GRSG 35th Conference 2024 Abstract

Title: Automated landslide identification using machine learning classification of Airborne LiDAR

Author: Jingru Wang

Organisation: University of Leeds

Airborne LiDAR point cloud data provides detailed, accurate elevation data, even in densely vegetated areas, which is valuable for the identification of landslide morphology. The assessment of landslide morphology is often a manual and time-consuming task, and machine learning algorithms can be used automate landslide detection. This study focuses on developing and applying point cloud filtering and landslide identification algorithms using airborne LiDAR, aimed at improving the precision and efficiency of landslide detection in geospatial analysis.

The proposed workflow starts with point cloud filtering which consists of several steps: preprocessing, filtering of non-ground points, the identification of slope clusters, and region growing segmentation of ground points. The point cloud filtering process uses voxel grid-based down sampling aimed at reducing data size and enhancing data quality. A radius-based outlier removal technique is also applied to identify and eliminate noise points. Following preprocessing, the point cloud is segmented into regular grids, and the lowest point in each grid is extracted.

This grid-based segmentation reduces data redundancy while preserving critical features. Slope consistency is then assessed by fitting a line through the neighborhood points of each target point, determining whether the slope is predominantly upward or downward. Points with positive slopes are grouped into the upward slope category, while those with negative slopes are grouped into the downward slope category. These groups are then subjected to Euclidean clustering, resulting in distinct clusters for both slope directions. This method enhances the identification of consistent slope trends by clustering points with similar slope directions, smoothing local variations, and capturing broader patterns in the data.

This spatial division simplifies subsequent processing steps and retains spatial structure information, facilitating easier analysis and interpretation of the point cloud data. A four-neighborhood region-growing algorithm is then applied to the spatial grid, effectively filtering and segmenting low-lying regions in the 2D plane of the point cloud. The algorithm expands the neighborhood in the horizontal direction, constructing connected low-lying regions while avoiding vertical expansion. Finally, the ground planes and their connected regions within the point cloud are extracted. Initial ground points are identified, and a plane is fitted to these points using the least squares method. The algorithm then evaluates other points in the excluded region by calculating their distance from the fitted plane and expanding the ground point set to include points that lie within a specified threshold distance. This approach effectively classifies ground regions through connected plane identification enabling more accurate terrain analysis and modeling.

The proposed landslide point cloud classification algorithm is based on the Random Forest algorithm. A comprehensive set of features are devised, and the accuracy of landslide identification is assessed. The features include eigenvalue-based metrics such as the sum of eigenvalues, omnivariance, eigenentropy, anisotropy, planarity, linearity, sphericity, and verticality. These metrics capture the geometric characteristics of the point cloud, including shape, distribution, and orientation, providing essential information for distinguishing landslide-related points.

Additionally, density features are computed to assess local point density within a spherical neighborhood, which is crucial for differentiating various terrain morphological features.

Intensity features, derived from the intensity values of the point cloud data, offer additional insights into the reflectance properties of the terrain. Metrics such as mean intensity, intensity deviation, and intensity range are employed to enhance the classification of different land cover types. Moreover, slope features, calculated by determining the gradient of the unit normal vector of the point cloud surface, are critical for identifying terrain variations and detecting landslides.

To maximize classification accuracy, feature selection is conducted using a Random Forest classifier, which evaluates the importance of each feature based on its contribution to the model’s performance. This feature selection processed revealed that omnivariance, verticality, local vertical variance, vertical variance, vertical height, vertical height coefficient, density ratio, and distance between point and plane are most impactful for the detection of landslide morphology. Further testing will be done to better identify landslide morphology and the accuracy will be compared against manual landslide segmentation.