Multi-Regional landslide detection using combined unsupervised and supervised machine learning

Abstract

Landslide detection is concerned with delineating the extent of landslides. Most of existing works on landslide detection have limited geographical extents. Therefore, the models developed out of these studies might perform poorly when applied to regions with different characteristics. This study investigates an Object-Based Image Analysis methodology built on unsupervised and supervised Machine Learning to detect the location of landslides occurred in multiple regions across the world. The utilized data includes Sentinel-2 multi-spectral satellite imagery and ALOS Digital Elevation Model. In the segmentation stage, pre and post-landslide images undergo segmentation using K-means clustering. Following the segmentation stage and dataset preparation and removing highly-correlated features from the dataset, two Random Forest classifiers (RF1 and RF2) are trained and tested on two different datasets to measure the generalization level of the algorithms with RF1 dataset spanning over more geographical diversities than RF2 dataset. The results show that the RF models can successfully detect landslide segments with test precision = 0.96 and recall = 0.96 for RF1 and test precision = 0.90 and recall = 0.87 for RF2. Further validation shows that, compared to RF2, RF1 results in less mislabelled non-landslide segments.

Keywords

• landslide
• change detection
• machine learning
• obia
• k-means
• random forest