A method for identifying active landslides

Abstract

The application discloses an active landslide identification method, and relates to the field of disaster identification, which comprises the following steps: acquiring image data of a region to be identified, and constructing an active landslide data set; constructing a context-aware adaptive fusion model; the context-aware adaptive fusion model comprises an encoder and a decoder; the encoder comprises a plurality of feature processing layers and a convolution layer connected in sequence; the feature processing layer comprises a convolution layer and a wavelet transform down-sampling WDB module connected in sequence; the decoder comprises a plurality of multi-branch scale adaptive aggregation MSA modules, and a plurality of feature extraction refinement layers and a convolution layer connected in sequence; the feature extraction refinement layer comprises a convolution layer and a convolution up-sampling CUB module connected in sequence; based on the active landslide data set, the context-aware adaptive fusion model is trained to obtain an active landslide identification model; the active landslide data set to be detected is input into the active landslide identification model to obtain an active landslide identification result.

Description

Technical Field

[0001] This invention relates to the field of disaster identification technology, and in particular to a method for identifying active landslides. Background Technology

[0002] As a common geological hazard, landslides pose a serious threat to life and cause significant property damage to people worldwide. Landslide events are typically accompanied by a slow, active deformation phase, which may manifest as the generation and expansion of cracks in the surface soil and rocks, forming large active areas over time, and eventually leading to landslide instability. Timely identification of active deformation areas of active landslides (ALs), and subsequent monitoring and early warning of high-risk areas, is crucial for disaster prevention and mitigation.

[0003] Synthetic Aperture Radar Interferometry (InSAR) technology can capture minute displacements of potential active landslides, thus distinguishing between unstable and stable slopes over a large area, and is currently widely used in the identification of active landslides (ALs). Existing InSAR-based methods for identifying active landslides (ALs) mainly include manual visual interpretation, cluster analysis, and thresholding. Among these, manual visual interpretation is time-consuming and labor-intensive and relies on expert knowledge; cluster analysis has poor adaptability to complex terrain; and thresholding relies on expert experience to determine the threshold, resulting in subjective results and difficulty in handling low-contrast areas, thus affecting the accuracy of active landslide identification. In addition, although deep learning methods can automatically identify ALs in large-scale terrain, current research still suffers from problems such as loss of small-scale landslide features and poor boundary extraction, leading to insufficient accuracy in identifying active landslides (ALs). Summary of the Invention

[0004] This invention provides a method for identifying active landslides (ALs) to address the aforementioned problems in the prior art, namely, how to improve the accuracy of active landslide (ALs) identification in the prior art. The method includes:

[0005] Acquire image data of the area to be identified and construct an active landslide dataset;

[0006] Construct a context-aware adaptive fusion model; the context-aware adaptive fusion model includes an encoder and a decoder;

[0007] The encoder includes multiple feature processing layers and a convolutional layer connected in sequence; the feature processing layer includes a convolutional layer and a wavelet transform downsampling (WDB) module connected in sequence.

[0008] The decoder includes multiple multi-branch scale adaptive aggregation (MSA) modules for receiving feature inputs from two adjacent layers, adaptively allocating layer weights according to the target features, and fusing them; as well as multiple feature extraction and refinement layers and a convolutional layer connected in sequence; the feature extraction and refinement layers include convolutional layers connected in sequence, and a convolutional upsampling (CUB) module for restoring the resolution of features and refining boundary information.

[0009] The multi-branch scale adaptive aggregation (MSA) module receives feature inputs from two adjacent layers, adaptively allocates layer weights based on the target features, and fuses them, including the following steps:

[0010] 1×1 convolution is used to enhance the low-level features of the input, and 1×1 convolution and bilinear interpolation are used to adjust the number of channels and spatial resolution of the high-level features of the input.

[0011] Based on the adjusted low-level and high-level features, feature components are determined, and the weights of the feature components are determined using the sigmoid function. Then, each group of weighted high-level features is concatenated with the corresponding weighted low-level features to generate fused features. Layer normalization is then applied to normalize each group of fused features. Next, dilated convolution is used to extract multi-scale features from the four normalized fused features. Finally, the extracted fused features are concatenated, and the number of channels is adjusted using a 1×1 convolution to output the fused features.

[0012] The convolutional upsampling CUB module restores the resolution of the features and refines the boundary information, including the following steps:

[0013] The input features are extracted and refined using a 3×3 convolutional layer, then upsampled by a factor of 2 before being input into two parallel depthwise separable convolutional branches. 1×3+3×1 and 3×1+1×3 convolutional kernels are used to extract the horizontal and vertical spatial features of the input features. The outputs of the parallel branches are then fused using a pixel-wise addition method, and the fused features are enhanced using the ReLU activation function. Finally, a 1×1 convolution is used to adjust the number of channels in the feature map.

[0014] The active landslide dataset is input into the context-aware adaptive fusion model for training, and an active landslide identification model for landslide identification is obtained.

[0015] Input the dataset of active landslides to be detected into the active landslide recognition model to obtain the active landslide recognition results.

[0016] Optionally, acquiring image data of the area to be identified and constructing an active landslide dataset specifically includes:

[0017] Image data of the area to be identified is obtained by using synthetic aperture radar interferometry (InSAR) technology. The average deformation rate is obtained by using InSAR stacking technology based on the surface deformation image data.

[0018] The average deformation rate value is mapped to the RGB color space to determine the raster image. The raster image is then segmented by sliding window to generate an active landslide dataset for active landslide identification.

[0019] Optionally, the downsampling of the wavelet transform downsampling (WDB) module includes the following steps:

[0020] The input features are decomposed into four frequency domain feature components by Haar wavelet transform. These four frequency domain components are concatenated along the channel dimension and then the feature representation is refined by a 1×1 convolution. The four frequency domain feature components include one low-frequency component and three high-frequency components.

[0021] Also includes:

[0022] The active landslide identification results are post-processed using the average slope value to obtain the landslide connected components. By performing connected component analysis on the landslide connected components, the set of slope values ​​of each landslide connected component in the corresponding slope map is extracted, and the average slope of the connected components is determined. Each landslide connected component is filtered by setting a slope threshold, and the filtered connected components are recombined to generate optimized active landslide identification results.

[0023] Optionally, the active landslide dataset can be filtered and denoised.

[0024] Compared to existing technologies, the beneficial effects of this invention are as follows: This invention provides an active landslide identification method. This method acquires image data of the area to be tested and constructs an active landslide dataset based on the image data. The context-aware adaptive fusion model constructed by this invention includes partial convolutional layers and an encoder and decoder obtained from three modules proposed in this invention: WDB, MSA, and CUB. Training the context-aware adaptive fusion model based on the active landslide dataset yields an active landslide identification model for landslide identification. This model can then identify deformation zones in the landslide dataset to be identified, resulting in accurate active landslide (ALS) identification results. Furthermore, the WDB module proposed in this invention is used for downsampling during the encoding stage, which preserves more detailed information while performing downsampling. This effectively solves the problem of small-scale landslide feature loss caused by pooling layers. Secondly, based on the output features in the encoding stage, the MSA module proposed in this invention receives feature inputs from two adjacent layers, adaptively assigns appropriate layer weights according to the target features, and fuses them, achieving more accurate feature alignment and effectively avoiding the loss of small-scale landslide features. In the decoding stage, an efficient CUB module is introduced to gradually restore the feature resolution and further refine the boundary information, thereby more robustly enhancing the feature expression capability and effectively improving the recognition accuracy of active landslides. In addition, by using the average slope value to perform connected component analysis on the recognition results, and then recombining the selected connected components, optimized ALs recognition results are generated, thereby effectively improving the accuracy of active landslide ALs recognition results. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0026] Figure 1 A flowchart illustrating an active landslide identification method provided in an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of the dataset structure and deformation zone identification of an active landslide identification method provided in an embodiment of the present invention;

[0028] Figure 3 A context-aware adaptive fusion model framework diagram of an active landslide identification method provided in an embodiment of the present invention;

[0029] Figure 4 This is a comparison chart of active landslide identification results provided by an embodiment of the present invention for an active landslide identification method;

[0030] Figure 5 This is a schematic diagram of the identification results of the Haiyuan migration test area for an active landslide identification method provided in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0032] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0033] Example 1

[0034] like Figure 1 and Figure 2 As shown in the figure, this embodiment illustrates a method for identifying active landslides, including:

[0035] S1: Obtain image data of the area to be identified and construct an active landslide dataset.

[0036] Optionally, image data of the area to be identified is obtained by using synthetic aperture radar interferometry (InSAR) technology. Based on the surface deformation image data, the average deformation rate is obtained using InSAR stacking technology. The average deformation rate value is mapped to the RGB color space to determine the raster image. The raster image is then segmented by sliding window to generate an active landslide dataset for active landslide identification.

[0037] For example, InSAR Stacking technology can be used to obtain the average deformation rate of the area to be identified in the line-of-sight (LOS) direction of the radar. The average deformation rate value is then mapped to the RGB color space and converted into a raster image. The raster image is then segmented by sliding window to generate an active landslide dataset for automatic identification of ALS.

[0038] S2: Construct a context-aware adaptive fusion model; the context-aware adaptive fusion model includes an encoder and a decoder; the encoder includes multiple feature processing layers and a convolutional layer connected in sequence; the feature processing layer includes a convolutional layer and a wavelet transform downsampling (WDB) module connected in sequence; the decoder includes multiple multi-branch scale adaptive aggregation (MSA) modules for receiving feature inputs from two adjacent layers, adaptively allocating layer weights according to the target features, and fusing them, as well as multiple feature extraction and refinement layers and a convolutional layer connected in sequence; the feature extraction and refinement layer includes a convolutional upsampling (CUB) module for restoring the resolution of features and refining boundary information.

[0039] For example, the wavelet transform downsampling (WDB) module first decomposes the input features into four frequency domain feature components using Haar wavelet transform. These four components include a low-frequency component containing global features and three high-frequency components containing local details in the horizontal, vertical, and diagonal directions, respectively. These four frequency domain components are concatenated along the channel dimension to aggregate multi-scale information and generate a new feature representation. Finally, a 1×1 convolution refines the feature representation while adjusting the channel dimension to filter out redundant information.

[0040] For example, the multi-branch scale adaptive aggregation (MSA) module takes two inputs: low-level features and high-level features. It uses 1×1 convolutions to enhance the low-level features, while simultaneously adjusting the channel count and spatial resolution of the high-level features using 1×1 convolutions and bilinear interpolation to align them with the low-level feature space. Then, the adjusted high-level and low-level features are mapped into four sets of feature components according to the channel dimension, and a sigmoid function is applied. Calculate the weights w1, w2, w3, and w4 for the four groups of high-level feature components, and the corresponding weights for the four groups of low-level feature components are 1-w1, 1-w2, 1-w3, and 1-w4. During the weight allocation process, when the weight of a certain set of high-level feature components is greater than 0.5, high-level features are used first to enhance contextual information; conversely, low-level features, that is, fine-grained features, are used first to enhance detailed information.

[0041] Subsequently, each weighted high-level feature group is concatenated with the corresponding weighted low-level feature group to generate four fused features, and layer normalization is applied to normalize each fused feature group.

[0042] Next, dilated convolution is used to extract multi-scale features from the four normalized fused features. The kernel size is fixed at 3, and the dilation rates are set to {1, 2, 4, 8} to capture contextual information at different scales. Finally, the four fused features are concatenated according to the channel dimension, and the number of channels is adjusted by a 1×1 convolution to achieve interaction and integration between features at different scales, ultimately outputting the fused features.

[0043] For example, the Convolutional Upsampling (CUB) module first extracts and refines the input features using a 3×3 convolutional layer, then upsamples them by a factor of 2 before inputting them into two parallel depthwise separable convolutional branches. These two branches use 1×3+3×1 and 3×1+1×3 convolutional kernels respectively, instead of the traditional 3×3 convolution, to extract spatial features in the horizontal and vertical directions, effectively reducing computational complexity while maintaining the receptive field of the original large convolutional kernel. The outputs of the parallel branches are fused pixel-by-pixel, thus integrating spatial features from different directions. The fused features are then normalized to ensure feature stability. Next, the ReLU activation function is applied to further enhance the features and increase boundary activation values. Finally, a 1×1 convolution is used to adjust the number of channels in the feature map, and edge information extraction is optimized through residual structures to enhance feature learning.

[0044] S3: Input the active landslide dataset into the context-aware adaptive fusion model for training to obtain an active landslide identification model for landslide identification.

[0045] S4: Input the active landslide dataset to be detected into the active landslide recognition model and obtain the active landslide recognition results.

[0046] For example, the average slope value can be used to post-process the active landslide identification results. Specifically, this includes performing connectivity analysis on the landslide connectivity based on the active landslide identification results, extracting the set of slope values ​​for each landslide connectivity in the corresponding slope map, determining the average slope of the connectivity, filtering each landslide connectivity by setting a slope threshold, and recombining the filtered connectivity to generate optimized active landslide identification results.

[0047] For example, the dataset of active landslides to be detected is input into the active landslide identification model for deformation zone identification. After obtaining the identification results, the average slope value can be used for post-processing. Specifically, connected component analysis is performed on the preliminary identification results to extract the slope value set of each target connected component in the corresponding slope map, and the average slope of the connected component is calculated. In this embodiment, the average slope of the connected component is selected as 5° as the judgment threshold. When the average slope of the connected component is greater than 5°, the area is considered a landslide and retained; otherwise, it is removed and determined to be a non-landslide area. The selection of the slope threshold can be adjusted according to the specific situation of the target area. Finally, the connected components filtered by slope constraints are recombined to generate optimized ALs identification results.

[0048] Example 2

[0049] Taking the upper reaches of a river in China as an example, a distribution map of ALs (alveolar sludge) locations in the upper reaches of the river was obtained. The locations of ALs were mainly detected based on millimeter- to meter-level deformation information, which determined the training data area, the marine migration test area, and the locations of active landslides for the river.

[0050] Figure 4 This is a comparison of the recognition results. The first column shows the deformation rate after color mapping, the second column shows the labels, and the third to sixth columns show the landslide extraction results of different networks. White indicates correctly extracted landslide areas, red indicates false detections (background incorrectly identified as landslides), and green indicates missed detections (landslide areas incorrectly identified as background). This study achieved a recognition rate of 83.3% on the active landslide deformation dataset, which is 4.4% higher than ResUNet and 8.9% higher than PSPNet. DeeplabV3+ performed the worst, with a recognition rate of only 66.6%, which is 16.7% lower than the recognition result of the method used in this invention. Figure 5 The image shows the results of identifying active landslides in the Haiyuan migration test area. The red dashed box in the image represents the boundary of the Haiyuan migration test area, and A, B, C, and D are magnified views of the results within the area.

[0051] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for identifying active landslides, characterized in that, include: Acquire image data of the area to be identified and construct an active landslide dataset; Construct a context-aware adaptive fusion model; the context-aware adaptive fusion model includes an encoder and a decoder; The encoder includes multiple feature processing layers and a convolutional layer connected in sequence; the feature processing layer includes a convolutional layer and a wavelet transform downsampling (WDB) module connected in sequence. The decoder includes multiple multi-branch scale adaptive aggregation (MSA) modules for receiving feature inputs from two adjacent layers, adaptively allocating layer weights according to the target features, and fusing them; as well as multiple feature extraction and refinement layers and a convolutional layer connected in sequence; the feature extraction and refinement layers include convolutional layers connected in sequence, and a convolutional upsampling (CUB) module for restoring the resolution of features and refining boundary information. The multi-branch scale adaptive aggregation (MSA) module receives feature inputs from two adjacent layers, adaptively allocates layer weights based on the target features, and fuses them, including the following steps: 1×1 convolution is used to enhance the low-level features of the input, and 1×1 convolution and bilinear interpolation are used to adjust the number of channels and spatial resolution of the high-level features of the input. Based on the adjusted low-level and high-level features, feature components are determined, and the weights of the feature components are determined using the sigmoid function. Then, each group of weighted high-level features is concatenated with the corresponding weighted low-level features to generate fused features. Layer normalization is then applied to normalize each group of fused features. Next, dilated convolution is used to extract multi-scale features from the four normalized fused features. Finally, the extracted fused features are concatenated, and the number of channels is adjusted through a 1×1 convolution to output the fused features. The convolutional upsampling CUB module restores the resolution of the features and refines the boundary information, including the following steps: The input features are extracted and refined using a 3×3 convolutional layer, then upsampled by a factor of 2 before being input into two parallel depthwise separable convolutional branches. 1×3+3×1 and 3×1+1×3 convolutional kernels are used to extract the horizontal and vertical spatial features of the input features. The outputs of the parallel branches are then fused using a pixel-wise addition method, and the fused features are enhanced using the ReLU activation function. Finally, a 1×1 convolution is used to adjust the number of channels in the feature map. The active landslide dataset is input into the context-aware adaptive fusion model for training, and an active landslide identification model for landslide identification is obtained. Input the dataset of active landslides to be detected into the active landslide recognition model to obtain the active landslide recognition results.

2. The active landslide identification method as described in claim 1, characterized in that, The process of acquiring image data of the area to be identified and constructing an active landslide dataset specifically includes: Image data of the area to be identified is obtained by using synthetic aperture radar interferometry (InSAR) technology. The average deformation rate is obtained by using InSAR stacking technology based on the surface deformation image data. The average deformation rate value is mapped to the RGB color space to determine the raster image. The raster image is then segmented by sliding window to generate an active landslide dataset for active landslide identification.

3. The active landslide identification method as described in claim 1, characterized in that, The downsampling of the wavelet transform downsampling (WDB) module includes the following steps: The input features are decomposed into four frequency domain feature components by Haar wavelet transform. These four frequency domain components are concatenated along the channel dimension and then the feature representation is refined by a 1×1 convolution. The four frequency domain feature components include one low-frequency component and three high-frequency components.

4. The active landslide identification method as described in claim 1, characterized in that, Also includes: The active landslide identification results are post-processed using the average slope value to obtain the landslide connected components. By performing connected component analysis on the landslide connected components, the set of slope values ​​of each landslide connected component in the corresponding slope map is extracted, and the average slope of the connected components is determined. Each landslide connected component is filtered by setting a slope threshold, and the filtered connected components are recombined to generate optimized active landslide identification results.

5. The active landslide identification method as described in claim 1, characterized in that, The active landslide dataset is then filtered and denoised.

Images