Detection of Flooded Areas in a Time Series of High Resolution Synthetic Aperture Radar Images using Curvelet Transform and Unsupervised Classification.

Due to their weather and illumination independence and due to their large area coverage at high spatial resolution, Synthetic Aperture Radar (SAR) images have been recognized as a valuable data source for the mapping and tracking of flooding events. In order to optimize mapping accuracy, a SAR-based flood detection approach should provide the following features: (1) due to the recent increase of flood situations and the rising number of high-resolution SAR datasets that require rapid processing, the generation of flood maps should be accomplished automatically and with robust results; (2) due to the large spatial extent of many recent flooding events, a flood detection approach should be computationally efficient and transferable to different environmental conditions (surface cover; scene heterogeneity) and data types. Considering these requirements and to improve upon the current state-of-the-art, we propose an unsupervised flood detection approach that is based on a time-series analysis of multi-temporal SAR images. Due to the inclusion of geometric and radiometric terrain correction, our method automatically adapts to changing environmental parameters. To preserve the full resolution of the original SAR data while reducing the influence of Speckle on flood detection performance, we integrate several recent image processing developments in our approach: We use curvelet filtering methods to suppress noise while preserving most relevant image details. Similarly, we perform a multi-scale decomposition of the input images to generate image instances with varying resolutions and signal-to-noise ratios. Finally, to increase the robustness to false alarms, we incorporate time-series analysis techniques in our workflow: We first generate a redundant set of change probability maps from a stack of repeated SAR images and then invert this redundant set using least-square techniques to arrive at the most likely time series of surface change. To demonstrate the performance of our approach we processed a series of seven TerraSAR-X data covering the Sagavanirktok River flooding in northern Alaska in Spring 2015. Using our method we mapped both the extent and progression of this event. Through the combination of several parallel tracks, an image time series with a three-day average sampling rate could be achieved.