Abstract

Pluvial (or surface water) floods occur typically during short-term high-intensity rainfall events and are characterized by extreme overland flow of rainwater causing various types of damages to the land surface. Comparably to fluvial floods (i.e. when rivers overflow), pluvial floods have been estimated accountable for around half of all flood damage claims each year. Intense surface runoff can potentially occur anywhere, especially outside the vicinity of watercourses, and over very short time periods. Consequently, observational data associated with pluvial floods are very hard to come by and most sources
cannot be considered exhaustive. This thesis aims at making the most of the growing availability of high-resolution satellite imagery to identify pluvial flood footprints on the ground in the days following a heavy weather event, not flooded water surfaces per se.

To start with, extended geo-referencing and labeling of areas affected by intense overland flow was carried out on three distinct events in the South of France. Then, open-source Sentinel-2 (S-2) optical products were considered for their fine spatial resolution, high worldwide revisit frequency and good spectral range. Change images were produced on each event from the closest cloud-free pre and post event data to determine specific statistical patterns in the temporal evolution of vegetation-/water-based spectral indices within affected areas. Initial works identified optimal combinations of S-2 indicators to successfully implement a transferable object-based Gaussian process classifier called SPCD, for Sentinel
Plot-based Change Detection. Detection rates greater than or equal to 70% and false positives lower than 12% were obtained on all three events using simple VNIR-based spectral indices like the NDVI or NDWI. Then, the resulting impact maps from SPCD were used to thoroughly evaluate the IRIP© susceptibility mapping method by considering rainfall intensity radar measurements of the events. The IRIP model was found very relevant as the proportions of damaged plots increased accordingly with higher susceptibility levels, and even more so when focusing on where precipitations were the heaviest. Structural improvements to IRIP were also suggested in a version called IRIP++. The main IRIP predictors for locating runoff damages turned out to be the topographic wetness index in association with the uphill runoff production susceptibility. Afterwards, a second generalizable method called FuSVIPR, for Fusion of Sentinel-2 and Very high resolution Imagery for Pluvial Runoff, was developed. It enriched the S-2 change images with post event very high spatial resolution optical imagery from Pléiades satellites or airborne sensors to identify land cover changes induced by pluvial floods more precisely and at the pixel level, using Random Forest or U-net algorithms. Submetric impact maps were eventually obtained on the three contrasted validation sites, displaying good accuracy (≥ 75% detection rates) and fewer false positives (≤ 2%) compared to SPCD. An additional recent validation event in South Africa with external ground truths confirmed the accuracy and transferability of the method in a suburban area with similar classification performances (77%). Then, IRIP as well as its suggested enhanced version IRIP++ were further demonstrated relevant, this time based on the FuSVIPR impact maps. In parallel, an original methodology was developed to support the downscaling of extreme rainfall using a stochastic simulator in order to produce fine-scale scenarios that are consistent with the spatial distribution of pluvial flood damages as identified in the FuSVIPR maps.

Cite the thesis

Arnaud Cerbelaud. Pluvial flood detection using satellite remote sensing and machine learning techniques for the evaluation of surface runoff susceptibility mapping. Physics [physics]. ISAE - Institut Supérieur de l'Aéronautique et de l'Espace, 2023. English. ⟨NNT : 2023ESAE0036⟩. ⟨tel-04569356v2⟩

Access manuscript on HAL thèses