Automatic Inundation Warning System Utilizing High-Precision Water Segmentation Deep Learning Model
Abstract
Inundation monitoring plays an important role in early warning and response. An automatic inundation mapping requires highly precise water segmentation results for collating with the public DB. Recently, research on water segmentation using deep learning shows significant performance improvements. Synthetic Aperture Radar (SAR) is widely used for water segmentation, as SAR is immune to the weather and time conditions of the flood. However, the main disadvantage of SAR is that speckle noises and water-like land cover (e.g. airport runway, golf course) make false positive errors, a decisive factor in impairing inundation mapping accuracy. Requiring post-processing is a major problem in processing an automatic inundation warning system, this paper attempts to develop a water segmentation model utilizing geospatial data without post-processing but with reduced false positive error. For constructing abundant and high-quality training data, we exploit a water layer of a landcover map produced by Korea's Ministry of Environment. As river flow changes over time, the time difference between SAR images and aerial orthoimages is limited to three days. Geospatial data of digital elevation map (DEM), slope, aspect, profile curvature, terrain ruggedness index (TRI), distance from the river, and topographic wetness index (TWI) are stacked with Sentinel-1 SAR VV images. We use the deeplabv3+ model for water segmentation and four layers (SAR VV and three geospatial layers) combinations were tested. Collating the segmented water image with the river boundary of the public database, we map the inundation area, limiting the minimum size of the inundation area considering the resolution of Sentinel-1. SAR VV image with TRI, distance from the river, and TWI layers get the highest f1score 96.63%, which is 4.1% higher than only SAR VV image used. As an actual case, water-lookalike areas like flat schoolyards are removed for the case of the flood occurred in August 2020 at Gyeongsangnam-do, Korea. In this paper, we improved the accuracy of water segmentation model without post-processing. Our model proves the potential of the automatic inundation warning system with Sentinel-1 images. On the basis of the future ultra-small satellite constellation, our work contributes to the automatic inundation warning system for real-time.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMNH45B0450K