Regionally Adapted Sentinel-1 Flood Detection in Northern Bangladesh For Index Insurance Applications

Abstract:

New types of financial instruments, such as index insurance, are being piloted in flood-prone Bangladesh. Current remote-sensing flood index insurance contracts in Bangladesh rely on optical satellite sensors (MODIS) for flood detection. Radar-based satellites can penetrate clouds, offering more consistent mapping during flood events by detecting changes in backscatter. Yet identifying meaningful backscatter deviations presents challenges in Bangladesh due to the presence and regional variation of widespread seasonal flooding, irrigation inundation, and permanent water. For example, an agriculturally “damaging” flood in Sylhet may be a pre-monsoon “flash” flood while in Jamalpur it may take the form of long-term monsoon-season inundation. Here we present a method to estimate inundated area from the Sentinel-1 Synthetic Aperture Radar (SAR) (every 6-12 days at 10m resolution since 2014) for different regions in Northern Bangladesh.

A Synthetic Aperture Radar time series is developed in Google Earth Engine from the Devries et al. (2020) Sentinel-1 flood mapping algorithm detecting statistical deviation of a target image from a baseline stack of dry-period images. The algorithm is adapted to produce a flooded area time series, then post-processed for increased accuracy. Due to regional variations in wintertime rice irrigation inundation, a dry-period baseline stack is derived regionally from the maximum 1.5 months of the inverse of soil moisture (NASA SMAP).

We compare the flooded area time series to localized river water level data to assess consistency in the signal. Preliminary results comparing a Sentinel-1 flood map to Sentinel-2 in April 2017 Sylhet floods indicate high accuracy (F1 = 0.92). We will present accuracy metrics for three additional large floods in Northern Bangladesh (August 2017, July 2019, and July 2020). Comparisons to river water level data and preliminary event imagery statistics reveal that despite limitations in observation frequency, Sentinel-1 provides an inundation time series with a consistent flood signal and accurate imaging, and is thus suitable for index insurance triggers. However, radar algorithms must be adapted regionally to ensure inundation dynamics are consistently represented for index insurance applications.