Large-scale flood mapping using SAR remote sensing data

Abstract

Flooding is one of the most serious and costly natural disasters worldwide. In order to monitor and map the global flood-affected regions, remote sensing data over a large scale can be considered. Compared with optical data, synthetic aperture radar (SAR) data is more preferred as it can be obtained regardless of solar illumination and weather conditions. Owing to the growing number of SAR satellite missions, large-scale flood mapping can be conducted with high revisit rates. However, previous flood mapping research focused and tested on specific sites and events while systematic and feasible applications of algorithms at large-scale are still rare.This thesis presents two automatic and reliable methods for large-scale flood mapping, i.e., one focused on rural areas based on historical SAR intensity data while the other focused on urban areas using multi-temporal SAR intensity and interferometric SAR (InSAR) coherence. Furthermore, it introduces a time-series Sentinel-1 based exclusion map (EX-map) indicating the region where SAR data cannot identify floods or non-floods.For rural flood mapping using historical data archives, automatic flood image selection has been proposed by selecting images whose bimodal distributions can always be identified in this specific image and different reference images. Then, a Euclidean distance based reference index is introduced to select an adequate pre-event reference image for change detection based flood mapping. At last, the incidence angle effect, wind effect and water-lookalike vegetation have been considered in the final flood mapping step (i.e., thresholding, region growing and change detection). Experiments covering the south-western United Kingdom with 74 ENVISAT ASAR wide-swath images acquired between 2005 and 2012 demonstrate the effectiveness, robustness and flexibility of the proposed method.The combination of multi-temporal SAR intensity and InSAR coherence in urban flood mapping is explored through U-Net and the proposed urban-aware module. The prior information of urban areas is introduced through urban-aware module to the multi-level features in order to help the model extract informative characteristics of different flood classes (i.e. flooded urban areas and flooded bare soils and sparsely vegetated regions). Experiments conducted on six flood events over three continents based on Sentinel-1 data demonstrate that the proper strategy of using prior information of urban areas can profitably complement intensity and coherence information for large-scale urban flood mapping, especially when it comes to a deep learning model trained via limited labelled training samples.The general assumption in the above-mentioned SAR-based flood mapping is that the appearance of floodwater leads to changes in backscatter. However, this common assumption is not always valid in some specific cases, such as radar shadow/layover regions, urban areas, arid regions and dense forests. Thus, we firstly defined an EX-map which should include all those specific regions in support of largescale flood mapping. A decision tree based method using four multi-temporal indicators is presented. Experiments over six study sites from four continents with totally different land cover scenarios and weather conditions based on 20m Sentinel-1 images from eight orbital tracks illustrate that the definition of the EX-map is more appropriate than the ones of other reference datasets such as global land cover maps and DEM-derived radar shadow/layover mask. Besides complementing flood maps, the proposed EX-map can be used in other applications such as large-scale soil moisture retrieval and assimilation of flood extent maps into hydrological-hydraulic models.