A new radar-based land-cover classification algorithm for accurate flood risk analysis

Abstract:

During a flooding event, the ability of the terrain to dissipate water flow energy depends on its land-cover type and the associated surface roughness. NOAA's Coastal Change Analysis Program (C-CAP) land-cover data are currently used in operational storm surge models for estimating surface roughness maps. C-CAP data were derived from optical imagery, and a new version is released every 5-6 years. Here we show that modern radar satellites provide a way to retrieve surface roughness maps with improved spatial and temporal coverage for improving the accuracy of storm surge modeling.

We developed a supervised land-cover classification algorithm and classified the land surface into 9 classes with distinct surface roughness using Polarimetric Synthetic Aperture Radar (PolSAR) and Interferometric Synthetic Aperture Radar (InSAR) measurements. We tested our algorithm using L-band ALOS data and produced a land-cover map over the Houston area. Our results show strong correlation (R=0.90 with p-value << 0.01) with those derived from NOAA’s C-CAP 2010 land-cover data. Our algorithm does not require an extensive human involvement in training sample selection, and less than 0.3% of radar pixels were used as training data in the test case. In addition, we employed a hierarchical classification strategy, which improves the classification accuracy by splitting a large multi-class classification problem into subsets that are easier to solve.

This new method can be applied to Sentinel-1 and the upcoming NISAR data for annual or bi-annual surface roughness retrievals. The radar-derived surface roughness can be integrated into storm surge models to improve the accuracy of flood risk analysis and the preparedness in future disasters.