Transfer Learning in Landslide Susceptibility Mapping: Bridging Data-Rich and Data-Scarce Regions in the Northwestern Himalayas

Landslides are increasing in mountainous regions of the world, causing significant threats to human lives and the economy. Landslide Susceptibility Mapping (LSM) is vital for delineating areas prone to future landslides. However, the LSM preparation often encounters limited data availability, causing a major obstacle in understanding the occurrence of landslides and the environmental factors responsible. This article attempts to bridge the gap of limited data availability using a transfer learning approach wherein a random forest approach is applied. The knowledge obtained from the data-rich region is applied in the scarce region. The results depicted that factors such as distance to road, distance to river and lineament were among major contributing factors. Furthermore, the value of statistical measures such as AUC-ROC, precision, and F-score increased significantly for the LSM when trained from source data. Overall, the study successfully addresses the role of transfer learning approach in LSM preparation, especially in data-scarce regions.