An automated approach to flood detection through a threshold based method and machine learning using Sentinel-1A/B SAR data
Abstract
Flooding is one of the most prevalent natural disasters in the world, causing many fatalities and high economic loss. Thus, there is a need to quantify flooding both rapidly and accurately in ways which can aid in disaster management and decision making. This includes quantifying the spatial extent of the flooding through time, as well as the depth of the water. The method used here implements SAR data provided by ESA Sentinel-1A/B which provides a large spatial extent ~100s of kms and 6 to 12 day repeat period which is ideal for map flooding. We use both SAR amplitude and coherence products to derive binary flooding maps as well as combining these products to improve accuracy. Each SAR image is split into n tiles which will be analyzed separately because the characteristics of a SAR scene vary across the swath. Each tile is split into a grid of s by s pixels (box) which are then each analyzed to check for a bimodal distribution using the maximum normalized between-class variance (BCV) threshold (Cao et al. 2019, Demirkaya et al. 2004). The s is determined by iterating through several values and determining which value yields the maximum number of bimodal distributed boxes within a tile. After s is determined, the boxes which were determined to have a bimodal distribution are used to automatically determine the optimal threshold for the tile, using either the mode of the distributions or a local minimum algorithm. This process is done for each tile and combining them produces a binary water coverage map. This map also includes permanent water bodies which need to be removed. To remove the permanent water bodies, a set of images through time are used and common classifications among images are removed assuming sufficient temporal coverage. The binary classification map with common classifications removed is the final flood map which can further be used with a high-resolution Digital Elevation Model (DEM) to produce a flood depth map. Using the same tiling methodology, we can also apply machine learning algorithms to the tiles to detect changes within a tile which are induced by flooding events. The test regions used in this study include Gujarat, Congo, Texas and Michigan.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMNH026..05W
- Keywords:
-
- 3360 Remote sensing;
- ATMOSPHERIC PROCESSES;
- 0240 Public health;
- GEOHEALTH;
- 4328 Risk;
- NATURAL HAZARDS;
- 4332 Disaster resilience;
- NATURAL HAZARDS