Machine learning and clustering for supporting the identification of active landslides at national scale

Abstract

Landslides are one of the major geohazards causing significant economic damage and loss of life, with impacts expected to increase due to climate change. Landslide inventory maps (LIMs) are essential for risk management and reduction, but they usually remain an overlooked issue, especially over very large areas, i.e. at a regional or national level. Nowadays, extensive interferometric satellite radar data with wide area coverage are profitably available, but their potential can be not fully exploited due to the challenge of managing them. In this context, we used space-borne advanced Interferometric Synthetic Aperture Radar (InSAR) data at a national scale, in order to create a useful database of active slope instability movement areas to rely on where the landslide inventory map is missing or largely incomplete. Specifically, we provide insights into a new approach, proposing a national-scale method that combines Machine Learning (ML) and clustering tools, which are crucial to manage a huge amount of data. The proposed methodology has been applied to Great Britain. The use of a ML algorithm, specifically Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) for noise filtering, has allowed the InSAR dataset to be reduced from approximately 6.5 million points to about 3.8 million points per component. Thus, implementing ML along with Slope Units for geomorphological reliability, and tools for identifying and classifying active deformation areas yields an InSAR landslide inventory map. Through this process, 336,557 Slope Units have been classified; of these, 5% show discrepancies between landslide inventory and InSAR data. Identifying these areas, along with those classified as landslides by both datasets, is crucial for risk management as it highlights areas that require closer inspections.