Predicting Landslide Risk in Post-Fire Zones of Northern Thailand Oumkrue, S.(ab*), Rojanavasu, P.(a), Chaiwongsai, J.(a), Kantawong, K.(a), Rachata, N.(a), Deeprasertkul, P.(b)
a)School of Information and Communication Technology, University of Phayao, Thailand
b)Hydro Informatics Institute (Public Organization) (HII), Thailand
Abstract
Landslides are frequently found in Northern Thailand and especially in the rainy season that follows periods of forest fires. Wildfires, generally occurred from January to May, and they damaged the foliage on mountainous terrains. The wildfire damage weakens the soil cohesion and increases the risk of landslides during heavy or prolonged rainfall. This study aims to develop a spatial prediction model for post-wildfire landslide risk using machine learning techniques. It integrates three years of geosptial data which are daily rainfall from real-time telemetry station satellite-derived thermal hotspots historical landslide occurrences slope gradients from DEMs and land-use land-cover data. First in data preparation we use cumulative rainfall over 1-day 3-day and 7-day periods, hotspot counts within 1-year and 2-year windows, slope steepness, and land-use types. Next, we classify landslide risk into 3 levels as follows low, medium, and high, which are defined by analyzing rainfall intensity, hotspot intensity, and the density of historical landslides. The proposed system uses classification methods and a GIS map tool to generate maps that show landslide risk levels. The model performs with high classification accuracy and can identify multiple landslide risk zones. Furthermore, the analysis results suggest that areas previously affected by wildfires tend to show higher landslide exposure under equivalent rainfall conditions compared to areas without wildfire history. The result also shows rainfall accumulation thresholds that are associated with increased landslide risk in post-fire environments. Lastly we found that integrating rainfall, hotspot density, slope gradient and land use can raise the accuracy of predicting landslide risk zones and help create a landslide risk map. In conclusion, this study presents an integration of wildfire history and multi-temporal rainfall data with geospatial features and uses machine learning to enhance landslide risk assessment.
