Landslide extraction method, system, device and medium based on geological features

Abstract

The application discloses a landslide extraction method and system based on geological characteristics, and relates to the technical field of landslide extraction. The method comprises the following steps: collecting remote sensing images, digital elevation models and terrain parameters of a region to be detected; inputting the remote sensing images, the digital elevation models and the terrain parameters into a landslide extraction model to output a landslide extraction result; the landslide extraction model is obtained by training a deep convolutional neural network according to a training set; when features are extracted in the deep convolutional neural network, a hollow convolution and a spatial pyramid pooling module are adopted; and the area of a landslide is determined according to the landslide extraction result and the digital elevation model. The application improves the accuracy of landslide extraction.

Description

Technical Field

[0001] This invention relates to the field of landslide extraction technology, and in particular to a landslide extraction method, system, equipment and medium based on geoscientific features. Background Technology

[0002] Landslides are highly dangerous geological hazards that impact human safety and economic development. Traditional landslide extraction methods rely on the expertise of geological specialists, which are ineffective and inefficient. Therefore, researchers have employed deep learning techniques to construct landslide extraction models to address the shortcomings of traditional methods. However, the accuracy of current deep learning-based landslide extraction models still needs improvement. Summary of the Invention

[0003] The purpose of this invention is to provide a landslide extraction method, system, equipment, and medium based on geoscientific features, which improves the accuracy of landslide extraction.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A landslide extraction method based on geoscientific features includes:

[0006] Collect remote sensing images, digital elevation models, and terrain parameters of the area to be detected;

[0007] The remote sensing image, the digital elevation model, and the terrain parameters are input into the landslide extraction model, and the landslide extraction result is output. The landslide extraction model is obtained by training a deep convolutional neural network based on a training set. When performing feature extraction in the deep convolutional neural network, dilated convolution and spatial pyramid pooling modules are used.

[0008] The landslide area is determined based on the landslide extraction results and the digital elevation model.

[0009] Optionally, the deep convolutional neural network includes an encoding structure and a decoding structure; the encoding structure adopts an improved deep residual network, wherein the first layer of the improved deep residual network uses a spatial pyramid pooling module to extract features from the input data, and the third and fourth layers of the improved deep residual network use dilated convolutions to replace the original pooling layers; the decoding structure is an upsampling process, wherein a flow alignment module is used to learn the semantic flow between feature maps of adjacent layers during the upsampling process.

[0010] Optionally, the improved deep residual network is the improved ResNet50.

[0011] Optionally, the terrain parameters are extracted from the digital elevation model, and the terrain parameters include slope and aspect.

[0012] Optionally, it also includes classifying landslides into small, medium, and large landslides based on their area, wherein the area of ​​the small landslides ranges from 0 to 10. 4 m 2 The area of ​​medium-sized landslides ranges from 10. 4 m 2 ~10 5 m 2 The area of ​​the large landslides ranges from 10 5 m 2 ~10 6 m 2 .

[0013] Optionally, each sample data in the training set includes samples from landslide areas and samples from non-landslide areas;

[0014] The landslide area sample includes input data and label data; the non-landslide area sample includes input data; the input data consists of remote sensing images, digital elevation models, and terrain parameters of the sample area, and the label data consists of landslide edge images of the sample area.

[0015] This invention also discloses a landslide extraction system based on geoscientific features, comprising:

[0016] The data acquisition module is used to acquire remote sensing images, digital elevation models, and terrain parameters of the area to be detected.

[0017] The landslide extraction module is used to input the remote sensing image, the digital elevation model, and the terrain parameters into the landslide extraction model and output the landslide extraction result. The landslide extraction model is obtained by training a deep convolutional neural network based on a training set. When performing feature extraction in the deep convolutional neural network, dilated convolution and spatial pyramid pooling modules are used.

[0018] The landslide area determination module is used to determine the landslide area based on the landslide extraction results and the digital elevation model.

[0019] The present invention also discloses an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the landslide extraction method based on the described geoscientific features.

[0020] The present invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the landslide extraction method based on geoscientific features.

[0021] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0022] This invention uses remote sensing imagery, digital elevation model (DEM), and terrain parameters as multi-source data inputs to a landslide extraction model. The landslide extraction model employs a well-trained deep convolutional neural network with dilated convolution and pyramid pooling, which improves the accuracy of landslide extraction and can determine the landslide area based on the landslide extraction results and the digital elevation model. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. 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.

[0024] Figure 1 A schematic diagram of a landslide extraction method based on geoscientific features provided in an embodiment of the present invention;

[0025] Figure 2 This is a general technical roadmap for automatic landslide extraction provided in the embodiments of the present invention;

[0026] Figure 3 This is a schematic diagram of the spatial pyramid pooling module structure provided in an embodiment of the present invention;

[0027] Figure 4 A schematic diagram of a deep convolutional neural network structure provided in an embodiment of the present invention;

[0028] Figure 5 This is a schematic diagram of a landslide extraction system based on geoscientific features, provided as an embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] The purpose of this invention is to provide a landslide extraction method, system, equipment, and medium based on geoscientific features, which improves the accuracy of landslide extraction.

[0031] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] Example 1

[0033] like Figure 1As shown in the figure, this embodiment provides a landslide extraction method based on geoscientific features, which includes the following steps.

[0034] Step 101: Collect remote sensing images, digital elevation models, and terrain parameters of the area to be detected.

[0035] Step 102: Input the remote sensing image, the digital elevation model, and the terrain parameters into the landslide extraction model, and output the landslide extraction result; the landslide extraction model is obtained by training a deep convolutional neural network based on the training set, and the deep convolutional neural network uses dilated convolution and spatial pyramid pooling modules for feature extraction.

[0036] Step 103: Determine the landslide area based on the landslide extraction results and the digital elevation model.

[0037] Landslide extraction model construction approach: Landslides are mostly distributed in steep mountainous areas and river valleys. For landslide areas, texture changes, tonal differences, and their relationship with surrounding land features are key features for identifying and extracting landslide areas, and are also crucial to the model training accuracy. Details are as follows... Figure 2 As shown. Figure 2 The spatial relationships are represented by label data, which includes not only the landslide itself but also the types of land features surrounding it. Figure 2 From left to right, the images are remote sensing images, the deep learning process, and the intelligent interpretation results.

[0038] like Figure 4 As shown, the deep convolutional neural network includes an encoding structure and a decoding structure. The encoding structure adopts an improved deep residual network, wherein the first layer of the improved deep residual network uses a spatial pyramid pooling module to extract features from the input data, and the third and fourth layers of the improved deep residual network replace the original pooling layers with dilated convolutions. The decoding structure is an upsampling process, in which a flow alignment module is used to learn the semantic flow between feature maps of adjacent layers.

[0039] The improved deep residual network is the improved ResNet50.

[0040] Landslides are classified into small, medium, and large landslides based on their area. Small landslides have an area ranging from 0 to 10 square meters. 4 m 2 The area of ​​medium-sized landslides ranges from 10. 4 m 2 ~10 5 m 2 Large landslides have an area greater than 10 5 m 2 .

[0041] Each sample in the training set includes samples from landslide areas and samples from non-landslide areas.

[0042] The landslide area sample includes input data and label data; the non-landslide area sample includes input data; the input data consists of remote sensing images, digital elevation models, and terrain parameters of the sample area, and the label data consists of landslide edge images of the sample area.

[0043] The training process of a deep convolutional neural network specifically includes:

[0044] First, high-resolution remote sensing images and digital elevation models (DEMs) of the sample area were acquired. Based on color variations and textures in the high-resolution remote sensing images, combined with visual interpretation, the boundaries of the landslide area were located. Then, topographic parameters of the landslide area were extracted based on DEM topographic analysis. Topographic parameters include slope and aspect.

[0045] High-resolution remote sensing imagery is fused with DEM, slope, and aspect parameters to generate multi-source fusion data for model training.

[0046] A labeled dataset was established based on multi-source fused data and corresponding labeled data, including samples from both landslide and non-landslide areas. Two-thirds of the samples were randomly selected from the dataset as the training set for training the model, while the remaining one-third was used as the test set to test model accuracy. During training, the input samples were used to generate landslide predictions through a constructed deep convolutional neural network (DCNN). A loss function was then calculated between the predictions and the labeled samples, and gradient backpropagation was performed to adjust the model parameters. After multiple rounds of training, the model's prediction accuracy on the test set was ensured to reach a certain level.

[0047] The essence of deep learning is to construct a neural network model with multiple hidden layers and learn more effective features using massive amounts of training data, thereby ultimately improving prediction accuracy. Therefore, deep models are the means, and feature learning is the goal. The landslide extraction model, by combining the spatial and spectral characteristics of the landslide area itself and incorporating geoscientific knowledge as prior information into the model, has developed a landslide extraction model with geoscientific knowledge-guided features, improving the model's geographical perception capabilities.

[0048] Key technologies in landslide extraction models include multi-scale adaptive land cover extraction technology and network structure simplification and optimization technology based on parameter reuse.

[0049] (1) Multi-scale adaptive extraction technology for ground features.

[0050] Geoscientific observations are characterized by multi-scale nesting; different resolutions can cause a "scale effect" on different land features, affecting the model's recognition capabilities. Even at the same resolution, land features exhibit different scale characteristics; for example, a landslide of a few square meters and a landslide of tens of thousands of square meters are both considered landslides in the land feature classification system. Therefore, this invention introduces a pyramid pooling (SPP) structure to further capture the characteristics of landslides at different scales.

[0051] from Figure 3 It can be observed that the pyramid pooling module integrates features from four different scales. Figure 3 The top layer after pooling is the coarsest level, outputting a single feature vector (bin) generated using global pooling. The remaining three layers divide the input feature map into several different sub-regions, pool each sub-region, and finally combine the pooled bins containing location information. To preserve the weights of the global features, a 1×1 convolution kernel is used after each pyramid level. Then, the low-dimensional feature map is upsampled using bilinear interpolation to make it the same scale as the original feature map. Finally, the feature maps from different levels are concatenated (CONCAT) to form the final pyramid-pooled global feature. This aligns with the multi-scale nesting in geoscientific features. Figure 3 CONV in the text represents convolution.

[0052] (2) Network structure simplification and optimization techniques based on parameter reuse.

[0053] The deep convolutional neural network adopts an encoder-decoder structure. In the encoder structure, an improved deep residual network is constructed, using ResNet50 as the backbone. The feature extraction module is also improved: since the resolution of input images may vary and the size of landslides differs, ResNet Layer 1 (the first layer) employs a scale pyramid + feature fusion structure to capture features at different scales. ResNet layers 3-4 (the third and fourth layers) replace the pooling layers with dilated convolutions to capture peripheral information of the feature images, capturing more semantic information without reducing image size, thus improving recognition accuracy. In the decoder structure, this invention improves the upsampling process by employing a flow alignment module (FAM) to learn the semantic flow between adjacent layer feature maps. FAM takes features from adjacent layers as input, generates a flow field, and finally warps coarse features to refined features with high resolution based on this flow field, effectively propagating high-level features to high-resolution features and aligning them. In the experiment, the extraction model was replaced with other models for comparison, including UNet, Attention-UNet, Deeplab, etc. The study shows that the feature extraction module is easier to optimize as a residual network, and combining dilated convolution and pyramid pooling (SPP) in feature extraction can improve the accuracy of landslide extraction.

[0054] In a specific implementation, the encoding structure comprises four convolutional groups, each containing multiple basic convolutional computation processes. The first convolutional group receives an input image of size 5×256×256, the second convolutional group receives a feature map of size 64×128×128, the third convolutional group receives a feature map of size 128×64×64, and the fourth convolutional group receives a feature map of size 256×32×32. The feature maps output from the third and fourth groups are input to a first-stream alignment module, which generates a first offset field. This first offset field warps the coarse features output from the fourth group to a first refined feature with high resolution. The first refined feature is input to the decoding structure. The feature map output from the second group and the output (64×64) of the first upsampling layer of the decoding structure are input to a second-stream alignment module, which generates a second offset field. This second offset field warps the coarse features from the previous group to a second refined feature with high resolution. The second refined feature undergoes two more upsampling operations, ultimately outputting a feature classification map.

[0055] In landslide extraction model applications, remote sensing images to be detected are fused with topographic parameters to generate multi-band fused data. Then, a trained landslide extraction model is used to classify and detect the data, achieving accurate segmentation and extraction of landslide areas.

[0056] The trained model is applied to new imagery to extract landslides and obtain their edges, size, and topographic information. Combining the identified location and edge information with the DEM (Digital Image Model) allows for landslide classification, providing further insight into their nature. Here, landslides are categorized into small landslides (~10 mm) based on their area size. 4 m 2 Medium-sized landslides (~10) 5 m 2 Large landslides (~10) 6 m 2 The area calculation formula is shown below. Furthermore, the landslide area topography can be combined with local soil types for further analysis.

[0057]

[0058] Where Area represents the landslide area, Cos represents the cosine function, and n represents the total number of pixels used to identify the landslide. pixel_k Slope represents the area corresponding to a single pixel. k This represents the terrain slope corresponding to a single pixel, and k represents the pixel number.

[0059] This invention is a deep learning-based computer vision method primarily used to identify and extract landslide areas from satellite imagery. This method is applicable to landslide extraction from a wide field of view, exhibiting high accuracy and efficiency.

[0060] The deep learning landslide extraction method based on prior geoscience knowledge can effectively extract landslide areas with an accuracy (mIOU) of 92.74%, and can capture landslides at different scales.

[0061] This invention combines landslide detection results with other parameters such as topography and geological conditions to classify landslides and extract different landslide types. For example, the detection results include slope parameters, which can be used to calculate the landslide area and classify landslides into small, medium, and large types. Combined with geological types, they can be classified into cohesive soil landslides, loess landslides, gravel landslides, and weathered crust landslides.

[0062] This invention incorporates prior geological knowledge about landslide extraction into the model training process, effectively improving the accuracy of landslide extraction. The network structure is improved, simplifying network complexity and enhancing parameter reuse capabilities. Furthermore, the model exhibits strong adaptability, further improving its feature extraction capabilities.

[0063] Example 2

[0064] like Figure 5 As shown, this embodiment provides a landslide extraction system based on geoscientific features. The system includes:

[0065] The data acquisition module 201 is used to acquire remote sensing images, digital elevation models, and terrain parameters of the area to be detected.

[0066] The landslide extraction module 202 is used to input the remote sensing image, the digital elevation model and the terrain parameters into the landslide extraction model and output the landslide extraction result. The landslide extraction model is obtained by training a deep convolutional neural network based on a training set. When performing feature extraction in the deep convolutional neural network, dilated convolution and spatial pyramid pooling modules are used.

[0067] The landslide area determination module 203 is used to determine the landslide area based on the landslide extraction results and the digital elevation model.

[0068] Example 3

[0069] This embodiment provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the landslide extraction method based on geoscientific features according to Embodiment 1.

[0070] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the landslide extraction method based on geoscientific features described in Embodiment 1.

[0071] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0072] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A landslide extraction method based on geoscientific features, characterized in that, include: Collect remote sensing images, digital elevation models, and terrain parameters of the area to be detected; The remote sensing image, the digital elevation model, and the terrain parameters are input into the landslide extraction model, and the landslide extraction result is output. The landslide extraction model is obtained by training a deep convolutional neural network based on a training set. When performing feature extraction in the deep convolutional neural network, dilated convolution and spatial pyramid pooling modules are used. The landslide area is determined based on the landslide extraction results and the digital elevation model. The deep convolutional neural network includes an encoding structure and a decoding structure. The encoding structure adopts an improved deep residual network, wherein the first layer of the deep residual network uses a spatial pyramid pooling module to extract features from the input data, and the third and fourth layers of the deep residual network use dilated convolutions to replace the original pooling layers. The decoding structure is an upsampling process, in which a flow alignment module is used to learn the semantic flow between feature maps of adjacent layers. ； in, Indicates the landslide area. Cos Represents the cosine function. n This indicates the total number of pixels used to identify the landslide. This represents the area corresponding to a single pixel. This indicates the terrain slope corresponding to a single pixel. k Indicates the pixel number.

2. The landslide extraction method based on geoscientific features according to claim 1, characterized in that, The improved deep residual network is the improved ResNet50.

3. The landslide extraction method based on geoscientific features according to claim 1, characterized in that, The terrain parameters are extracted from the digital elevation model, and the terrain parameters include slope and aspect.

4. The landslide extraction method based on geoscientific features according to claim 1, characterized in that, Also includes: Landslides are classified into small, medium, and large landslides based on their area. Small landslides have an area ranging from 0 to 10 square meters. 4 m 2 The area of ​​medium-sized landslides ranges from 10. 4 m 2 ~10 5 m 2 The area of ​​the large landslides ranges from 10 5 m 2 ~10 6 m 2 .

5. The landslide extraction method based on geoscientific features according to claim 1, characterized in that, Each sample data in the training set includes samples from landslide areas and samples from non-landslide areas; The landslide area sample includes input data and label data; the non-landslide area sample includes input data; the input data consists of remote sensing images, digital elevation models, and terrain parameters of the sample area, and the label data consists of landslide edge images of the sample area.

6. A landslide extraction system based on geoscientific features, characterized in that, include: The data acquisition module is used to acquire remote sensing images, digital elevation models, and terrain parameters of the area to be detected. The landslide extraction module is used to input the remote sensing image, the digital elevation model, and the terrain parameters into the landslide extraction model and output the landslide extraction result. The landslide extraction model is obtained by training a deep convolutional neural network based on a training set. When performing feature extraction in the deep convolutional neural network, dilated convolution and spatial pyramid pooling modules are used. A landslide area determination module is used to determine the landslide area based on the landslide extraction results and the digital elevation model. The deep convolutional neural network includes an encoding structure and a decoding structure. The encoding structure adopts an improved deep residual network, wherein the first layer of the deep residual network uses a spatial pyramid pooling module to extract features from the input data, and the third and fourth layers of the deep residual network use dilated convolutions to replace the original pooling layers. The decoding structure is an upsampling process, in which a flow alignment module is used to learn the semantic flow between feature maps of adjacent layers. ； in, Indicates the landslide area. Cos Represents the cosine function. n This indicates the total number of pixels used to identify the landslide. This represents the area corresponding to a single pixel. This indicates the terrain slope corresponding to a single pixel. k Indicates the pixel number.

7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the landslide extraction method based on geoscientific features according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the landslide extraction method based on geoscientific features as described in any one of claims 1 to 5.

Images