An old landslide identification method based on LiDAR multi-dimensional observation features

Abstract

This invention discloses an ancient landslide identification method based on LiDAR multidimensional observation features, comprising the following identification process: acquiring airborne LiDAR multidimensional observation data; acquiring a persistence map and extracting landslide morphological topological features; acquiring another persistence map and extracting landslide texture topological features; extracting ancient landslide spectral features; vectorizing the topological features; constructing a deep learning convolutional network; fusing the vectorized topological features with the orthophoto spectral characteristics of the ancient landslide; and obtaining the ancient landslide image feature classification result based on multi-feature fusion. This invention utilizes persistence homology theory to generate a persistence map, extracts topological features from it, establishes the relationship between topological features and landslide evolution, combines the topological features and spectral features, and dynamically performs nonlinear transformations and combinations of each feature during training using a constructed convolutional neural network to obtain a more comprehensive representation of landslide features and thus obtain the ancient landslide identification result.

Description

Technical Field

[0001] This invention relates to a method for identifying ancient landslides, and more particularly to a method for identifying ancient landslides based on LiDAR multidimensional observation features. Background Technology

[0002] In recent years, due to factors such as extreme weather, earthquakes, and engineering construction, a large number of ancient landslides have reactivated, causing serious losses to people's lives and property. The concealed nature of ancient landslides and the complexity of their reactivation and evolution make identification difficult, hindering definitive decisions and making prevention challenging. Utilizing LiDAR technology to visualize the entire landslide body, characterize its micro-topography, and interpret its geological features provides a new solution for geological hazard identification.

[0003] However, for ancient landslides that have experienced instability and destruction in the past, the spectral characteristics of the landslide boundary are not significantly different from the surrounding environment due to the long time elapsed, resulting in a low accuracy rate in identifying ancient landslides. Currently, existing research based on airborne LiDAR technology mainly utilizes high-precision DEM data and derived data, as well as high-resolution RGB orthophotos, to extract geometric features, textures, and tonal characteristics. These methods essentially only focus on the geometric and statistical features of the landslide, ignoring the global topological features. This makes it difficult to effectively distinguish features in complex mountainous environments, such as those with significant shadow variations, field ridges, and steep walls, which are similar to the landforms of ancient landslides. Summary of the Invention

[0004] To address the shortcomings of the aforementioned technologies, this invention provides a method for identifying ancient landslides based on LiDAR multidimensional observation features.

[0005] To address the above technical problems, the technical solution adopted in this invention is: a method for identifying ancient landslides based on LiDAR multidimensional observation features, comprising the following identification process:

[0006] Step 1: Acquire airborne LiDAR multidimensional observation data;

[0007] Step 2: Construct a grayscale function for the mountain shadow using a DEM to obtain a persistent map and extract the topological features of the landslide morphology;

[0008] Step 3: Use the LiDAR intensity grayscale function to obtain the persistence map and extract the landslide texture topological features;

[0009] Step 4: Extract spectral features of ancient landslides from orthophoto data;

[0010] Step 5: Vectorize the topological features;

[0011] Step 6: Construct a deep learning convolutional network;

[0012] Step 7: Perform data fusion of the vectorized topological features from Step 5 and the spectral features of ancient landslides obtained from the LiDAR orthophoto imagery in Step 4;

[0013] Step 8: Using the deep learning convolutional neural network constructed in Step 6, obtain the feature classification results of ancient landslide images based on multi-feature fusion.

[0014] Furthermore, in step one, high-precision DEM, intensity data, and orthophoto data are acquired using LiDAR technology; and the slope and aspect are calculated using the high-precision DEM.

[0015] Furthermore, in step two, multiple mountain shadow maps are created from different directions to characterize the micro-topography of the landslide. For the shadow images generated by illumination from multiple angles, grayscale functions of the landslide shadows in multiple directions are constructed, and the calculation formula is as follows:

[0016] S ei =sin H ei ·cos H s +cos H ei ·sin H s ·cos(A zi -A s )

[0017] Among them, S ei H represents the grayscale value of the mountain shadow. ei H is the solar altitude angle. s For the slope of the grid cell, A zi Let A be the solar azimuth angle. s The slope aspect of the grid cell.

[0018] Furthermore, in step two, the grayscale function S of the mountain shadow is used. ei (x,y) represents the topological features extracted from the research object; s is set as the grayscale threshold, and the grayscale value function is filtered to record the times when the topological features appear and disappear as the grayscale threshold s increases, thereby obtaining a persistence map and extracting the topological features of the landslide morphology.

[0019] Furthermore, in step two, the solar altitude angle is specifically set at 15°, 30°, 45°, and 60°, and the shadow grayscale function S on azimuth angle i is obtained by transforming the solar azimuth angle. ei (x,y), the shadow grayscale function S at different azimuth angles ei (x,y) is used to extract topological features of the research object, and the features from multiple directions are combined to obtain local contextual information of the morphology. Finally, the topological features of different azimuth angles of 0, 60, 120, 180, 240, 300, and 359 are combined to obtain the topological features of landslide morphology from multiple directions.

[0020] Furthermore, in step three, the topological features of the study object are extracted using the LiDAR intensity gray value function G(x,y); the intensity threshold is set to t, the intensity value function is filtered, and the times when the topological features appear and disappear as the intensity threshold t increases are recorded to obtain the persistence map and the topological features of the landslide texture are obtained.

[0021] Furthermore, in step five, the vectorization process is as follows:

[0022] The topological features are sorted in descending order based on their lifespan. The top N features are selected as fixed-dimensional inputs. The x and y coordinates of each feature are then transformed to obtain a feature vector. The above topological feature persistence graph is then vectorized to obtain topological features that can be input into a deep learning convolutional neural network.

[0023] Furthermore, in step six, a U-shaped encoder-decoder structure is constructed, with the backbone network of the encoding part being Mobilenetv2, and the decoding part utilizing deconvolution and inverted residual structures to obtain rich hierarchical features.

[0024] Furthermore, deep feature extraction is performed based on the deep learning convolutional neural network constructed in step six to obtain a new abstract image feature layer of the ancient landslide. The new feature layer is then input into the fully connected layer and the Softmax layer for prediction and classification, resulting in a feature classification result of the ancient landslide image based on multi-feature fusion.

[0025] This invention discloses an ancient landslide identification method based on LiDAR multidimensional observation features. It utilizes persistent homology theory to analyze mountain shadow and intensity data to generate a persistence map, extracts topological features of ancient landslides from the persistence map, and establishes the relationship between these topological features and landslide evolution. These topological and spectral features are combined, and a constructed convolutional neural network dynamically performs nonlinear transformations and combinations on each feature during training to obtain a more comprehensive representation of landslide features. The representation is then input into fully connected layers and a softmax layer for prediction and classification, yielding the ancient landslide identification result and effectively improving the accuracy of ancient landslide identification. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0027] Figure 2 This is a LiDAR orthophoto image of the present invention.

[0028] Figure 3 This is a LiDAR intensity data diagram of the present invention.

[0029] Figure 4 This is a mountain shadow diagram when the azimuth angle is 300 degrees and the elevation angle is 45 degrees.

[0030] Figure 5 This is a mountain shadow diagram when the azimuth angle is 240 degrees and the elevation angle is 45 degrees. Detailed Implementation

[0031] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0032] This implementation discloses a method for identifying ancient landslides based on LiDAR multidimensional observation features, such as... Figure 1 The overall technical flowchart is shown below, including the following identification process:

[0033] Step 1: Acquire airborne LiDAR multidimensional observation data;

[0034] High-precision DEM, intensity data, and orthophoto data were acquired using LiDAR technology; the slope and aspect were further calculated using the high-precision DEM. The intensity data and orthophoto data are shown below. Figure 3 and Figure 2 As shown.

[0035] Step 2: Construct a grayscale function for the mountain shadow using a DEM to obtain a persistent map and extract the topological features of the landslide morphology;

[0036] Multiple mountain shadow maps are created from different directions to characterize the micro-topography of the landslide. Based on shadow images generated by illumination from multiple angles, grayscale functions of the landslide shadows in multiple directions are constructed, and the calculation formula is as follows:

[0037] S ei =sin H ei ·cos H s +cos H ei ·sin H s ·cos(A zi -A s )

[0038] Among them, S ei H represents the grayscale value of the mountain shadow. ei H is the solar altitude angle. s For the slope of the grid cell, A zi Let A be the solar azimuth angle. s The slope aspect of the grid cell.

[0039] Using the grayscale function S of the mountain shadow ei (x,y) represents the topological features extracted from the research object; the grayscale threshold is set to s, and the grayscale value function is filtered to record the times when the topological features appear and disappear as the grayscale threshold s increases, thereby obtaining the persistence map and extracting the topological features of the landslide morphology.

[0040] Specifically, the solar altitude angle is set at 15°, 30°, 45°, and 60°, and the shadow grayscale function S on azimuth angle i is obtained by transforming the solar azimuth angle. ei (x,y), the shadow grayscale function S at different azimuth angles ei (x,y) is used to extract topological features of the research object, and features from multiple directions are combined to obtain local contextual information of the morphology. Finally, topological features at different azimuth angles of 0, 60, 120, 180, 240, 300, and 359 are combined to obtain multi-directional landslide morphological topological features. For example Figure 4 The image shown is a mountain shadow diagram with an azimuth angle of 300 degrees and an elevation angle of 45 degrees, as processed by this invention. Figure 5 The mountain shadow map processed by this invention has an azimuth angle of 240 degrees and an elevation angle of 45 degrees.

[0041] Step 3: Use the LiDAR intensity grayscale function to obtain the persistence map and extract the landslide texture topological features;

[0042] The topological features of the study object are extracted using the LiDAR intensity gray value function G(x,y). The intensity threshold is set to t. The intensity value function is filtered, and the times when the topological features appear and disappear as the intensity threshold t increases are recorded to obtain the persistence map and the topological features of the landslide texture.

[0043] Step 4: Extract spectral features of ancient landslides from orthophoto data;

[0044] Step 5: Vectorize the topological features;

[0045] The topological features are sorted in descending order based on their lifespan. The top N features are selected as fixed-dimensional inputs. The x and y coordinates of each feature are then transformed to obtain a feature vector. The above topological feature persistence graph is then vectorized to obtain topological features that can be input into a deep learning convolutional neural network.

[0046] Step 6: Construct a deep learning convolutional network;

[0047] A U-shaped encoder-decoder structure is constructed. The backbone network of the encoding part is Mobilenetv2, which obtains 5 effective feature layers in 5 stages. The decoding part uses deconvolution and inverse residual structure to obtain rich hierarchical features.

[0048] Step 7: Perform data fusion of the vectorized topological features from Step 5 and the spectral features of ancient landslides obtained from the LiDAR orthophoto imagery in Step 4;

[0049] Step 8: Using the deep learning convolutional neural network constructed in Step 6, obtain the feature classification results of ancient landslide images based on multi-feature fusion.

[0050] The deep learning convolutional neural network constructed in step six is ​​used to extract deep features and obtain a new abstract image feature layer of the ancient landslide. The input is then used to perform prediction and classification in the fully connected layer and the Softmax layer to obtain the feature classification result of the ancient landslide image based on multi-feature fusion.

[0051] This invention addresses the challenge of distinguishing between features in complex mountainous environments, such as those with significant shadow variations, field ridges, and steep cliffs, which resemble ancient landslide landforms. It proposes using persistent homology theory to analyze mountain shadow and intensity data to generate a persistence map. Topological features of ancient landslides are then extracted from this persistence map. These topological features are robust to noise and deformation, providing a new perspective on landslide characteristics and improving the accuracy of ancient landslide identification.

[0052] The above embodiments are not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the technical solution of the present invention are also within the protection scope of the present invention.

Claims

1. A method for identifying ancient landslides based on LiDAR multidimensional observation features, characterized in that: The identification process includes the following: Step 1: Acquire airborne LiDAR multidimensional observation data; Step 2: Construct a grayscale function for the mountain shadow using a DEM to obtain a persistent map and extract the topological features of the landslide morphology; In step two, multiple mountain shadow maps are created from different directions to depict the micro-topography of the landslide. For the shadow images generated by illumination from multiple angles, a grayscale function for the landslide shadow in multiple directions is constructed, calculated using the following formula: ， in, This represents the grayscale value of the mountain shadow. The solar altitude angle, For the slope of the grid cells, The azimuth of the sun. The slope aspect of the grid cells; using the grayscale function of the mountain shadow. Extract topological features from the research object; set the grayscale threshold as... s Filter the grayscale value function and record the changes with the grayscale threshold. s By increasing the occurrence and disappearance of topological features, a persistence map is obtained, and landslide morphological topological features are extracted. Step 3: Use the LiDAR intensity grayscale function to obtain the persistence map and extract the landslide texture topological features; In step three, the topological features of the research object are extracted using the LiDAR intensity grayscale value function G(x,y); let the intensity threshold be... t Filter the intensity value function and record the changes with the intensity threshold. t By adding the times when topological features appear and disappear, a persistence map is obtained, and landslide texture topological features are acquired. Step 4: Extract spectral features of ancient landslides from orthophoto data; Step 5: Vectorize the topological features; Step 6: Construct a deep learning convolutional network; Step 7: Perform data fusion of the vectorized topological features from Step 5 and the spectral features of ancient landslides obtained from the LiDAR orthophoto imagery in Step 4; Step 8: Using the deep learning convolutional neural network constructed in Step 6, obtain the feature classification results of ancient landslide images based on multi-feature fusion.

2. The method for identifying ancient landslides based on LiDAR multidimensional observation features according to claim 1, characterized in that: In step one, LiDAR technology is used to acquire high-precision DEM, intensity data, and orthophoto data; and the slope and aspect are calculated using the high-precision DEM.

3. The method for identifying ancient landslides based on LiDAR multidimensional observation features according to claim 1, characterized in that: In step two, the solar altitude angle is specifically set at 15°, 30°, 45°, and 60°, and the azimuth angle is obtained by changing the solar azimuth angle. The grayscale function of the shadow on Shadow grayscale function at different azimuth angles Topological features were extracted for the research object, and features from multiple directions were combined to obtain local contextual information of the morphology. Finally, topological features from different azimuth angles of 0°, 60°, 120°, 180°, 240°, 300°, and 359° were combined to obtain multi-directional landslide morphological topological features.

4. The method for identifying ancient landslides based on LiDAR multidimensional observation features according to claim 1, characterized in that: In step five, the vectorization process is as follows: The topological features are sorted in descending order based on their lifespan. The top N features are selected as fixed-dimensional inputs. The x and y coordinates of each feature are then transformed to obtain a feature vector. The above topological feature persistence graph is then vectorized to obtain topological features that can be input into a deep learning convolutional neural network.

5. The method for identifying ancient landslides based on LiDAR multidimensional observation features according to claim 1, characterized in that: In step six, a U-shaped encoder-decoder structure is constructed. The backbone network of the encoding part is Mobilenetv2, and the decoding part uses deconvolution and inverted residual structure to obtain rich hierarchical features.

6. The method for identifying ancient landslides based on LiDAR multidimensional observation features according to claim 1, characterized in that: Based on the deep learning convolutional neural network constructed in step six, deep feature extraction is performed to obtain a new abstract image feature layer of the ancient landslide. The input is then used to perform prediction and classification in a fully connected layer and a softmax layer to obtain the image feature classification result of the ancient landslide based on multi-feature fusion.

Images