Rapid Land Cover Change Detection with Optical and SAR Satellite Data after Tropical Cyclone Seroja - Case Study in Dili, Timor-Leste Pedro Junior Fernandes (a*), Masahiko Nagai (b)
Yamaguchi University
Abstract
This study utilizes optical PlanetScope imagery and Synthetic Aperture Radar (SAR) data from Sentinel-1 to assess land cover changes in Dili, Timor-Leste, following Tropical Cyclone Seroja. The primary aim is to investigate how combining these data types enhances disaster monitoring and response in areas affected by flooding. Using the Random Forest classifier, known for its effectiveness with high-dimensional and noisy datasets, we categorized land cover into six classes: vegetation, water, built-up areas, bare soil, clouds, and shadows. This categorization was conducted across pre-disaster, post-disaster, and recovery phases using Google Earth Engine (GEE). To improve the delineation of water bodies, we applied binary segmentation through Otsu thresholding on the SAR images. The classification achieved impressive accuracy, with overall accuracy scores ranging from 97% to 98.7% and Kappa indices between 0.947 and 0.968, indicating strong model performance. Notably, the study revealed a significant increase in water bodies, considerable damage to vegetation and built-up areas after the disaster, and a gradual recovery over time. However, challenges were encountered in urban classification, especially in distinguishing between built-up and bare soil areas. This research emphasizes the value of integrating optical and SAR data with machine learning techniques for effective land cover monitoring in post-disaster contexts, contributing to enhanced disaster preparedness, improved recovery planning, and vital insights for managing flood impacts in regions like Timor-Leste, where technical resources and data may be limited.
Keywords: land cover change- change detection- disaster monitoring- Random Forest classifier- Otsu Thresholding- SAR data- optical imagery- Timor-Leste
