Flood Event Detection and Assessment using Sentinel-1 SAR-C Time Series and Machine Learning Classifiers Impacted on Agricultural Area, Northeastern, Thailand

Abstract

This study presents image classification techniques using Sentinel-1A microwave SAR-C imagery to detect agricultural vulnerability area resulting from a massive flood in Ubon Ratchathani province, Thailand, which occurred in 2019. Two time series of selected images were used in analytical processes: namely S1A_IW_GRDH acquired on August 10th, 2019, representing the pre-flood event, and S1A_IW_GRDH acquired on 9th September 2019 represents the massive flood in this area. Prior to the classification, these data were preformed pre-processing processes, such as calibration, speckle filtering and terrain correction. The preprocessed data were then classified using 3 machine learning classifier algorithms, namely, Random Forest (RF), K-Dimensional Tree (KDTree KNN), and Maximum Likelihood for comparing classification accuracy derived from each classifier. There are 4 land use/land cover (LULC) classes derived from the dataset, i.e., (1) paddy rice, (2) water body, (3) residential area, and (4) vegetation, respectively. The second map was used to determine the extent of flooding and non-water area based on backscattering coefficient derived from Sigma0_VV polarization using band math calculation obtained from the histogram. The extracted flooded area aimed at creating the flooded water mask for overlaying with the classified LULC maps derived from each classifier. Finally, the LULC maps were overlaid with flooded event map that occurred on September 9, 2019, for quantifying affected area. The results indicated that paddy rice was damaged by flooded with the area of 98 km2 classified by RF achieving the overall accuracy of 94.60%. The KDTree KNN classifier identified the affected area of 85 km2 with the overall accuracy of 93%, while the Maximum Likelihood classifier detected the flooded area of 91 km2 with the overall accuracy of 93.36%, respectively.