FLOOD INUNDATION EXTRACTION BASED ON DECISION-LEVEL DATA FUSION: A CASE IN PERU
Keywords: flood, flood inundation extraction, deep learning, decision-level data fusion
Abstract. Every year, millions of people affected and huge property losses by floods were recorded in many parts of the world. Accurately flood inundated areas extraction is essential for disaster reduction. Existed studies have used multi-spectral (MS) data and synthetic-aperture radar (SAR) data or the fusion data to extract flood inundated areas. However, most data fusion methods think less about regional difference and the complementarities between different models. This study explores a new decision-level data fusion method, which pays more attention to the complementarities between models. First, we construct models trained by diverse bands of Sentinel- 1/2 and water indices. Then, divide the whole study area into three parts, cloud-free & non-water area, cloud-free & flood area and cloud area, and select the models suitable for the three areas. Third, combine water extents extracted by selected models with decision tree to obtain water extents before and after disaster. Finally, subtract the water extent before disaster from the water extent after disaster to get flood inundated areas. The experiments in Peru indicated that our method increases the Intersection over Union (IoU) of water extraction to 0.69. Moreover, our method successfully reduces the impact of cloud and shadow owing to the fusion of different features.