Mapping Outburst Floods Using a Collaborative Learning Method Based on Temporally Dense Optical and SAR Data: A Case Study with the Baige Landslide Dam on the Jinsha River, Tibet

Abstract

Outburst floods resulting from giant landslide dams can cause devastating damage to hundreds or thousands of kilometres of a river. Accurate and timely delineation of flood inundated areas is essential for disaster assessment and mitigation. There have been significant advances in flood mapping using remote sensing images in recent years, but little attention has been devoted to outburst flood mapping. The short-duration nature of these events and observation constraints from cloud cover have significantly challenged outburst flood mapping. This study used the outburst flood of the Baige landslide dam on the Jinsha River on 3 November 2018 as an example to propose a new flood mapping method that combines optical images from Sentinel-2, synthetic aperture radar (SAR) images from Sentinel-1 and a Digital Elevation Model (DEM). First, in the cloud-free region, a comparison of four spectral indexes calculated from time series of Sentinel-2 images indicated that the normalized difference vegetation index (NDVI) with the threshold of 0.15 provided the best separation flooded area. Subsequently, in the cloud-covered region, an analysis of dual-polarization RGB false color composites images and backscattering coefficient differences of Sentinel-1 SAR data were found an apparent response to ground roughness’s changes caused by the flood. We carried out the flood range prediction model based on the random forest algorithm. Training samples consisted of 13 feature vectors obtained from the Hue-Saturation-Value color space, backscattering coefficient differences/ratio, DEM data, and a label set from the flood range prepared from Sentinel-2 images. Finally, a field investigation and confusion matrix tested the prediction accuracy of the end-of-flood map. The overall accuracy and Kappa coefficient were 92.3%, 0.89 respectively. The full extent of the outburst floods was successfully obtained within five days of its occurrence. The multi-source data merging framework and the massive sample preparation method with SAR images proposed in this paper, provide a practical demonstration for similar machine learning applications using remote sensing.