Landslide susceptibility mapping using XGBoost machine learning method
Shubham Badola, Varun Narayan Mishra, Surya Parkash
Abstract
The present work demonstrates a landslide hazard zonation technique based on a machine learning algorithm of Extreme Gradient Boosting (XGboost). The XGboost algorithm uses a multi-parameter indices function as an input, namely Slope, Aspect, Curvature, topographic wetness index (TWI), Stream Power Index (SPI), Rainfall and Normalized Difference Vegetation Index (NDVI). These indices were systematically investigated using ALOS PALSAR DEM image (12.5m resolution), rainfall from WorldClim data and Landsat-5 data for NDVI mapping. Initially, the hierarchical selection of the causative factors was carried out within the XGBoost algorithm, which revealed the variable importance of each causative factor to be used for the parameterization of the landslide potential index (LPI). LPI mapping indicated five different classes of landslide zonation, which were tested using the ROC (receiver operating characteristic) curve. The AUC (Area under the curve) yielded a value of 0.91, indicating a robust performance of the ROC in terms of a true positive rate. A significantly larger spatial extent of the study area was found in moderate to high zones along the road, indicating the potential impact of governing causative factors of landslide vulnerability.