USING NDVI DIFFERENCE AS A LANDSLIDE SUSCEPTIBILITY MAPPING SUPPLEMENT A case study on the Western Hills area near urban Beijing, China

Abstract

As the field of landslide susceptibility mapping gradually moves away from methodology based in field work to that of remote sensing, the temporal aspect of slope stability has become relatively neglected. This is in part caused by a lack of easily accessible datasets from which land use and vegetation change can be derived; to address this, this study provides an initial cost-free, publicly available, and easy to utilize method for including land use and vegetation change in machine learning based landslide susceptibility mapping. This study uses some of the most commonly employed landslide susceptibility mapping conditioning factors in elevation, slope, and aspect, together with normalized difference vegetation index (NDVI) to train machine learning models using the random forest classifier. A base model is compared to three additional models that incorporate an additional, new, conditioning factor in “NDVI difference over time” (NDVIΔ). This new conditioning factor is created by calculating the difference in NDVI value from the study year as well as the NDVI value 1 year (NDVIΔ1), 3 years (NDVIΔ3), and 5 years (NDVIΔ5) prior. To analyze the effectiveness of the addition of the NDVIΔ conditioning factor, this study conducts a landslide susceptibility mapping case study on the Western Hills area of Beijing to look at (i) the benefits, drawbacks, and impact of using the NDVIΔ conditioning factors when creating landslide susceptibility maps, as well as (ii) analyzing the impact the NDVIΔ conditioning factors have on the machine learning model performances. The resulting landslide susceptibility maps show only minor visual differences as a result of the addition of NDVIΔ conditioning factors, and model evaluation establishes the four models created for this study perform at relatively comparable levels. The model incorporating the NDVIΔ1 conditioning factor is the best performing model in the study, with an AUC-score (0.72 ± 0.01) 2.86% higher than the base model (0.70 ± 0.01), indicating a small improvement to landslide prediction strength. In its current state, the NDVIΔ conditioning factor does not effectively represent the temporal dimension of land use and vegetation change, however, the addition of the NDVIΔ conditioning factor has few to no drawbacks while showing some benefits, indicating possible growth and future method development.