High-Resolution Flood Monitoring Based on Advanced Statistical Modeling of Sentinel-1 Multi-Temporal Stacks

Abstract

High-resolution flood monitoring can be achieved relying on multi-temporal analysis of remote sensing SAR data, through the implementation of semi-automated systems. Exploiting a Bayesian inference framework, conditioned probabilities can be estimated for the presence of floodwater at each image location and each acquisition date. We developed a procedure for efficient monitoring of floodwaters from SAR data cubes, which adopts a statistical modelling framework for SAR backscatter time series over normally unflooded areas based on Gaussian processes (GPs), in order to highlight flood events as outliers, causing abrupt variations in the trends. We found that non-parametric time series modelling improves the performances of Bayesian probabilistic inference with respect to state-of-the-art methodologies using, e.g., parametric fits based on periodic functions, by both reducing false detections and increasing true positives. Our approach also exploits ancillary data derived from a digital elevation model, including slopes, normalized heights above nearest drainage (HAND), and SAR imaging parameters such as shadow and layover conditions. It is here tested over an area that includes the so-called Metaponto Coastal Plain (MCP), in the Basilicata region (southern Italy), which is recurrently subject to floods. We illustrate the ability of our system to detect known (although not ground-truthed) and smaller, undocumented inundation events over large areas, and propose some consideration about its prospective use for contexts affected by similar events, over various land cover scenarios and climatic settings.