Automated Extraction of Inundated Areas from Multi-Temporal Dual-Polarization RADARSAT-2 Images of the 2011 Central Thailand Flood

Abstract

This study examines a novel extraction method for SAR imagery data of widespread flooding, particularly in the Chao Phraya river basin of central Thailand, where flooding occurs almost every year. Because the 2011 flood was among the largest events and of a long duration, a large number of satellites observed it, and imagery data are available. At that time, RADARSAT-2 data were mainly used to extract the affected areas by the Thai government, whereas ThaiChote-1 imagery data were also used as optical supporting data. In this study, the same data were also employed in a somewhat different and more detailed manner. Multi-temporal dual-polarized RADARSAT-2 images were used to classify water areas using a clustering-based thresholding technique, neighboring valley-emphasis, to establish an automated extraction system. The novel technique has been proposed to improve classification speed and efficiency. This technique selects specific water references throughout the study area to estimate local threshold values and then averages them by an area weight to obtain the threshold value for the entire area. The extracted results were validated using high-resolution optical images from the GeoEye-1 and ThaiChote-1 satellites and water elevation data from gaging stations.

Fast and Automatic Data-Driven Thresholding for Inundation Mapping with Sentinel-2 Data

“…The vast majority of local automatic thresholding approaches have been developed for utilizing radar images [22,[26][27][28][29], while the proposed approach uses as input optical satellite images to detect water extent. The aim of this paper is to introduce an unsupervised approach capable of generating inundation maps relying on preprocessed S2 data without the need for simultaneous ground truth data.…”

“…This method has been used in multiple approaches in the literature aiming at separating pixels into inundated and non-inundated classes [10,[22][23][24]42]. In this experiment, MCET is replaced by the Otsu method, while the rest of the steps of the unsupervised approach are kept the same.…”

“…The proposed approach efficiently exploits optical data, e.g., Sentinel-2 bands and extracted indices, while the vast majority of local automatic thresholding approaches are developed for utilizing radar data [22,24,[26][27][28][29]. Moreover, it estimates local thresholds relying on patches of expanding size, contrary to most of the approaches that use image subsets of a fixed size [24,[26][27][28].…”

“…Several of the methods rely on image subsets of fixed size [24,[26][27][28] to estimate local thresholds, while there is a limited number of recent approaches that adapt the size of the subregions to the information content retrieved from the image [22,29]. The approach in Chini et al [29] estimates subregions of variable size, which allow the parameterization of the distribution of water and non-water classes, while the approach in Nakmuenwai et al [22] selects subregions of elliptical shape having a water and non-water cover ratio of nearly 1:1.…”

“…Various algorithms have been used in the literature for estimating thresholds partitioning water and non-water pixels inside an image subset based on their class distribution. Most prevalent algorithms include Otsu's algorithm [21] used in [10,13,[22][23][24] and Kittler and Illingworth's algorithm [25], which was used in [26][27][28]. Automatic thresholding approaches are distinguished into global and local thresholding approaches:…”