Modeling method and application of global landslide remote sensing rapid identification model based on improved DeepLab-V3 +
By improving the DeepLab-V3+ model and combining ASPP and CBAM mechanisms, a rapid identification model for global landslide remote sensing was constructed, which solved the problems of low efficiency and low accuracy in landslide surveys in existing technologies, and achieved efficient identification and accurate location of global landslides.
Patent Information
- Application Number
- CN202410853929.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-28
- Publication Date
- 2025-10-28
AI Technical Summary
Existing landslide survey methods are inefficient and inaccurate, failing to achieve efficient landslide identification on a global scale. Furthermore, the lack of rich datasets affects the learning ability and identification accuracy of the models.
The DeepLab-V3+ model was improved by increasing the dilated convolutional dilation rate of the ASPP module and introducing the CBAM attention mechanism in the model decoding stage to optimize the information content of the feature map and build a rapid global landslide remote sensing identification model.
It improves the accuracy and efficiency of landslide identification, enabling accurate identification of landslide areas worldwide and providing basic information to support disaster area reconstruction and recovery, hazard assessment, and early warning and forecasting.
Smart Images
Figure CN120852983A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of landslide disaster investigation, specifically to a modeling method and application of a global landslide remote sensing rapid identification model based on an improved DeepLab-V3+. Technical Background
[0002] Landslides are one of the most common natural disasters worldwide. Conducting a global survey of landslide hazard points, compiling a landslide catalog, and investigating the spatial distribution characteristics of landslides can provide basic information for disaster area reconstruction and recovery, landslide hazard assessment, and landslide early warning and forecasting worldwide.
[0003] Existing landslide survey methods mainly include manual field surveys and remote sensing image interpretation. Manual field surveys are time-consuming, inefficient, provide limited information, have poor timeliness, and require a large amount of manpower and resources. Remote sensing technology has the advantages of speed, large-scale coverage, and no physical contact, and can obtain remote sensing images of large areas. In recent years, domestic and foreign researchers have proposed a number of landslide interpretation methods based on optical remote sensing images and deep learning. However, these methods have been limited to landslide surveys in a certain region and cannot be applied to global areas. There are two main reasons for this: (1) The feature extraction accuracy of existing landslide identification models is low. Simply applying deep learning models to the field of landslide identification cannot fully utilize the learning ability of the models. (2) The richness of the dataset is insufficient. On the one hand, the types and number of landslide samples are insufficient, making it impossible for the model to learn the deep feature information of global landslides from the samples. On the other hand, the data categories are insufficient, and geological hazards with similar appearance characteristics to landslides will interfere with the accuracy of landslide identification models. Summary of the Invention
[0004] The purpose of this invention is to provide a modeling method and application for a global landslide remote sensing rapid identification model based on an improved DeepLab-V3+, so as to achieve effective identification of landslides in global land areas, compile a global landslide inventory, and provide basic information for studying the spatial distribution characteristics of landslides worldwide, disaster area reconstruction and recovery, landslide hazard assessment, and landslide early warning and forecasting.
[0005] To achieve the above objectives, the technical solution of the present invention is as follows:
[0006] A modeling method for rapid identification of global landslides based on improved DeepLab-V3+ remote sensing includes the following steps:
[0007] Step S1, Landslide Remote Sensing Data Acquisition: Collect landslide remote sensing data within the pre-selected target area, identify and extract the apparent features of landslide target categories as a dataset;
[0008] Step S2, landslide remote sensing dataset creation: Use annotation tools to classify the extracted target appearance features and organize the dataset in a data format suitable for training image semantic segmentation algorithm models;
[0009] Step S3, Improve the landslide identification algorithm: Based on the DeepLab-V3+ model, increase the dilation rate of the atrous convolution in the ASPP module during the model encoding stage, and introduce an attention mechanism during the model decoding stage to optimize the information content of the feature map;
[0010] Step S4, Construct a rapid landslide identification model: Train and test the model until the accuracy and mean intersection-over-union (mIOU) of the test reach the predetermined standards, and use this model for rapid identification of landslide areas;
[0011] Furthermore, in step S1, the categories in the target area landslide remote sensing dataset include landslides, rainfall-induced debris flows, and freeze-thaw-induced debris flows.
[0012] Furthermore, in step S3, during the model encoding stage, the following is included:
[0013] Step S301: Increase the dilation rate of the dilated convolution in the ASPP module, and input the feature map output by the dilated convolution into multiple deep dilated convolution (DDW) configurations, including DDW 1×1rate 1, DDW 3×3rate 2, DDW 3×3rate 4, DDW 3×3rate 8, and DDW 3×3rate 16.
[0014] Step S302: Perform 1×1 pointwise convolution, batch normalization and ReLU6 nonlinear activation on each DDW output, and stitch the processed feature maps into a comprehensive high-order feature map.
[0015] Furthermore, in step S3, during the model decoding stage, the following are included:
[0016] Step S303: Adjust the number of channels in the stride-4, stride-8, and stride-16 feature maps output by the backbone network to an appropriate level using 1×1 convolution, such as 48 or other channels adjusted according to actual application requirements, and apply the CBAM attention mechanism to filter background information and highlight the target region.
[0017] Step S304: Upsample the high-order feature map processed in step S303 by 4 times, upsample the lower resolution to a higher resolution, and superimpose it with the low-order feature map to form a fused feature map. For example, upsample the 32×32 feature map to 128×128 and superimpose it with the low-order feature map to form a fused feature map of 128×128×304.
[0018] Step S305: Use a 3×3 convolution to adjust the number of channels to an optimization level, for example, to 128 channels, and perform another 4x upsampling to complete the feature map optimization.
[0019] Further, in step S301, the plurality of DDW configurations include DDW 1×1rate 1, DDW 3×3rate 2, DDW 3×3rate 4, DDW 3×3rate 8 and DDW 3×3rate 16.
[0020] Further, step S4 includes:
[0021] Step S401, Model Training: The improved landslide identification algorithm in step S3 is trained using the landslide remote sensing dataset generated in step S2. At the same time, pre-trained model weights obtained from public databases are introduced to optimize training efficiency and initial model performance, and the maximum number of iterations is set to 300.
[0022] Step S402, Model Testing: Perform model performance testing. If the accuracy and mIOU of the test do not reach the preset threshold of 0.85, continue training. When the indicators reach the preset threshold, select the model as the rapid landslide identification model.
[0023] According to a second aspect of the present invention, the present invention provides an application of the above-mentioned global landslide remote sensing rapid identification model in the field of landslide disaster investigation, comprising the following steps:
[0024] Step S501: Acquire global remote sensing imagery and crop it into image patches of suitable size;
[0025] Step S502: Use the landslide fast identification model improved and trained based on steps S1 to S4 to identify landslide areas in these image blocks;
[0026] Step S503: Integrate the identified landslide area information to create a global landslide distribution map.
[0027] The present invention has the following advantages:
[0028] 1. Strong Feature Extraction Capability: The improved ASPP uses dilation rates of 2, 4, 8, and 16 for dilation convolution, which, compared to the original dilation rates of 6, 12, and 18, focuses on the feature information captured at dilation rates of 2 and 4. The dilation rates of 6, 12, and 18 are replaced with dilation rates of 8 and 16 to capture the information captured by the original dilation rates. These improvements can more effectively capture the contextual information of landslide images and enrich the semantic information of landslides in high-order feature maps.
[0029] 2. Rich information in low-level feature maps: stride-4, stride-8, and stride-16 feature maps are selected as low-level feature maps, instead of just stride-4 / stride-8 feature maps. This allows for the fusion of low-level feature map information at different scales, which helps the model to more accurately capture the positional information of large, medium, and small targets, refine the target boundaries, and improve the target recognition accuracy.
[0030] 3. High-quality feature maps: The CBAM attention mechanism is used to filter background information in both high-order and low-order feature maps, resolve noise points caused by different spectra of the same object in landslide images, highlight the target area, and improve the quality of positional and detail information in low-order feature maps and semantic information in high-order feature maps respectively. This allows the fused feature map of high-order and low-order features to contain rich positional and semantic information at the same time, thereby improving the accuracy of target recognition.
[0031] 4. High accuracy: By improving ASPP and using CBAM to filter information from high-order / low-order feature maps, the quality of the feature maps is significantly improved. This allows the fused feature maps to take into account both high-quality location and semantic information, which helps the model identify the irregular boundaries of landslides and thus identify the landslide area, thereby improving the recognition accuracy. Attached Figure Description
[0032] Figure 1 This is the overall technical roadmap of the model of this invention.
[0033] Figure 2 This is a data collection point for landslide datasets from the Qinghai-Tibet Plateau.
[0034] Figure 3 This is an example of a landslide dataset from the Qinghai-Tibet Plateau.
[0035] Figure 4 This is the CBAM model structure used in the model of this invention.
[0036] Figure 5 This invention uses an improved DeepLab-V3+ model structure.
[0037] Figure 6 This is an example of global landslide identification based on the application of this invention in the field of landslide disaster investigation.
[0038] Figure 7 This is a global landslide distribution map constructed based on the application of this invention in the field of landslide disaster investigation. Detailed Implementation
[0039] The preferred embodiments of the present invention mainly include a remote sensing dataset of landslides on the Qinghai-Tibet Plateau and an improved deep learning target recognition method. Figure 1 This is the overall technical roadmap of the system of the present invention. Figures 2-5 These are the various sub-modules of the present invention. Figure 6 This is a global example of landslide identification. Figure 7 This invention provides a global landslide distribution map obtained by applying the system to a global region. The technical solution of this invention will be further described below with reference to the accompanying drawings.
[0040] Step S1, Landslide Remote Sensing Data Acquisition: Collect landslide remote sensing data within the pre-selected target area, identify and extract the apparent features of landslide target categories as a dataset.
[0041] First, the categories in the Tibetan Plateau landslide remote sensing dataset were determined as landslides, rainfall-induced debris flows, and freeze-thaw-induced debris flows. Then, target locations were determined through field investigations and literature review, and the appearance characteristics of the three types of targets were identified in the remote sensing images. Next, based on the basic appearance characteristics of the three types of targets, relevant targets were collected from remote sensing images of the Tibet region and extracted as images (JPG format) for dataset creation. Data collection points are as follows: Figure 2 As shown.
[0042] Step S2, landslide remote sensing dataset creation: Use annotation tools to classify the extracted target appearance features and organize the dataset in a data format suitable for training image semantic segmentation algorithm models.
[0043] First, install and use LabelMe to annotate the target appearance features extracted in step S1, with each JPG file corresponding to a PNG file. Then, create a dataset according to the VOC data format. An example of the final landslide dataset is shown below. Figure 3 As shown.
[0044] Step S303: Adjust the number of channels in the stride-4, stride-8, and stride-16 feature maps output by the backbone network to an appropriate level using 1×1 convolution, such as 48 or other channels adjusted according to actual application requirements, and apply the CBAM attention mechanism to filter background information and highlight the target region.
[0045] The model decoding stage includes:
[0046] Step S301: Increase the dilation rate of the dilated convolution in the ASPP module, and input the feature map output by the dilated convolution into multiple deep dilated convolution (DDW) configurations, including DDW 1×1rate 1, DDW 3×3rate 2, DDW 3×3rate 4, DDW 3×3rate 8, and DDW 3×3rate 16.
[0047] Step S302: Perform 1×1 pointwise convolution, batch normalization and ReLU6 nonlinear activation on each DDW output, and stitch the processed feature maps into a comprehensive high-order feature map.
[0048] The model decoding stage includes:
[0049] Step S303: Adjust the number of channels in the stride-4, stride-8, and stride-16 feature maps output by the backbone network to an appropriate level using 1×1 convolution, such as 48 or other channels adjusted according to actual application requirements, and apply the CBAM attention mechanism to filter background information and highlight the target region.
[0050] Step S304: Upsample the high-order feature map processed in step S303 by 4 times, upsample the lower resolution to a higher resolution, and superimpose it with the low-order feature map to form a fused feature map. For example, upsample the 32×32 feature map to 128×128 and superimpose it with the low-order feature map to form a fused feature map of 128×128×304.
[0051] Step S305: Use a 3×3 convolution to adjust the number of channels to an optimization level, for example, to 128 channels, and perform another 4x upsampling to complete the feature map optimization.
[0052] CBAM structure as Figure 4 As shown, the improved model is as follows Figure 5 As shown.
[0053] Step S4: Construct a rapid landslide identification model: Train and test the model until the accuracy and mean intersection-over-union (mIOU) of the test reach the predetermined standards, and use this model for rapid identification of landslide areas.
[0054] Step S401, Model Training: The improved landslide identification algorithm in step S3 is trained using the landslide remote sensing dataset generated in step S2. At the same time, pre-trained model weights obtained from public databases are introduced to optimize training efficiency and initial model performance, and the maximum number of iterations is set to 300.
[0055] Step S402, Model Testing: Perform model performance testing. If the accuracy and mIOU of the test do not reach the preset threshold of 0.85, continue training. When the indicators reach the preset threshold, select the model as the rapid landslide identification model.
[0056] First, the model is trained. An improved landslide identification algorithm is trained using a plateau landslide remote sensing dataset, employing pre-trained model weights and setting a maximum of 300 iterations. Then, the model is tested. The trained model is tested, and training continues if the accuracy and mIOU are below 0.85. When the accuracy and mIOU are above 0.85, the model is selected as the fast landslide identification model.
[0057] According to the application of the global landslide remote sensing rapid identification model provided by the present invention in the field of landslide disaster investigation, a global landslide distribution map is constructed, including the following steps:
[0058] Step S501: Acquire global remote sensing imagery and crop it into image patches of suitable size;
[0059] Step S502: Use the landslide fast identification model improved and trained based on steps S1 to S4 to identify landslide areas in these image blocks;
[0060] Step S503: Integrate the identified landslide area information to create a global landslide distribution map.
[0061] Specifically, global remote sensing imagery is acquired from Google Earth, cropped into 640*640 pixel image patches, and input into a rapid landslide identification model to identify landslide areas. An example of such identification is shown below. Figure 6 As shown. The recognition results are as follows. Figure 7 As shown, this is integrated into a global landslide distribution map.
[0062] Globally, a total of 10,016 landslide areas have been identified. Of these, 2,304 were identified on the Tibetan Plateau, and 7,712 were identified in other regions. Their specific distribution is as follows:
[0063] Africa: 520 landslide areas
[0064] Asia (excluding the Tibetan Plateau): 5,772 landslide areas
[0065] Europe: 180 landslide areas
[0066] North America: 844 landslide areas
[0067] South America: 344 landslide areas
[0068] Australia: 572 landslide areas
[0069] The statistics on landslide numbers show that the Tibetan Plateau region accounts for 28.5% of all landslides identified in Asia and 23.1% of all landslides identified globally. The Tibetan Plateau is not only the region with the most widespread landslide distribution in Asia, but also a major concentration area of landslide disasters worldwide.
Claims
1. A method for modeling a rapid global landslide remote sensing identification model based on an improved DeepLab-V3+, characterized in that, Includes the following steps: Step S1, Landslide Remote Sensing Data Acquisition: Collect landslide remote sensing data within the pre-selected target area, identify and extract the apparent features of landslide target categories as a dataset; Step S2, landslide remote sensing dataset creation: Use annotation tools to classify the extracted target appearance features and organize the dataset in a data format suitable for training image semantic segmentation algorithm models; Step S3, Improve the landslide identification algorithm: Based on the DeepLab-V3+ model, increase the dilation rate of the atrous convolution in the ASPP module during the model encoding stage, and introduce an attention mechanism during the model decoding stage to optimize the information content of the feature map; Step S4: Construct a rapid landslide identification model: Train and test the model until the accuracy and mean intersection-over-union (mIOU) of the test reach the predetermined standards, and use this model for rapid identification of landslide areas.
2. The method for rapid identification and modeling of global landslides using remote sensing according to claim 1, characterized in that, In step S1, the categories in the target area landslide remote sensing dataset include landslides, rainfall-induced debris flows, and freeze-thaw-induced debris flows.
3. The method for rapid identification and modeling of global landslides using remote sensing according to claim 1, characterized in that, Step S3, in the model encoding stage, includes: Step S301: Increase the dilation rate of the dilated convolution in the ASPP module, and input the feature map output by the dilated convolution into multiple deep dilated convolution (DDW) configurations respectively; Step S302: Perform 1×1 pointwise convolution, batch normalization and ReLU6 nonlinear activation on each DDW output, and stitch the processed feature maps into a comprehensive high-order feature map.
4. The method for rapid identification and modeling of global landslides using remote sensing according to claim 1, characterized in that, Step S3, in the model decoding stage, includes: Step S303: Adjust the channels of the stride-4, stride-8, and stride-16 feature maps output by the backbone network to an appropriate level using 1×1 convolution, and apply the CBAM attention mechanism to filter background information and highlight the target region; Step S304: Upsample the high-order feature map processed in step S303 by 4 times, upsample the lower resolution to a higher resolution, and superimpose it with the low-order feature map to form a fused feature map; Step S305: Adjust the number of channels to an optimization level using a 3×3 convolution and perform another 4x upsampling to complete the feature map optimization.
5. The method for rapid identification and modeling of global landslides using remote sensing according to claim 3, characterized in that, In step S301, the plurality of DDW configurations include DDW 1×1rate 1, DDW 3×3rate 2, DDW 3×3rate 4, DDW 3×3rate 8 and DDW 3×3rate 16.
6. The method for rapid identification and modeling of global landslides using remote sensing according to claim 1, characterized in that, Step S4 includes: Step S401, Model Training: Use the landslide remote sensing dataset generated in step S2 to train the improved landslide identification algorithm in step S3, while introducing pre-trained model weights obtained from public databases and setting the maximum number of iterations; Step S402, Model Testing: Perform model performance testing. If the accuracy and mIOU of the test do not reach the preset threshold, continue training. When the indicators reach the preset threshold, select the model as the rapid landslide identification model.
7. The application of the global landslide remote sensing rapid identification model as described in claim 1 in the field of landslide disaster investigation, characterized in that, Includes the following steps: Step S501: Acquire global remote sensing imagery and crop it into image patches of suitable size; Step S502: Use the landslide fast identification model improved and trained based on steps S1 to S4 to identify landslide areas in these image blocks; Step S503: Integrate the identified landslide area information to create a global landslide distribution map.