacks the dual-polarization features into two channels. In order to balance the data set, the data is randomly sampled to ensure that the target ratio in each image is greater than 0.3 to prevent the model from overfitting. In order to use the pre-trained model, a blank channel is superimposed on the third channel. The PSPNet, UNet++, and DeepLabV3+ deep learning networks are operated respectively. Some experimental results are shown in the figure. Figure 5 By adjusting the learning rate and the number of training times, the accuracy of the model can be gradually improved.

[0086] The core architectural innovation of PSPNet lies in the PSP module, which enhances the network's ability to perceive multi-scale objects through a multi-level context aggregation mechanism. The PSPNet model structure can be divided into two parts: a backbone feature extraction network that extracts image features through downsampling; and an enhanced feature extraction structure that extracts multi-scale features, then upsamples and restores the image size to 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. The 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 to the feature map in sequence - the image is divided into 1×1, 2×2, 3×3, and 6×6 small blocks and pooled separately - to capture global and local contextual information. Each pooling branch is then compressed to 1 / 4 of the original number of channels through a 1×1 convolution, 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 spliced ​​with the retained original feature parts in the channel dimension and further fused through subsequent convolutions, making the model sensitive to detailed textures while maintaining a good perception of the overall structure, thereby completing high-quality feature extraction.

[0088] In this experiment, PSPnet used a ResNet34 backbone network with an encoder depth of three levels. A convolutional network was used to extract a multi-scale feature pyramid. The encoder then compressed the feature maps layer by layer through a series of 3×3 convolutions and downsampling operations, outputting feature maps at three spatial resolutions: 1 / 2, 1 / 4, and 1 / 8 of the input size. To accelerate convergence and improve feature representation quality, the encoder was initialized with weights pre-trained on ImageNet.

[0089] After the pyramid pooling module (PPM), the decoder first batch normalizes the number of channels (512) of the multi-scale feature map to improve training stability and suppress overfitting, and sets a dropout rate of 0.2 in the spatial pyramid branch to enhance robustness. PPM contains four parallel adaptive average pooling branches (1×1, 2×2, 3×3, 6×6 grids). After the output of each branch is reduced in dimension by 1×1 convolution, it is upsampled to the original feature map size by bilinear interpolation and spliced ​​with the third-level features of the encoder along the channel direction. It is then compressed by 1×1 convolution to achieve multi-scale fusion of global-local context. The decoder performs a series of convolution and upsampling operations, and the training and validation loss curves are shown as follows. Figure 6 shown.

[0090] Based on the core idea of ​​the UNet encoding-decoding symmetric structure and skip connection, UNet++ innovatively introduces the dense connection mechanism of DenseNet to build a nested multi-level feature fusion architecture, achieving an exponential expansion of feature expression. The encoder part consists of multiple downsampling modules, which extract low-resolution semantic features of the image step by step. The decoder adopts a nested branch design. Each decoding layer receives the features of all previous encoding layers through skip connections, and upsamples step by step to restore the resolution. In the task of landslide extraction and semantic segmentation, its nested connection enables the model to perceive local details (such as landslide boundary texture in SAR images) and global context (such as regional geological structure background) at the same time, and can better fuse feature information at different levels, significantly improving the semantic segmentation accuracy and feature expression capabilities in complex scenes.

[0091] In this experiment, UNet++ uses ResNet34 as its backbone network, leveraging its residual structure to effectively mitigate vanishing gradients and enhance feature expression. The encoder extracts feature maps at five scales through continuous convolution and downsampling operations, gradually reducing their spatial size by 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32. This constructs a multi-scale feature pyramid to meet the multi-scale fusion requirements of the decoder. To accelerate convergence and improve feature quality, the encoder is initialized with weights pre-trained on ImageNet.

[0092] The decoder uses normalization at each level to improve training stability, suppress overfitting and accelerate convergence. The channel configuration of the decoder corresponds to the depth of the encoder, which is (256, 128, 64, 32, 16) in sequence. The feature map output at each level is restored to size (×2) through upsampling operations until it is finally consistent with the input resolution. In each decoding unit, the input of the current layer not only fuses the jump connections from the same layer and deeper layer encoders, but also forms a dense connection with the output of the previous decoding layer; then cross-layer feature fusion is completed through convolution and upsampling. At the end of the decoder, 1×1 convolution is used to map the last level of features to a single channel output, and Sigmoid activation is used to generate a landslide binary segmentation mask to meet the requirements of the two-category segmentation task. The training and verification loss curves are as follows: Figure 7 shown.

[0093] As an improved semantic segmentation model, UNet++ demonstrates significant advantages in remote sensing image object classification and target extraction tasks with its dense skip connection structure and multi-scale feature fusion mechanism. The model enhances feature expression through nested decoder paths, effectively improving segmentation accuracy in complex scenarios. However, its dense connection mechanism leads to a high number of parameters and computational complexity, places strict requirements on hardware resources and the completeness of annotated data, and there is still a risk of losing details in the segmentation of small targets. In response to the challenges of SAR remote sensing images with significant multi-scale characteristics, complex spatial structure and high resolution, UNet++'s hierarchical feature aggregation capability can adaptively capture object information of different granularities, thereby optimizing segmentation efficiency and accuracy. In practical applications, it is necessary to balance model performance and computational cost, and adjust network depth and parameter configuration based on task requirements.

[0094] The innovative architecture of UNet++ reconstructs the skip path between the encoder and decoder and introduces dense convolution blocks to bridge the feature semantic gap, thereby reducing the complexity of the optimization process. At the same time, the deep supervision mechanism embeds auxiliary loss functions in each stage of the decoder, forcing the model to learn multi-level discriminant features, significantly improving the segmentation consistency of multi-scale landslides. To verify its effectiveness, this embodiment conducts comparative experiments based on the Luding earthquake landslide SAR image dataset. UNet++ performs better in the fineness of landslide area boundaries and the detection of small-area landslides. This illustrates the effectiveness of nested skip paths for remote sensing landslide segmentation, and also provides a reference for the subsequent use of appropriate pruning or lightweight transformation in resource-constrained environments.

[0095] DeepLab VThe 3+ network consists of two parts: an encoder-ASPP module and a lightweight decoder. On the encoder side, the input image first passes through a ResNet34 backbone network to extract shallow and deep features. The deep features are then fed into the ASPP module, which consists of multi-branch dilated convolutions (commonly used with dilation rates of 6, 12, and 18) and global average pooling. These are then concatenated and fused using 1×1 convolutions for dimensionality reduction. This ensures that multi-scale context information is captured while maintaining resolution.

[0096] On the decoder side, the deep semantic features output by ASPP are first upsampled to one-quarter of their original resolution using fourfold bilinear interpolation, significantly restoring spatial detail. Simultaneously, shallow, high-resolution features from the backbone network are compressed using 1×1 convolutions before being concatenated and fused with the upsampled deep features in the channel dimension. They are then refined using one or two 3×3 convolutions and finally restored to the original image size using fourfold interpolation to generate a pixel-level segmentation map. This design balances wide-area context capture with detailed boundary recovery, demonstrating high accuracy and excellent real-time performance in scenarios requiring fine boundaries, such as remote sensing landslide extraction.

[0097] In the experiment based on the SAR image landslide dataset, DeepLab V The 3+ model significantly improves the performance of landslide extraction through multi-level structural optimization: the backbone network adopts ResNet34, and its residual connection design enhances the feature reuse capability through jump gradient propagation, effectively alleviating the gradient degradation problem in deep network training, especially in complex terrain areas, it can retain high-resolution features and support the refined segmentation of landslide boundaries; the ASPP module integrates multi-scale hole convolution (expansion rate 6 / 12 / 18) with the global average pooling branch to simultaneously capture local scattering anomalies (such as deformation of the trailing edge of the landslide body) and wide-area terrain correlation, and combines depthwise separable convolution to reduce computational complexity, suppress SAR image speckle noise interference, and improve the robustness of multi-scale landslide detection; the decoder uses a cross-level feature fusion mechanism to splice the deep semantic features with the shallow high-resolution features after bilinear upsampling, optimizes the boundary response using 3×3 convolution, and restores details in combination with a progressive upsampling strategy, achieving sub-pixel segmentation accuracy under weak annotation conditions, providing an efficient and robust end-to-end solution for landslide monitoring in complex terrain scenarios. The training and verification loss curves are shown in Figure 2. Figure 8 shown.

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

[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 jump-connected to the ASPP output, achieving complementary spatial details and semantic information through cascaded weighted fusion. The fused features are then subjected to 3×3 convolution to refine boundary information, and then bilinear interpolation is used to restore the spatial resolution to 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 of 1×1 convolution, mapping the 256-dimensional features to a single channel (number of categories = 1) and generating a binary landslide segmentation mask using a sigmoid activation function. This configuration addresses the dual requirements of accuracy and efficiency in remote sensing image segmentation tasks.

[0100] Because deep learning model training requires a large amount of data, to mitigate the impact of a small dataset on network segmentation performance, data augmentation was performed by rotating, horizontally flipping, and vertically flipping the original and labeled images. This resulted in a dataset of 5,000 dual-polarization original SAR images and labeled images. This dataset was randomly divided into a 70% training set, a 15% test set, and a 15% validation set. The training set was used to train model parameters, the validation set was used to evaluate the model's generalization after each batch of training and hyperparameter adjustments, and the test set was used to assess the network's final segmentation results. This data augmentation method increases the sample size and improves network model training and generalization.

[0101] The three segmentation models proposed in this example are all implemented using the PyTorch deep learning framework. Model training uses the stochastic gradient descent Adam optimizer to optimize the classification network. This optimizer combines momentum with an adaptive learning rate approach, adjusting the learning rate of each parameter by calculating the first-order moment (mean) and second-order moment (variance) of the gradient, thereby accelerating training and improving convergence. The initial learning rate, lr, is 0.0001, and the poly strategy is used to dynamically adjust the learning rate. Weight_decay = 0.01, and the batch size is set to 4. Each cycle randomly selects 3,500 samples (128 × 128 pixels) 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 indicator for measuring the difference between the model prediction and the true label. During the training process, the model parameters are updated by minimizing the function, thereby achieving good generalization of new data. This embodiment adopts a composite loss function (BCEDiceLoss) that combines binary cross entropy (BCE) and the Dice coefficient to integrate the advantages of pixel-level classification accuracy and overall regional overlap. In particular, in the landslide extraction of binary SAR images with imbalanced foreground and background, the foreground (landslide area) often occupies a small number of pixels, and the background noise is strong and easily confused. BCEDiceLoss can significantly improve the model's detection accuracy for landslide boundaries and small landslide bodies, reducing false alarms and missed detections. The mathematical form of the composite loss function is defined as follows:

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

[0106] Where α and β are weight coefficients (equal weights are usually used, i.e. α = β = 0.5). BCE focuses on the classification accuracy of each pixel and can provide a stable gradient signal for fine-grained errors. Dice Loss pays more attention to the overall shape consistency and regional level matching. It is particularly effective for situations where foreground pixels are sparse or categories are unbalanced. It can effectively alleviate the problem of small targets or a few categories being ignored. Usually α = β = 0.5 is taken to balance pixel-level and regional-level errors. BCEDiceLoss considers both local classification accuracy and global regional overlap, which helps the model remain robust when the foreground and background are unbalanced.

[0107] This example 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 Intersection over Union (IoU) calculations for each category, measuring the average degree of overlap between the predicted segmented area and the true segmented area. Its value ranges from 0 to 1, with higher values ​​indicating more accurate segmentation results. Accuracy reflects the proportion of pixels correctly classified by the model as a whole (all pixels). Precision focuses on the proportion of true landslide pixels among the results determined by the model as "positive examples," reflecting the reliability of the predicted positive examples. A higher Precision means a lower false positive rate. Recall measures the proportion of all true landslide pixels successfully detected by the model, reflecting the model's detection capability. A high Recall indicates fewer missed detections, meaning the model can identify most true landslide areas. The F1 Score comprehensively considers the model's positive detection performance by taking the harmonic average of Precision and Recall.

[0108] In the task of landslide extraction from SAR images, PSPNet, UNet++ and DeepLabV3+ models all showed their respective advantages in multiple evaluation indicators (as shown in Table 2). Overall, UNet++ performed best in mIoU and Precision, PSPNet was slightly better in Recall and F1, and DeepLab V 3+ is highly dependent on global context, so it faces challenges in dealing with noise and small-scale objects in SAR images. The specific challenges are as follows:

[0109] In terms of overall segmentation accuracy and area coverage: in terms of mIoU index, UNet++ (86.05%) is slightly higher than PSPNet (84.05%), while DeepLab V 3+ is relatively low (78.75%), indicating that UNet++ is the best in terms of spatial overlap between the predicted mask and the real landslide area; PSPNet is second, and DeepLab V 3+ emphasizes the trade-off between global semantics and detail recovery, so its multi-scale feature fusion effect is slightly inferior.

[0110] In terms of boundary detail capture and error control: In terms of Precision, UNet++ (93.31%) and PSPNet (93.26%) are almost on par, both outperforming DeepLab. V 3+ (88.36%), indicating that the first two are better at avoiding false positives; DeepLab V 3+, due to its dilated convolution and decoder design, focuses more on capturing large-scale context, resulting in more misclassifications of small objects 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%), which shows that its pyramid pooling module has a stronger ability to suppress missed detection in detecting landslide edges and fragmented areas; 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: On F1Score, PSPNet (92.20%) and UNet++ (90.96%) are significantly better than DeepLab V 3+ (87.72%), which comprehensively reflects their balance between precision and recall. DeepLab V Although 3+ has the advantages of multi-scale context modeling and lightweight in model architecture, its end-to-end decoder's robustness to noise, scarce annotations and other issues needs to be further optimized in this SAR landslide extraction dataset.

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

[0114]

[0115]

[0116] By comparing and analyzing the training and validation loss curves of the three models PSPNet, UNet++ and DeepLabV3+ ( Figure 6-Figure 8 ). In terms of training loss, UNet++ converged the fastest, rapidly decreasing from 0.62 to 0.09. PSPNet declined most steadily, with no significant fluctuations throughout. DeepLabV3+ exhibited an exponential decay trend. Validation loss performance varied significantly: DeepLabV3+ performed best, reaching a minimum of 0.18 with a fluctuation of only ±0.04. UNet++'s validation loss suddenly spiked to 0.57 at epoch 40, indicating the highest risk of overfitting. PSPNet experienced unusual fluctuations mid-term but ultimately stabilized at 0.24.

[0117] In terms of model characteristics, DeepLabV3+'s dilated convolution and encoder-decoder architecture effectively balance multi-scale features, demonstrating optimal generalization. UNet++'s dense skip connections enhance feature reuse, but regularization strategies need to be optimized. PSPNet's pyramid pooling module needs improvements in feature fusion weight distribution. Optimization suggestion: DeepLabV3+ can experiment with data augmentation.

[0118] UNet++ should introduce dropout or early stopping; PSPNet can extend the training cycle and optimize the weights of the pooling layer. In practical applications, DeepLabV3+ is suitable for scenarios requiring high precision, while UNet++ needs to focus on overfitting, and PSPNet needs to optimize its feature fusion mechanism.

[0119] In summary, DeepLabV3+ has the most stable validation loss and strong generalization capabilities, but its mIOU and Accuracy are slightly lower. PSPNet's training loss decreases steadily, but its validation loss fluctuates significantly in the middle, with all evaluation metrics falling in the middle. UNet++ has the fastest decline in training loss, but its validation loss increases significantly in the later stages, which may pose a risk of overfitting. However, its mIOU, Accuracy, Precision, Recall, and F1 values ​​are the highest, and its segmentation performance is the strongest. In summary, if you are looking for model stability and generalization capabilities, DeepLabV3+ is the best choice. If you need higher segmentation accuracy, UNet++ performs best after controlling overfitting. PSPNet is suitable for tasks that require a higher level of global context understanding.

[0120] The experimental results of dual-polarization SAR image landslide data show that UNet++ performs best in mIoU and Precision, PSPNet is slightly better in Recall and F1, and DeepLab V 3+ Due to its high dependence on global context, it poses certain challenges to noise and small-scale targets in SAR images.

[0121] PSPNet's training loss curve is the most stable, with little fluctuation. This shows that the pyramid pooling module has a natural noise suppression effect in multi-scale feature extraction, making it suitable for tasks that require higher global context understanding. Its stable training process and good robustness give it certain advantages when dealing with complex scenes.

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

[0123] DeepLabV3+, on the other hand, exhibits a clear exponential attenuation of training loss, which decreases rapidly at the beginning and then levels 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 enable it to perform excellently when dealing with uncertainty and noise.

[0124] Future work can explore hybrid architectures that integrate local-global feature interactions to further optimize the detection sensitivity of small landslides.

[0125] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for extracting landslide information from SAR images based on a deep learning algorithm, characterized in that: include: Acquire SAR and optical images of the earthquake area; Mapping the landslide boundary marked in the optical image to the SAR image to mark the SAR image to obtain a label image; Constructing a data set based on the original image and the labeled image of the SAR image; Using the dataset to train a pre-built semantic segmentation model; A segmentation map is generated based on the trained semantic segmentation model to extract the landslide area.

2. The method for extracting landslide information from SAR images based on a deep learning algorithm according to claim 1, wherein: Constructing a training sample of a preset size based on the original image and the labeled image of the SAR image includes: Preprocessing the original image, adjusting the image to a unified format and performing normalization processing; Annotate the SAR image pixel by pixel according to the landslide boundary, and use 8-bit color map for the corresponding label; Based on the original image and the corresponding label image and the original image and the corresponding label image that are synchronously geometrically transformed, training samples of preset sizes are constructed to obtain a data set.

3. The method for extracting landslide information from SAR images based on a deep learning algorithm according to claim 1, wherein: The target ratio in each image of the data set is greater than a preset threshold, and the pixel block size of the training sample is 128×128.

4. The method for extracting landslide information from SAR images based on a deep learning algorithm according to claim 1, wherein: 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 the third channel to construct a multi-channel input; Based on the multi-channel input, a pre-built semantic segmentation model is trained until the loss function converges.

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

6. The method for extracting landslide information from SAR images based on a deep learning algorithm according to claim 1, wherein: The selection method of the semantic segmentation model includes: If the required segmentation accuracy is higher than a preset threshold, 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 complex scenes, PSPNet is determined as the semantic segmentation model.

7. The method for extracting landslide information from SAR images based on a deep learning algorithm according to claim 6, wherein: When PSPNet is determined as the semantic segmentation model: For the input image, a proportionally reduced feature map is extracted through the backbone network; The pyramid pooling module (PPM) applies four different sizes of adaptive average pooling to the feature map in sequence. After each pooling branch, the number of channels is compressed to 1 / 4 of the original through 1×1 convolution. Then use bilinear interpolation to upsample each branch feature map back to the same spatial resolution as the input feature map; The multi-scale pooled and up-sampled feature map is concatenated with the original feature map in the channel dimension, and convolution fusion is performed on the concatenated feature map; The fused low-resolution feature map is mapped to the original size of the input image to output the predicted segmentation map.

8. The method for extracting landslide information from SAR images based on a deep learning algorithm according to claim 6, wherein: When DeepLabV3+ is determined as the semantic segmentation model, the DeepLabV3+ network includes an encoder-ASPP module and a lightweight decoder; On the encoder side, the input image is passed through the ResNet34 backbone network to extract shallow features and deep features, and then the deep features are sent to the ASPP module; On the decoder side, the deep semantic features output by the ASPP module are first upsampled to one-quarter of the original resolution using fourfold bilinear interpolation. In parallel, the high-resolution shallow features taken from the backbone network are compressed through a 1×1 convolution channel and then concatenated with the upsampled deep features in the channel dimension. The features are then refined through one or two 3×3 convolutions and finally restored to the original image size using fourfold interpolation to generate a pixel-level segmentation map.

9. The method for extracting landslide information from SAR images based on a deep learning algorithm according to claim 6, wherein: When UNet++ is determined as the semantic segmentation model: For the input image, the encoder extracts feature maps of five scales in sequence through continuous convolution and downsampling operations. The spatial size of the feature maps is reduced step by step by 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 to construct a multi-scale feature pyramid. The decoder uses normalization at each level. The decoder channel configuration corresponds to the encoder depth, which is (256, 128, 64, 32, 16). The feature map output by each level is restored to size (×2) through upsampling until it is consistent with the input resolution. Each decoding layer receives the features of all previous encoding layers through skip connections. At the end of the decoder, 1×1 convolution is used to map the last level features to a single channel output, and Sigmoid activation is used to generate a landslide binary segmentation mask.

Images