Landslide range extraction method and device for background ground feature noise suppression, and model training method and device

Abstract

The invention provides a landslide range extraction method for background ground feature noise suppression, and the method comprises the steps: firstly, obtaining a dual-time-phase remote sensing image, and inputting the dual-time-phase remote sensing image into a generator; and then, extracting space and spectral features of the dual-time-phase remote sensing image under different scales through an encoder of the generator, and aggregating the two features to suppress background ground feature noise and capture landslide change features under different scales. And extracting time sequence difference characteristics of the landslide change characteristics under different scales, and capturing a semantic relationship of the time sequence difference characteristics between the landslide areas which are not communicated in space by introducing a non-local context enhancement strategy. And finally, decoding the time sequence difference feature map through a decoder of the generator, and dynamically adjusting the weight of a convolution kernel according to the form of the landslide in the decoded time sequence difference feature map so as to refine the boundary of the landslide and complete the extraction of the landslide. The method can effectively reduce the problems of leak detection and boundary fuzziness of the landslide.

Description

Technical Field

[0001] This invention relates to the fields of geological disaster detection and artificial intelligence technology, and in particular to a method and apparatus for extracting the landslide range by suppressing background ground noise, and a model training method and apparatus. Background Technology

[0002] Landslides occur when rock and soil masses on a slope slide downwards along a sloping surface under the influence of gravity, influenced by geological, climatic, and human factors. As a common and highly destructive geological hazard, landslides severely threaten human life and property, infrastructure stability, and the balance of the ecological environment. Landslide extraction technology can accurately identify and monitor landslide areas, providing a scientific basis for disaster early warning, risk assessment, and post-disaster recovery.

[0003] The method of landslide extraction through change detection aims to capture the difference information of images of the same area at different time points, identify the characteristics and extent of surface changes, and thus determine whether geological disasters such as landslides have occurred. In early studies, simple geometric operations (such as image differencing, image ratio analysis, regression analysis, etc.) were usually used to obtain difference maps. Some scholars used the difference method to detect changes in the large landslide in Wulong, Chongqing, and compared the detection effects of three different threshold selection methods: automatic thresholding, cyclic segmentation, and categorical variance method. The results showed that the automatic thresholding method can better separate the landslide from the background and does not require iteration. Although these methods are simple, they are greatly affected by noise and it is difficult to extract complete changes. Later, Principal Component Analysis (PCA), Kauth-Thomas (KT) transformation, Gram-Schmidt (GS) transformation, and Chi-square transformation were proposed to improve these shortcomings. These methods usually require thresholds to determine the changed and unchanged areas, and the accuracy of change detection is greatly affected by the threshold. In recent years, deep learning has developed rapidly, achieving breakthroughs in landslide information extraction thanks to its superior feature extraction and discrimination capabilities. Deep learning can extract more complex and deeper features from images and achieve efficient network optimization through backpropagation. Therefore, deep learning-based change detection landslide extraction methods have emerged, significantly improving the accuracy of landslide extraction and making the process more intelligent and automated. However, existing deep learning methods also face many challenges, especially in addressing issues such as missed detection of small-area landslides, inaccurate edge recognition, and extraction accuracy in complex and changing scenarios. Summary of the Invention

[0004] This invention describes a method and apparatus for extracting the landslide range for background ground noise suppression, and a model training method and apparatus, which can solve the above-mentioned technical problems.

[0005] According to the first aspect, a method for landslide extent extraction with background noise suppression is provided. In this method, firstly, dual-temporal remote sensing images of the target area are acquired, where the target area refers to the region where landslides may occur. Secondly, the dual-temporal remote sensing images are input into a generator. The encoder of the generator extracts the spatial and spectral features of the dual-temporal remote sensing images at different scales, and aggregates these features to suppress background noise and capture landslide change features at different scales. Temporal difference features of landslide change features at different scales are extracted. A non-local context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide regions, resulting in a first difference feature map. Finally, the decoder of the generator decodes the first difference feature map to obtain a second difference feature map. A landslide adaptive convolution strategy is introduced to dynamically adjust the convolution kernel weights based on the landslide morphology in the second difference feature map, refining the landslide boundaries in the second difference feature map to obtain a predicted landslide mask. The landslide mask is used to mark the pixel locations where landslides occur in the second difference feature map.

[0006] In some embodiments, the bi-temporal remote sensing images of the target area are preprocessed before being input into the generator. In this preprocessing method, firstly, areas unrelated to landslides are removed to reduce background interference. Secondly, the bi-temporal remote sensing images are subjected to differential processing of corresponding bands to obtain a spectral change feature map. Next, a threshold segmentation method is used to extract pixels with significant changes in the spectral change feature map, and potential landslide areas are screened based on a preset threshold. Finally, the 4-neighborhood connected component method is used to analyze and optimize the screening results, removing noise areas with areas smaller than a preset minimum threshold and non-landslide features with areas larger than a preset maximum threshold.

[0007] In some embodiments, the decoder employs a skip connection strategy to gradually restore the spatial resolution of the first difference feature map by fusing multi-scale features of the corresponding layers of the encoder.

[0008] In some embodiments, the encoder employs a dual-branch structure, consisting of a max-pooling layer, a multi-scale separable spatial-spectral convolutional module containing a spatial spectral aggregation module, and a non-local context enhancement module. The max-pooling layer and the multi-scale separable spatial-spectral convolutional module containing the spatial spectral aggregation module constitute a combined block, and the encoder includes multiple combined blocks, which are cascaded with the non-local context enhancement module. The multi-scale separable spatial-spectral convolutional module containing the spatial spectral aggregation module is used to extract spatial and spectral features of the dual-temporal remote sensing images at different scales and aggregate these features to capture landslide change characteristics at different scales. The non-local context enhancement module is used to capture the semantic relationships between spatially disconnected landslide areas with temporal differences, obtaining a first difference feature map.

[0009] In some embodiments, the decoder comprises a cascaded deconvolutional upsampling layer, a multi-scale separable spatial-spectral convolutional module including a spatial spectral aggregation module, and a landslide adaptive convolutional module. The deconvolutional upsampling layer and the multi-scale separable spatial-spectral convolutional module are used to decode the first difference feature map to obtain a second difference feature map. The landslide adaptive convolutional module dynamically adjusts the convolutional kernel weights according to the landslide morphology in the second difference feature map to refine the landslide boundary and obtain a predicted landslide mask.

[0010] In some embodiments, the method of dynamically adjusting the convolution kernel weights according to the landslide morphology in the second difference feature map to refine the landslide boundary in the second difference feature map and obtain a predicted landslide mask includes: performing a convolution operation on the second difference feature map to obtain a coarse landslide mask; performing feature enhancement on the second difference feature map based on the coarse landslide mask; performing spatial average pooling on the feature-enhanced second difference feature map; generating an adaptive convolution kernel from the spatially average pooled second difference feature map through the convolution kernel; and performing depth convolution on the second difference feature map through the adaptive convolution kernel to generate the predicted landslide mask, wherein the adaptive convolution kernel is generated according to the landslide morphology in the second difference feature map.

[0011] According to the second aspect, a model training method is provided. This model training method is applied to a landslide extraction model, which includes a generator and a discriminator. The generator includes an encoder and a decoder. In this model training method, firstly, a training set of dual-temporal remote sensing images is acquired. The dual-temporal remote sensing image training set is input into the generator, and the encoder of the generator extracts the spatial and spectral features of the dual-temporal remote sensing images at different scales. The spatial and spectral features are then aggregated, and background noise is suppressed to capture landslide change features at different scales. The temporal difference features of landslide change features at different scales are extracted, and a nonlocal context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide areas with temporal difference features, resulting in a first difference feature map. Secondly, the generator's decoder decodes the first difference feature map to obtain the second difference feature map. A landslide adaptive convolution strategy is then introduced to dynamically adjust the convolution kernel weights based on the landslide morphology in the second difference feature map, refining the landslide boundaries and obtaining a predicted landslide mask. This landslide mask is used to mark the pixel locations where landslides occur in the second difference feature map. Finally, based on the temporal difference image, the predicted landslide mask, and the ground truth landslide annotations, the discriminator compares the differences between the predicted landslide mask and the ground truth landslide annotations to optimize the generator, resulting in a trained generator. The temporal difference image is obtained by taking the absolute value of the difference sum of the two temporal remote sensing images, and the ground truth landslide annotations are obtained based on the ground truth annotations of the two temporal remote sensing images.

[0012] According to a third aspect, a landslide extent extraction device for background noise suppression is provided. This landslide extent extraction device includes a first acquisition module, a first feature map extraction module, and a first landslide extraction module. The first acquisition module acquires dual-temporal remote sensing images of a target area, where the target area refers to an area where landslides may occur. The first feature map extraction module inputs the dual-temporal remote sensing images into a generator, extracts spatial and spectral features of the dual-temporal remote sensing images at different scales through the generator's encoder, and aggregates the spatial and spectral features to suppress background noise and capture landslide change features at different scales. Temporal difference features of landslide change features at different scales are extracted, and a non-local context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide areas, resulting in a first difference feature map. The first landslide extraction module decodes the first difference feature map using the generator's decoder to obtain a second difference feature map, and introduces a landslide adaptive convolution strategy to dynamically adjust the convolution kernel weights based on the landslide morphology in the second difference feature map to refine the landslide boundaries in the second difference feature map, thus obtaining a predicted landslide mask. The landslide mask is used to mark the pixel locations where landslides occur in the second difference feature map.

[0013] According to the fourth aspect, a model training device is provided. This device includes a second acquisition module, a second feature map extraction module, and a generator optimization module. The second acquisition module acquires a training set of dual-temporal remote sensing images. The second feature map extraction module inputs the dual-temporal remote sensing image training set into the generator, extracts spatial and spectral features of the dual-temporal remote sensing images at different scales through the generator's encoder, aggregates the spatial and spectral features, and suppresses background noise to capture landslide change features at different scales. Temporal difference features of landslide change features at different scales are extracted, and a non-local context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide regions, resulting in a first difference feature map. The second landslide extraction module decodes the first difference feature map using the generator's decoder to obtain a second difference feature map. An adaptive convolution strategy for landslides is introduced to dynamically adjust the convolution kernel weights based on the landslide morphology in the second difference feature map, refining the landslide boundaries to obtain a predicted landslide mask. The landslide mask is used to mark the pixel locations where landslides occur in the second difference feature map. The generator optimization module optimizes the generator by comparing the predicted landslide mask with the actual landslide annotations using a discriminator, based on temporal difference images, predicted landslide masks, and actual landslide annotations. The temporal difference images are obtained by subtracting and taking the absolute value of the difference between the two temporal remote sensing images, and the actual landslide annotations are obtained from the actual annotations in the two temporal remote sensing images.

[0014] This application uses a generator to predict binarized landslide masks from bi-temporal remote sensing imagery, while a discriminator distinguishes the predicted landslide masks from ground truth annotations based on shape, extent, and boundary consistency, thereby guiding the generator to produce more accurate and structurally consistent results. Adversarial training between the generator and discriminator effectively reduces missed detections and boundary ambiguity. By fully utilizing bi-temporal remote sensing imagery and combining non-local context enhancement and landslide adaptive convolution strategies, the model can adapt to landslide extraction tasks in different regions, time periods, and complex environmental conditions, further enhancing its generalization ability. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This diagram illustrates a two-phase landslide extraction model structure provided in an embodiment of this specification.

[0017] Figure 2This diagram illustrates a multi-scale spatial spectrum convolution structure including a spatial spectrum aggregation module, as provided in an embodiment of this specification.

[0018] Figure 3 This diagram illustrates a spatial spectral aggregation module structure provided in an embodiment of this specification.

[0019] Figure 4 This diagram illustrates a nonlocal context enhancement module structure provided in an embodiment of this specification.

[0020] Figure 5 This diagram illustrates a landslide adaptive convolution module structure provided in an embodiment of this specification.

[0021] Figure 6 This diagram illustrates a method for extracting the landslide range for background noise suppression, as provided in an embodiment of this specification.

[0022] Figure 7 This diagram illustrates a dual-temporal remote sensing image preprocessing workflow provided in an embodiment of this specification.

[0023] Figure 8 This document shows a schematic diagram of a model training method provided in an embodiment of the present specification.

[0024] Figure 9 This diagram illustrates a landslide range extraction device for background noise suppression provided in an embodiment of this specification.

[0025] Figure 10 This diagram illustrates a model training device structure provided in an embodiment of this specification. Detailed Implementation

[0026] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be described below with reference to the accompanying drawings.

[0028] In the description of the embodiments of this application, the words "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the words "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.

[0029] In the description of the embodiments of this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, B existing alone, and A and B existing simultaneously. Furthermore, unless otherwise stated, the term "multiple" means two or more.

[0030] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and their variations all mean "including but not limited to," unless otherwise specifically emphasized.

[0031] To facilitate understanding of the solutions provided in the embodiments of this application, some of the terms involved in this solution will be briefly introduced first.

[0032] Dual-temporal remote sensing images: Dual-temporal remote sensing images refer to two or more remote sensing images of the same area acquired at different times.

[0033] GAN (Generative Adversarial Network) model: The core idea of ​​GAN is to allow two neural networks (i.e., generator and discriminator) to compete and evolve together in a game-like process.

[0034] Background features: refers to the sum of all other features, environment, and information in remote sensing images or research scenes, excluding the target features that need to be focused on.

[0035] Non-landslide features: These refer to other surface features or artificial features that are easily misidentified as landslides in remote sensing images or field surveys, but are not actually landslides.

[0036] Existing technical solutions offer a dual-branch matrix segmentation network (Matrix SegNet) for potential landslide detection in images before and after an event to remove background objects. This model is trained on image patches cropped from the original image based on connected contours. This dual-branch matrix segmentation network has a dual-tower structure, uses ResNet-50 for encoding, an SE module for feature enhancement, a matrix convolution module for multi-scale feature learning, and a focus loss function to handle the problem of imbalanced sample distribution. However, directly segmenting large-scale remote sensing images is not the most ideal method because remote sensing images are large in scale and complex in information.

[0037] Existing technical solutions also offer another dual-branch network (GDSNet), where the CPSConv branch preserves the specific semantic features of landslides while extracting key local landslide features, ensuring the accuracy of landslide extraction. The other branch introduces gated convolution to weight the features and suppress non-landslide information, resulting in excellent extraction capabilities in complex environments. Existing landslide extraction models mainly distinguish landslides from the background by analyzing image features through semantic segmentation, but they still face problems such as inaccurate landslide edge extraction, missed extractions, and large model parameters.

[0038] In addition, existing technical solutions also include several methods, such as feature engineering-based methods, deep learning methods, and generative adversarial networks (GANs).

[0039] Feature-engineering-based methods rely on experienced experts to define feature types and parameters. These features are often derived from cognitive theories and empirical models, making them difficult to adapt to different environments and data variations. They also tend to have lower accuracy and are susceptible to noise interference. In change detection, especially when background features are complex or landslide features are not obvious, feature-engineering-based methods often fail to provide sufficient accuracy.

[0040] Deep learning methods automatically extract effective features through convolutional layers, but convolutional neural networks (CNNs) suffer from a limited receptive field, making it difficult to fully capture the correlations between distant features and global semantic information. This results in insufficient understanding of global information, affecting the accuracy and reliability of landslide extraction. To overcome the limitations of CNNs in terms of receptive field size, the Transformer implements global feature modeling through a self-attention mechanism, but its computational complexity is high, especially when processing large-scale remote sensing imagery, potentially leading to longer training times and increased demands on computational resources.

[0041] Generative Adversarial Networks (GANs) offer new possibilities for improving the generalization ability of models under conditions of scarce or heterogeneous data. In the field of landslide research, GAN applications have mainly focused on sample generation and mitigation of domain bias. However, current GAN-based applications are still limited to auxiliary functions, such as style transfer and sample augmentation, and have not become a core component of the landslide extraction workflow. The structural advantages of GANs in improving boundary accuracy and preserving spatial morphology have not been fully utilized in landslide extraction. Furthermore, directly applying standard GAN architectures often leads to unstable training, poor convergence, and ambiguous outputs, mainly due to factors such as high class imbalance and complex backgrounds in landslide data. Because some background targets (such as rivers, bare land, and buildings) are similar to landslides in shape, color, and texture, they can easily interfere with automated extraction, leading to mis-extraction or missed extraction.

[0042] This invention proposes a landslide extent extraction method with background noise suppression. This method can be trained using the dual-temporal landslide extraction model (BTLE-cGAN) constructed in this invention to obtain an optimal generator. This optimal generator can then be used to directly extract landslides from new dual-temporal remote sensing images. During the adversarial network training process, firstly, the generator's encoder extracts spatial and spectral features from the dual-temporal remote sensing images at different levels and aggregates these features to suppress background noise and capture landslide variation characteristics at different scales. Secondly, it extracts temporal difference features at different scales of the dual-temporal remote sensing images. Next, a nonlocal context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide regions based on the temporal difference features. Then, the generator's decoder decodes the temporal difference feature map and introduces a landslide adaptive convolution strategy to dynamically adjust the convolution kernel weights according to the landslide morphology in the decoded temporal difference feature map, refining the landslide boundaries of the second difference feature map to obtain a predicted landslide mask. Then, the discriminator can optimize the generator by comparing the predicted landslide mask with the actual landslide annotations based on temporal difference images, landslide masks, and actual landslide annotations. Finally, landslide extraction can be performed on newly acquired dual-temporal remote sensing images using the optimized generator. This method introduces a non-local context enhancement strategy into the adversarial network to capture the semantic relationships between spatially disconnected landslide regions. Simultaneously, it also introduces a landslide adaptive convolution strategy, which dynamically adjusts the convolution kernel to better handle the boundaries of irregular landslide features.

[0043] Next, the method for extracting the landslide range for background feature noise suppression provided in the embodiments of this application will be described.

[0044] This invention proposes a dual-temporal landslide extraction model (BTLE-cGAN). Figure 1 A schematic diagram of a dual-temporal landslide extraction model (BTLE-cGAN) architecture is shown. Figure 1 As shown, the model may include a generator and a discriminator. Furthermore, the generator may include an encoder and a decoder.

[0045] The generator can be used to receive dual-temporal remote sensing imagery. Specifically, the generator's encoder can employ a dual-branch network architecture, with one branch used to receive earlier remote sensing imagery of the target area (e.g., ...). Figure 1 Mid-phase 1), another branch is used to receive late-stage remote sensing images of the target area (e.g., Figure 1 Phase 2 in the middle.

[0046] The encoder can be used to extract spatial and spectral features from dual-temporal remote sensing images at different scales, aggregate these features, mine the correlations between them at different scales, suppress background noise, and capture landslide change characteristics at different scales. It can also be used to extract temporal difference features from dual-temporal remote sensing images at different scales. Furthermore, by introducing nonlocal context enhancement strategies, the semantic relationships between spatially disconnected landslide areas can be captured.

[0047] In one implementation, reference Figure 1 The encoder can include convolutional layers. After inputting bi-temporal remote sensing images into a bi-branch decoder network, they can first pass through this convolutional layer. This can map the number of channels (e.g., 3 or more) of the input bi-temporal remote sensing images to a new, fixed number of feature channels, serving as a unified input dimension for subsequent networks. It can also capture spatial features such as edges, textures, and color patches, providing richer input for subsequent deep networks.

[0048] For example, the convolutional layer can be designed with a 3×3 kernel, a stride of 1, padding of 1, and an activation function of ReLU, outputting a feature map of size 256×256×32.

[0049] In one implementation, continue to refer to Figure 1 The encoder may further include a max-pooling layer, a multi-scale separable spatial-spectral convolutional module containing a spatial spectral aggregation module, and a non-local context enhancement module. Specifically, the max-pooling layer and the multi-scale separable spatial-spectral convolutional module containing the spatial spectral aggregation module can constitute a combined block, and the encoder may include multiple combined blocks. Preferably, the encoder may include four combined blocks, which are connected in series, and the last combined block may also be cascaded with the non-local context enhancement module.

[0050] Specifically, max pooling layers can be used to max pool bi-temporal remote sensing images or bi-temporal remote sensing images processed by convolutional layers, increasing the receptive field of the temporal remote sensing images to capture higher-level semantic information. The receptive field refers to the size of the region in the original image corresponding to a point on the feature map. After downsampling through max pooling (e.g., from 256x256 to 128x128), a smaller convolutional kernel (e.g., 3x3) in subsequent convolutional layers can focus on a larger area in the original image.

[0051] The multi-scale separable spatial-spectral convolution module, which includes a spatial-spectral aggregation module, can be used to extract spatial and spectral features from bi-temporal remote sensing images at different scales. It can also aggregate spatial and spectral features and suppress background noise to uncover hidden information in bi-temporal remote sensing images at different scales.

[0052] Figure 2 This diagram illustrates the structure of a multi-scale separable spatial spectrum convolution module that includes a spatial spectrum aggregation module, as shown below. Figure 2 As shown, firstly, the multi-scale separable spatial-spectral convolution module can use point convolution to generate several (e.g., 16) local feature maps from the input dual-temporal remote sensing image feature maps. Then, dilation convolution can be used with different dilation rates (e.g., 1, 3, 6) to expand the spatial correlation of these local feature maps, thereby obtaining several (e.g., 16) auxiliary feature maps. Subsequently, a spatial-spectral aggregation module can be introduced to aggregate spatial and spectral features to model the dependency between space and spectrum, suppress background noise, and mine the implicit information in the dual-temporal remote sensing images at different scales.

[0053] Figure 3 A structural diagram of a spatial spectral aggregation module is shown. Figure 3 As shown, firstly, global average pooling can be performed on the auxiliary feature map after dilated convolution. Secondly, the similarity between the global mean (Q) of the channel dimension and the spatial location feature (K) can be modeled using the Gaussian kernel function. The similarity calculation formula (1) after modeling is as follows:

[0054]

[0055] Where Q represents the global mean of the channel dimension. K represents the input spatial location feature map F. s ∈R C×H×W ;σ 2 The variance σ represents the channel dimension. 2 ∈R C×1×1 β represents the empirical bias (the value can be 0.5); This represents the similarity between the global mean (Q) of the channel dimension and the spatial location feature (K).

[0056] Next, after modeling the similarity between the global mean (Q) of the channel dimension and the spatial location features (K), an attention weighting mechanism can be used to enhance important spatial regions or channel features, thereby completing the aggregation of spatial and spectral features of dual-temporal remote sensing images.

[0057] Back Figure 1After aggregating spatial and spectral features of dual-temporal remote sensing images, temporal differences at different scales can be extracted. Specifically, subtraction and absolute value operations can be performed on each layer of the encoder's feature map to obtain a difference feature map. Pixel-level subtraction can be performed on each layer of the feature map corresponding to the dual-temporal remote sensing images (temporal phase 1 and temporal phase 2) in the dual-branch encoder. The result may produce positive or negative values; a positive value indicates that the pixel value in temporal phase 2 is higher than that in temporal phase 1 (e.g., an increase in the vegetation index), and a negative value indicates that the pixel value in temporal phase 2 is lower than that in temporal phase 1 (e.g., a decrease in the vegetation index). After pixel-level subtraction of the two images, the absolute value can be taken. The purpose of taking the absolute value is to ignore the direction of change and only focus on the magnitude of the change.

[0058] Continue to refer to Figure 1 After extracting temporal difference features from different levels of dual-temporal remote sensing images, a nonlocal context enhancement module can be used to capture the semantic relationships between spatially disconnected landslide areas based on these temporal difference features. Specifically, refer to... Figure 4 The Nonlocal Context Enhancement (NCEM) module uses a 1x1 convolutional kernel to process temporal difference features, resulting in three feature maps, denoted as v, g, and u. Feature maps v and u are then multiplied element-wise to obtain autocorrelation information. Next, normalization is performed using the softmax function. Then, it is multiplied element-wise with feature map g. The multiplied feature maps are then up-dimensioned and the residuals are added to the original temporal difference feature maps to obtain the final result.

[0059] It is worth noting that the original design of traditional nonlocal modules uses a 1x1x1 convolutional kernel, processing both the spatial (H×W) and channel (C) dimensions. Unlike traditional nonlocal modules, the Nonlocal Context Enhancement Module (NCEM) of this application uses a 1x1 convolutional kernel, focusing only on the spatial dimension (H×W), simplifying computation and reducing complexity, while still effectively capturing the semantic relationships between spatially disconnected landslide regions that exhibit temporal differences.

[0060] The previous section mainly explained the encoder in the generator; the following section will discuss the decoder in the generator.

[0061] Continue to refer to Figure 1 After capturing the semantic relationships between spatially disconnected landslide areas by using a nonlocal context enhancement module, the temporal difference features can be decoded by a decoder. A landslide adaptive convolution strategy is introduced to refine the landslide boundaries of the decoded temporal difference features, thereby predicting the landslide mask.

[0062] In one implementation, the decoder can employ a skip connection strategy to progressively restore the spatial resolution of the temporal difference feature map by fusing multi-scale features from corresponding layers of the encoder. The encoder progressively downsamples to extract multi-scale features, but this results in a loss of spatial detail. The decoder then progressively upsamples to restore spatial resolution. The skip connection strategy fuses features from corresponding layers of the encoder with features from the decoder at the same scale, thereby helping the decoder recover more spatial detail.

[0063] In one implementation, return to Figure 1 The decoder may include an upsampling layer, a multi-scale separable spatial-spectral convolutional module containing a spatial spectral aggregation module, and a slippery adaptive convolutional module. Its upsampling layer and the multi-scale separable spatial-spectral convolutional module containing the spatial spectral aggregation module can form a combinatorial block. The decoder may include multiple combinatorial blocks, the number of which corresponds to the number of combinatorial blocks in the encoder. This allows for the implementation of a skip connection strategy, enabling the fusion of features from the corresponding layer of the encoder with features of the same scale in the decoder, thereby helping the decoder recover more spatial details.

[0064] Continue to refer to Figure 1 The decoder can decode temporal difference features to gradually restore the spatial resolution of the image. In other words, it can progressively upsample low-resolution, high-semantic temporal difference features and restore spatial details, ultimately outputting a feature map with the same size as the dual-temporal remote sensing image. After decoding the temporal difference features, a landslide adaptive convolution module can be used to dynamically adjust the convolution kernel weights based on the landslide morphology in the decoded temporal difference feature map, refining the landslide boundaries of the decoded temporal difference features and thus predicting the landslide mask.

[0065] In one implementation, the landslide adaptive convolution module first performs a convolution operation on the decoded temporal difference feature map to obtain a coarse landslide mask. Then, feature enhancement is performed on the decoded temporal difference feature map based on the coarse landslide mask (e.g., multiplying the coarse landslide mask with the decoded temporal difference feature map). The feature-enhanced temporal difference feature map can then undergo spatial average pooling. Finally, the spatially average pooled temporal difference feature map is used to generate an adaptive convolution kernel, which performs a depthwise convolution on the spatially average pooled temporal difference feature map to generate the predicted landslide mask. It is easy to understand that the adaptive convolution kernel is generated based on the landslide morphology in the decoded temporal difference feature map, because the coarse landslide mask can reflect the landslide morphology.

[0066] For example, refer to Figure 5 First, the decoded temporal difference feature map can be analyzed. Perform a convolution operation to generate a coarse mask. For example, feature maps are obtained after processing through a 3x3 convolutional layer. Feature map After further processing with a 1x1 convolutional layer and a sigmoid layer, a coarse mask is generated. Then apply the coarse mask Timing difference feature map after decoding Multiply. Then, spatial average pooling can be performed on the product, progressively downsampling it into a feature map of size 3×3×C (C... dk ×3×3), the downsampled feature map (C dk A 3x3 convolutional layer can generate an adaptive convolutional kernel. To improve efficiency and save computational resources, each convolutional filter in the adaptive convolutional kernel can be applied to one channel of the feature map. Therefore, the adaptive convolutional kernel can be applied to the decoded temporal difference feature map. The residual feature map is obtained by channel-wise convolution. and combine it with feature maps Adding them together yields a refined feature map. Finally, a refined landslide mask M can be generated using 1×1 convolutional layers and sigmoid layers.

[0067] In this implementation, the landslide adaptive convolution combines pooling and downsampling operations, adjusting the convolution kernel weights through an adaptive learning mechanism. This approach allows the model to dynamically adjust the convolution kernel response based on the landslide morphology in the input feature map, thereby improving the model's adaptability to different feature map inputs and avoiding the drawbacks of globally fixed weights. Incorporating landslide morphology features into the convolution kernel design enhances the detection and recognition capabilities of landslide phenomena.

[0068] The above content describes the structure and function of the generator in the dual-temporal landslide extraction model (BTLE-cGAN). The following section will describe the structure and function of the discriminator.

[0069] Back Figure 1 Its discriminator can be based on temporal difference images and predicted landslide masks (i.e., Figure 1 The generator is optimized by comparing the predicted landslide mask with the actual landslide label, which is generated by the generator.

[0070] Temporal difference images can be obtained from dual-temporal remote sensing images. Specifically, two temporal phases (e.g., Figure 1Subtraction and absolute value operations are performed on the remote sensing images of time phase 1 and time phase 2 to obtain a time-series difference image. Pixel-level subtraction can be performed between the remote sensing images of time phase 1 and time phase 2. The result may be positive or negative; a positive value indicates that the pixel value increased in time phase 2 compared to time phase 1 (e.g., an increase in the vegetation index), and a negative value indicates that the pixel value decreased in time phase 2 compared to time phase 1 (e.g., a decrease in the vegetation index). After pixel-level subtraction of the two images, the absolute value can be taken. The purpose of taking the absolute value is to ignore the direction of change and only focus on the magnitude of the change.

[0071] Accurate annotations of landslides can be obtained through manual or semi-manual interpretation of dual-temporal remote sensing images. The annotation method is a common technique in this field and will not be elaborated upon here.

[0072] Specifically, please refer to Figure 1 The discriminator network can be composed of three stacked convolutional-normalization-activation layers. Each layer in the discriminator network can have a 3×3 convolutional kernel. For example, the stride can be set to 2, the padding to 1, and the activation function to LeakyReLU. Finally, a 1×1 convolutional layer can be used for dimensionality reduction, and the probability value is output using the sigmoid function.

[0073] The above content describes the structure and function of the dual-temporal landslide extraction model (BTLE-cGAN). The following section will describe the training process of the landslide change detection model.

[0074] The dual-temporal landslide extraction model (BTLE-cGAN) can be trained using a training set (dual-temporal remote sensing images). The total loss function during training can include adversarial loss L. adv Cross-entropy loss L bce and perceived loss L per Its total loss function is defined as shown in formula (2).

[0075] L(G,D)=L adv (G, D) + αL bce (G)+βL per (G) (2)

[0076] In the formula, α is a weight hyperparameter used to balance the cross-entropy loss, and β is a weight hyperparameter used to balance the perceptual loss. For example, they can be set to α = 1 and β = 1e, respectively. -3 G stands for generator, and D stands for discriminator.

[0077] The main function of adversarial loss is to train the generator and discriminator, enabling the generator to not only successfully deceive the discriminator but also generate images that are visually similar to the real landslide annotations. Temporal difference images |x1-x2| can be introduced as conditions for the generator and discriminator; the representation of adversarial loss is shown in formula (3). Perceptual loss measures the perceptual difference between the generated landslide image and the real landslide image, rather than the pixel-level difference. This loss can be calculated using the feature layers of a VGG pre-trained network to extract the perceptual features of the image, as shown in formula (4). The calculation of cross-entropy loss is shown in formula (5).

[0078]

[0079] In the formula, y represents the actual landslide label, x1 represents the remote sensing image of time phase 1, x2 represents the remote sensing image of time phase 2, |x1-x2| represents the time-series difference image, and E represents the expectation.

[0080] L per =‖φ(G(x))-φ(y)‖2 (4)

[0081] In the formula, φ represents the feature extraction of the ReLU3_3 layer of the pre-trained VGG network, and G(x) represents the predicted landslide mask output by the generator.

[0082] L bce =-ylog2(G(x))-(1-y)log2(1-G(x)) (5)

[0083] For example, during the training of the dual-temporal landslide extraction model (BTLE-cGAN), landslides can be labeled as 1, and background features can be labeled as 0. Training uses the Adam optimizer with weight decay set to 5e. -4 The momentum parameters β1 and β2 were set to 0.5 and 0.999, respectively, and the initial learning rate was 1e. -4 Meanwhile, a cosine annealing learning rate scheduler is used to gradually decrease the learning rate during training.

[0084] The optimal generator network is trained, and the test set is input into the trained optimal generator network to obtain the landslide extraction results of the test set. In addition, this application also calculates the Intersection over Union (IOU), F1 score, recall, and precision for the entire test sample set to evaluate its accuracy, and obtains the statistical results shown in Table 1. The calculation method of IOU is shown in Equation (6), where TP represents the number of real pixels extracted as landslides; TN represents the number of real pixels extracted as background; FP represents the number of real background features misclassified as landslides; and FN represents the number of real landslides misclassified as background features. The specific calculation methods of Precision, Recall, and F1 score are shown in Equations (7)-(9).

[0085]

[0086] IOU Recall Precision F1 80.22％ 88.69％ 89.36％ 89.02％

[0087] Table 1. Statistical Table for Verification of Landslide Extraction Accuracy

[0088] Based on the statistical results of the above experiments, the accuracy index IOU is greater than 80%, and the other three accuracy indices are all greater than 88%. The dual-temporal landslide extraction model (BTLE-cGAN) proposed in this invention effectively improves the accuracy of landslide extraction, effectively solves the common problems of missed detection and false detection in landslides, improves the accuracy of landslide boundary extraction, enhances the adaptability of the model in complex geological environments, and improves the generalization ability of the model.

[0089] In this embodiment, a bi-temporal landslide extraction model framework based on a generator-discriminator mechanism (BTLE-cGAN) is introduced. The generator predicts binarized landslide masks from bi-temporal remote sensing images, while the discriminator distinguishes the predicted landslide masks from the ground truth labels based on shape, extent, and boundary consistency, thereby guiding the generator to produce more accurate and structurally consistent results. Adversarial training between the generator and discriminator effectively reduces missed detections and boundary ambiguity. Simultaneously, a non-local context enhancement module is introduced, which can deeply explore the correlations between distant landslides in the images, effectively solving the common problem of missed landslide extraction. Furthermore, an adaptive landslide convolution module is employed, dynamically adjusting the convolution kernel according to the landslide morphology. This not only improves the model's adaptability to different bi-temporal remote sensing images but also accurately captures landslide edge features, effectively solving the problem of inaccurate landslide area boundaries. By making full use of dual-temporal remote sensing imagery and combining the Non-Local Context Enhancement (NCEM) module and the Landslide Adaptive Convolution (LAC) module, the model can adapt to landslide extraction tasks in different regions, time periods, and complex environmental conditions, further enhancing the model's generalization ability.

[0090] This invention also provides a method for landslide extent extraction with background feature noise suppression. This method can be applied to the dual-temporal landslide extraction model (BTLE-cGAN) proposed in the above embodiments. Figure 6 As shown, the method may include the following steps.

[0091] Step S601: Acquire dual-temporal remote sensing images of the target area. The target area refers to the region where landslides may occur.

[0092] For example, Mentougou District in Beijing can be taken as the target area of ​​the study. Multispectral data before and after the landslide can be obtained. The dual-temporal remote sensing image can be received by the generator for processing.

[0093] In one implementation, the bi-temporal remote sensing images can be preprocessed before the generator receives them of the target area. The preprocessing procedure can be found in [reference needed]. Figure 7 ,like Figure 7 As shown, the preprocessing workflow for dual-temporal remote sensing images may include the following steps.

[0094] First, areas unrelated to landslides can be removed from dual-temporal remote sensing images to reduce interference from background features.

[0095] Specifically, spectral thresholding can be used to remove landslide-independent areas from dual-temporal remote sensing images. These areas can include built-up areas, roads, and other landslide-independent locations to reduce background interference.

[0096] Secondly, differential processing of corresponding bands can be performed on the dual-temporal remote sensing images to obtain the spectral variation characteristics of the dual-temporal remote sensing images.

[0097] Specifically, differential processing can be performed on corresponding bands of two images to highlight the spectral variation characteristics between the two images and preliminarily identify potential landslide areas. Differential processing refers to subtracting the two images pixel by pixel and taking the absolute value.

[0098] Then, a threshold segmentation method can be used to extract pixels with significant changes in the spectral change feature map, and potential landslide areas can be screened out based on a preset threshold.

[0099] Specifically, thresholding can be used to extract pixels showing significant changes. By setting appropriate thresholds, potential landslide areas can be identified. Thresholding involves classifying pixels in an image into different categories by setting one or more grayscale thresholds. This method can be used to separate the changed areas of a landslide from those with no related changes.

[0100] Next, the 4-neighborhood connected component method can be used to analyze and optimize the screening results, removing noise regions with an area smaller than a preset minimum threshold and non-slide features with an area larger than a preset maximum threshold. 4-neighborhood connected component analysis refers to a technique used to identify and label independent connected regions in binary images.

[0101] For example, 4-neighborhood connectivity analysis can be used to further optimize the extraction results, removing noise areas with an area of ​​less than 5 pixels and non-landslide features with an area of ​​more than 500 pixels, to ensure that the extracted landslide areas conform to optical response and spatial distribution characteristics.

[0102] After preprocessing, the dual-temporal remote sensing images can be cropped into standardized sample blocks (e.g., 32×32 pixels) based on the geometric center of each potential landslide area, and integrated into a high-quality landslide change detection dataset. The dataset is divided into training, validation, and test sets according to a certain ratio (e.g., 7:1:2) to ensure the reliability of the model evaluation. Image enhancement techniques such as translation, rotation, scaling, random cropping, and contrast adjustment can also be used to expand the dataset. Finally, the image data is resampled using quadratic linear interpolation and resized (e.g., to 256×256) to meet the generator's input requirements.

[0103] Most existing landslide datasets are generated by sequentially cropping images. However, this method fails to adequately consider the salience of changes in different regions within the data, resulting in the incorrect inclusion of many irrelevant or noisy areas in the dataset. This not only degrades the quality of the training data but may also affect the model's learning performance, causing it to be unable to accurately identify landslide areas. In this implementation, a sample balancing strategy based on connected component analysis was developed, thereby constructing a high-quality landslide change detection dataset. By analyzing the spectral differences in multi-temporal cloudless remote sensing image data, significantly changing landslide areas can be successfully identified and screened. The construction of this dataset provides the model with rich and accurate training data, improving the model's accuracy and reliability in landslide change detection in complex environments.

[0104] In step S602, dual-temporal remote sensing images can be input into the generator. The encoder of the generator extracts the spatial and spectral features of the dual-temporal remote sensing images at different scales, and aggregates the spatial and spectral features to suppress background noise, thereby capturing landslide change features at different scales. Temporal difference features of landslide change at different scales can be extracted, and a non-local context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide areas, resulting in a first difference feature map.

[0105] In one embodiment, the encoder may employ a dual-branch structure, comprising a max-pooling layer, a multi-scale separable spatial-spectral convolutional module including a spatial spectral aggregation module, and a non-local context enhancement module. The max-pooling layer and the multi-scale separable spatial-spectral convolutional module can form a combined block, and the encoder may include multiple combined blocks, which can be cascaded with the non-local context enhancement module.

[0106] Specifically, the multi-scale separable spatial-spectral convolution module can be used to extract spatial and spectral features of dual-temporal remote sensing images at different levels. The spatial-spectral aggregation module can be used to aggregate spatial and spectral features and suppress background noise to mine implicit information in dual-temporal remote sensing images at different levels. The non-local context enhancement module can be used to capture the semantic relationships between spatially disconnected landslide areas with temporal differences, obtaining the first difference feature map.

[0107] In step S603, the first difference feature map can be decoded by the generator's decoder to obtain the second difference feature map. A landslide adaptive convolution strategy is then introduced to dynamically adjust the convolution kernel weights based on the landslide morphology in the second difference feature map, thereby refining the landslide boundaries and obtaining a predicted landslide mask. The landslide mask is used to mark the pixel locations in the second difference feature map where landslides occur.

[0108] In one implementation, the decoder can employ a skip connection strategy to gradually recover the spatial resolution of the first difference feature map by fusing multi-scale features from the corresponding layers of the encoder.

[0109] In one implementation, the decoder may consist of cascaded deconvolutional upsampling layers, a multi-scale separable spatial spectrum convolutional module including a spatial spectrum aggregation module, and a landslide adaptive convolutional module.

[0110] Specifically, the deconvolutional upsampling layer and the multi-scale separable spatial-spectral convolutional module containing a spatial spectral aggregation module can be used to decode the second difference feature map, gradually restoring the spatial resolution of the first difference feature map, thereby generating the second difference feature map. The landslide adaptive convolutional module can be used to dynamically adjust the convolutional kernel weights according to the landslide morphology in the second difference feature map to refine the landslide boundary of the second difference feature map and obtain the predicted landslide mask.

[0111] In one implementation, a method for dynamically adjusting the convolution kernel weights based on the landslide morphology in the second difference feature map to refine the landslide boundary in the second difference feature map and obtain a predicted landslide mask may include the following steps.

[0112] First, a coarse landslide mask can be obtained by performing a convolution operation on the second difference feature map using a landslide adaptive convolution module.

[0113] Secondly, feature enhancement can be performed on the second difference feature map based on the coarse landslide mask. Specifically, the coarse landslide mask can be multiplied with the second difference feature map using a landslide adaptive convolution module to enhance the features of the second difference feature map.

[0114] Then, spatial average pooling can be performed on the second difference feature map after feature enhancement.

[0115] Finally, the second difference feature map after spatial average pooling is used to generate an adaptive convolution kernel. This adaptive convolution kernel is then used to perform depthwise convolution on the second difference feature map to generate the predicted landslide mask. The adaptive convolution kernel is generated based on the landslide morphology in the second difference feature map.

[0116] This application also provides a model training method. This model training method can be applied to the dual-temporal landslide extraction model (BTLE-cGAN) proposed in the above embodiments. The dual-temporal landslide extraction model may include a generator and a discriminator, and the generator may include an encoder and a decoder. Figure 8 As shown, the model method may include the following steps.

[0117] Step S801: Obtain the training set of dual-temporal remote sensing images.

[0118] In step S802, the training set of dual-temporal remote sensing images can be input into the generator. The encoder of the generator extracts the spatial and spectral features of the dual-temporal remote sensing images at different scales, and aggregates the spatial and spectral features to suppress background noise, so as to capture landslide change features at different scales. The temporal difference features of landslide change features at different scales are extracted, and a non-local context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide areas with temporal difference features, resulting in the first difference feature map.

[0119] In step S803, the first difference feature map can be decoded by the generator's decoder to obtain the second difference feature map. A landslide adaptive convolution strategy is then introduced to dynamically adjust the convolution kernel weights based on the landslide morphology in the second difference feature map, thereby refining the landslide boundaries and obtaining a predicted landslide mask. This landslide mask can be used to mark the pixel locations in the second difference feature map where landslides occur.

[0120] Step S804 involves using a discriminator to compare the differences between the predicted landslide mask and the actual landslide annotations based on the temporal difference image, the predicted landslide mask, and the actual landslide annotations, thereby optimizing the generator and obtaining a trained generator. The temporal difference image is obtained by subtracting and taking the absolute value of the difference between the two temporal remote sensing images, and the actual landslide annotations are obtained based on the actual annotations of the two temporal remote sensing images.

[0121] Specifically, subtraction and absolute value operations can be performed on remote sensing images from two different time periods to obtain a temporal difference image. The specific implementation methods for subtraction and absolute value operations have been described above and will not be repeated here. Accurate annotation of landslides can be obtained through manual or semi-manual interpretation of the dual-temporal remote sensing images; the annotation method is a common technique in this field and will not be elaborated upon here.

[0122] In one implementation, the bi-temporal remote sensing images in the training set can be preprocessed before acquiring the training set. The preprocessing method is the same as the bi-temporal remote sensing image preprocessing method described above, and will not be repeated here.

[0123] Based on the above description, this application also provides a landslide range extraction device for background feature noise suppression. For example... Figure 9 As shown, the landslide extent extraction device may include:

[0124] The first acquisition module 901 is used to acquire dual-temporal remote sensing images of the target area. The target area refers to the area where landslides may occur.

[0125] The first feature map extraction module 902 is used to input dual-temporal remote sensing images into the generator. The encoder of the generator extracts the spatial and spectral features of the dual-temporal remote sensing images at different scales, and aggregates the spatial and spectral features to suppress background noise, so as to capture landslide change features at different scales. The temporal difference features of landslide change features at different scales are extracted, and a non-local context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide areas with temporal difference features, resulting in the first difference feature map.

[0126] The first landslide extraction module 903 is used to decode the first difference feature map through the decoder of the generator to obtain the second difference feature map. It then introduces a landslide adaptive convolution strategy to dynamically adjust the convolution kernel weights based on the landslide morphology in the second difference feature map, thereby refining the landslide boundaries and obtaining a predicted landslide mask. The landslide mask is used to mark the pixel locations in the second difference feature map where landslides have occurred.

[0127] Based on the above description, embodiments of this application also provide a model training apparatus. For example... Figure 10As shown, the model training device may include: a second acquisition module 1001, used to acquire a training set of dual-temporal remote sensing images.

[0128] The second feature map extraction module 1002 is used to input the dual-temporal remote sensing image training set into the generator. The encoder of the generator extracts the spatial and spectral features of the dual-temporal remote sensing images at different scales, and aggregates the spatial and spectral features to suppress background noise, so as to capture landslide change features at different scales. The temporal difference features of landslide change features at different scales are extracted, and a non-local context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide areas with temporal difference features, resulting in the first difference feature map.

[0129] The second landslide extraction module 1003 is used to decode the first difference feature map through the decoder of the generator to obtain the second difference feature map. It then introduces a landslide adaptive convolution strategy to dynamically adjust the convolution kernel weights based on the landslide morphology in the second difference feature map, thereby refining the landslide boundaries and obtaining a predicted landslide mask. The landslide mask is used to mark the pixel locations in the second difference feature map where landslides have occurred.

[0130] The generator optimization module 1004 is used to optimize the generator by comparing the differences between the predicted landslide mask and the actual landslide annotations based on temporal difference images, predicted landslide masks, and actual landslide annotations, thereby obtaining a trained generator. The temporal difference images are obtained by taking the absolute value of the difference sum of the two temporal remote sensing images, and the actual landslide annotations are obtained based on the actual annotations of the two temporal remote sensing images.

[0131] Those skilled in the art will readily recognize that, based on the units and algorithm steps described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is implemented in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0132] Corresponding to the methods provided by the present invention, the present invention also provides a system. This system includes the generator and discriminator described in any of the above embodiments, and is used to execute the methods described in any of the above embodiments.

[0133] Corresponding to the methods provided by the present invention, the present invention also provides a computer-readable storage medium. This computer-readable storage medium includes a stored program, wherein the program, when executed by a processor, implements the methods described in any of the above embodiments.

[0134] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0135] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for landslide range extraction to suppress background ground noise, characterized in that, The method includes: Acquire dual-temporal remote sensing images of the target area; wherein, the target area refers to an area where landslides may occur; The dual-temporal remote sensing images are input into the generator. The encoder of the generator extracts the spatial and spectral features of the dual-temporal remote sensing images at different scales, and aggregates the spatial and spectral features to suppress background noise and capture landslide change features at different scales. The temporal difference features of the landslide change features at different scales are extracted, and a nonlocal context enhancement strategy is introduced to capture the semantic relationships between spatially disconnected landslide areas, resulting in a first difference feature map. The generator decodes the first difference feature map to obtain a second difference feature map. A landslide adaptive convolution strategy is introduced to dynamically adjust the convolution kernel weights according to the landslide shape in the second difference feature map to refine the landslide boundary of the second difference feature map and obtain a predicted landslide mask. The landslide mask is used to mark the pixel positions where landslides occur in the second difference feature map.

2. The method according to claim 1, characterized in that, Before inputting the dual-temporal remote sensing image of the target area into the generator, the dual-temporal remote sensing image of the target area is preprocessed: Areas unrelated to the landslide were removed to reduce background disturbance. The dual-temporal remote sensing images are subjected to differential processing of corresponding bands to obtain spectral variation feature maps; A threshold segmentation method is used to extract pixels with significant changes in the spectral change feature map, and potential landslide areas in the spectral change feature map are screened out based on a preset threshold. The 4-neighborhood connectivity method was used to analyze and optimize the screening results, and noise areas with an area smaller than the preset minimum threshold and non-landslide features with an area larger than the preset maximum threshold were removed.

3. The method according to claim 1, characterized in that, The decoder employs a skip connection strategy, gradually restoring the spatial resolution of the first difference feature map by fusing multi-scale features from the corresponding layers of the encoder.

4. The method according to claim 1, characterized in that, The encoder employs a dual-branch structure, comprising a max-pooling layer, a multi-scale separable spatial-spectral convolutional module including a spatial spectral aggregation module, and a non-local context enhancement module. The max-pooling layer and the multi-scale separable spatial-spectral convolutional module containing the spatial spectral aggregation module constitute a combined block. The encoder includes multiple combined blocks, and these combined blocks and the non-local context enhancement module are cascaded. The multi-scale separable spatial-spectral convolution module, which includes a spatial spectral aggregation module, is used to extract the spatial and spectral features of the dual-temporal remote sensing images at different scales, and aggregate the spatial and spectral features to capture landslide change features at different scales. The nonlocal context enhancement module is used to capture the semantic relationships between spatially disconnected landslide areas with the temporal difference features, and obtain the first difference feature map.

5. The method according to claim 1, characterized in that, The decoder consists of cascaded deconvolutional upsampling layers, a multi-scale separable spatial spectrum convolutional module containing a spatial spectrum aggregation module, and a landslide adaptive convolutional module; wherein, The deconvolutional upsampling layer and the multi-scale separable spatial spectrum convolutional module containing the spatial spectrum aggregation module are used to decode the first difference feature map to obtain the second difference feature map. The landslide adaptive convolution module is used to dynamically adjust the weights of the convolution kernel according to the shape of the landslide in the second difference feature map, so as to refine the landslide boundary of the second difference feature map and obtain the predicted landslide mask.

6. The method according to claim 5, characterized in that, The convolution kernel weights are dynamically adjusted based on the landslide morphology in the second difference feature map to refine the landslide boundary, resulting in a predicted landslide mask including: Perform a convolution operation on the second difference feature map to obtain a coarse landslide mask; Feature enhancement is performed on the second difference feature map based on the coarse landslide mask; Spatial average pooling is performed on the second difference feature map after feature enhancement; The second difference feature map after spatial average pooling generates an adaptive convolution kernel through a convolution kernel. The second difference feature map is then depthwise convolved with the adaptive convolution kernel to generate a predicted landslide mask. The adaptive convolution kernel is generated based on the landslide morphology in the second difference feature map.

7. A model training method, characterized in that, The model training method is applied to a landslide extraction model, which includes a generator and a discriminator. The generator includes an encoder and a decoder. The model training method includes: Acquire a training set of dual-temporal remote sensing images; The training set of the dual-temporal remote sensing images is input into the generator. The encoder of the generator extracts the spatial and spectral features of the dual-temporal remote sensing images at different scales, and aggregates the spatial and spectral features to suppress background noise, so as to capture landslide change features at different scales. The temporal difference features of the landslide change features at different scales are extracted, and a nonlocal context enhancement strategy is introduced to capture the semantic relationship between spatially disconnected landslide areas, so as to obtain the first difference feature map. The generator decodes the first difference feature map to obtain a second difference feature map. A landslide adaptive convolution strategy is introduced to dynamically adjust the convolution kernel weights according to the landslide shape in the second difference feature map to refine the landslide boundary of the second difference feature map and obtain a predicted landslide mask. The landslide mask is used to mark the pixel positions where landslides occur in the second difference feature map. Based on the temporal difference image, the predicted landslide mask, and the actual landslide annotations, the discriminator compares the differences between the predicted landslide mask and the actual landslide annotations to optimize the generator and obtain a trained generator. The temporal difference image is obtained by taking the absolute value of the difference sum of the two-temporal remote sensing images, and the actual landslide annotations are obtained by taking the actual annotations of the two-temporal remote sensing images.

8. The method according to claim 7, characterized in that, Before acquiring the training set of dual-temporal remote sensing images, the dual-temporal remote sensing images in the training set are preprocessed: Areas unrelated to the landslide were removed to reduce background disturbance. The dual-temporal remote sensing images are subjected to differential processing of corresponding bands to obtain spectral variation feature maps; A threshold segmentation method is used to extract pixels with significant changes in the spectral change feature map, and potential landslide areas in the spectral change feature map are screened out based on a preset threshold. The 4-neighborhood connectivity method was used to analyze and optimize the screening results, and noise areas with an area smaller than the preset minimum threshold and non-landslide features with an area larger than the preset maximum threshold were removed.

9. A landslide range extraction device for background noise suppression, characterized in that, The landslide extent extraction device includes: The first acquisition module is used to acquire dual-temporal remote sensing images of the target area; wherein, the target area refers to an area where landslides may occur. The first feature map extraction module is used to input the dual-temporal remote sensing images into the generator, extract the spatial and spectral features of the dual-temporal remote sensing images at different scales through the encoder of the generator, and aggregate the spatial and spectral features to suppress background noise and capture landslide change features at different scales; extract the temporal difference features of the landslide change features at different scales, and capture the semantic relationships between spatially disconnected landslide areas by introducing a nonlocal context enhancement strategy to obtain the first difference feature map; The first landslide extraction module is used to decode the first difference feature map through the decoder of the generator to obtain the second difference feature map, and introduces a landslide adaptive convolution strategy to dynamically adjust the weight of the convolution kernel according to the shape of the landslide in the second difference feature map, so as to refine the landslide boundary of the second difference feature map and obtain a predicted landslide mask; wherein, the landslide mask is used to mark the pixel position of the landslide in the second difference feature map.

10. A model training device, characterized in that, The model training device includes: The second acquisition module is used to acquire a training set of dual-temporal remote sensing images. The second feature map extraction module is used to input the dual-temporal remote sensing image training set into the generator, extract the spatial and spectral features of the dual-temporal remote sensing images at different scales through the encoder of the generator, and aggregate the spatial and spectral features to suppress background noise in order to capture landslide change features at different scales; extract the temporal difference features of the landslide change features at different scales, and capture the semantic relationships between spatially disconnected landslide areas by introducing a nonlocal context enhancement strategy to obtain the first difference feature map; The second landslide extraction module is used to decode the first difference feature map through the decoder of the generator to obtain the second difference feature map, and introduces a landslide adaptive convolution strategy to dynamically adjust the weight of the convolution kernel according to the shape of the landslide in the second difference feature map, so as to refine the landslide boundary of the second difference feature map and obtain a predicted landslide mask; wherein, the landslide mask is used to mark the pixel position of the landslide in the second difference feature map; The generator optimization module is used to optimize the generator by comparing the difference between the predicted landslide mask and the actual landslide label using a discriminator, based on the temporal difference image, the predicted landslide mask, and the actual landslide label, to obtain a trained generator; wherein the temporal difference image is obtained by taking the absolute value of the difference sum of the two temporal remote sensing images, and the actual landslide label is obtained by taking the actual label of the two temporal remote sensing images.

Images