Computing flood frequency and Duration from Earth Observation data

Kurzfassung

Among different natural disasters, flood occurs more often and is considered as one of the deadliest disasters. Vital information of the flood such as flood frequency and duration of inundation provides useful insights for assessing the vulnerability and further helps in assisting the concerned authorities for planning and management. Therefore, this study aims to develop an optimized automatic algorithm to compute flood frequency and duration in the user defined area. It further investigates the quality of the computed durations for which three different quality layers are generated. This study also explores ways to reduce the ambiguity through combining different satellite products which increase the revisit frequencies and the data availability. Two different sites in India namely Thanjavur and Bihar are studied for 2017. Developed automatic methodology in Python computes the output on pixel level using either flood or water masks. These masks are generated from automatic flood services of German Aerospace Center (DLR) by processing different satellite data. In Thanjavur computations are based on Sentinel-1 and TerraSAR-X flood masks whereas in Bihar water masks from Sentinel-1 and Sentinel-2 are used. The computation involves three major processes: data preparation, multi-processing and output. During data preparation, the masks are clipped by user defined shape file then it is merged if multiple tiles are covering the area and then all the masks are upscaled to the finest available spatial resolution. Frequency and relative frequency layers are computed using these resampled masks; it depicts the frequency with which the pixels are flooded. Similarly, duration and quality layers are computed using parallel processing by creating HDF5 database from the resampled masks. Durations are computed from the database in two different ways: backward duration from latest acquisition date and total inundation period for a certain time period. The satellite acquisitions are limited with its temporal resolution due to unavailability of the daily acquisitions. Thus, there is the possibility of ambiguity in computed durations. These uncertainties are represented in different quality layers where quality layer A indicates the uncertainty of exact end date of flood, quality layer B explains the uncertainty of the start of flood and quality layer C describes the reliability of the computed duration. All three quality layers are computed for total duration whereas backward duration consists of only two quality layers (B and C). Therefore, for each run, nine raster layers are generated which contribute to broader understanding of flood information in the study sites. In 2017, more than 70% and 40 % of flooded area in Thanjavur and Bihar are inundated up to 10 days. In both the study sites, most of the areas have the uncertainty of start and end of flood of less than 20 days and have the relative frequency of less than 20%. However, the results indicate that Bihar is more frequently flooded and is more prone to flood than Thanjavur. Additionally, the reliability of computed duration and frequency is increased by combining different satellite products. Furthermore, interpretation of DEM shows that most of the inundated area and areas with longer duration are at lower elevation. Various factors impact directly on the computational time, but this study optimizes the computation by using three parallel processes and Cython compiler. Since, the masks from DLR are of high accuracies, thus the output of this study is very reliable. In future, more satellite products can be combined to increase the frequency of acquisitions.