Flood Area Prediction in Urban Watershed of Chiang Mai Province Using Remote Sensing and Regression Learning

Abstract

The current state of global warming has led to more frequent and severe flooding. Accurate flood warnings in advance can help mitigate damage to lives and property. This research presents a modeling approach for predicting flood-prone areas using satellite imagery data extraction. The study focuses on urban areas in Chiang Mai Province, covering a total area of 902.025 square meters. Six key factors influencing flood prediction are considered: water balance, evapotranspiration, land cover, rainfall, NDVI, and flood extent. These factors are derived from satellite images captured by Sentinel, MODIS, and Landsat satellites via Google Earth Engine. Satellite imagery provides extensive and continuous spatial coverage, while radar satellite (SAR) data remains operational even in cloudy weather or at night, reducing the cost and time required for field surveys. The flood prediction models were developed using seven regression methods: LR, SVR, Decision Tree, Random Forest, Gradient Boosting, XGBoost, and LightGBM. These methods were selected for their high efficiency and computational speed, making them suitable for large datasets without requiring high processing power. The Random Forest method demonstrated the highest accuracy in flood area prediction, with lower RMSE and higher R-squared values compared to XGBoost, LightGBM, Decision Tree, Gradient Boosting, SVR, and LR, with average improvements of 3%, 12%, 19%, 31%, 51%, and 65%, respectively. This research holds significant potential for practical application, particularly through the development of an early warning system for flood hazards at the local level. Such a system would be especially beneficial in urban areas with high population density, enhancing the efficiency and timeliness of emergency response planning and evacuation efforts by relevant agencies in a systematic and coordinated manner.