| Résumé : |
(auteur) Floods are a hazard of major concern, causing substantial fatalities and eco-nomic losses. These losses are expected to further accumulate in the future, as both the frequency and magnitude of flood events are projected to increase dueto climate change. Insights into the occurrence and dynamics of these disastrous events are thus of paramount importance for the protection of livelihoods across the world, both in the near and far future.Synthetic Aperture Radar (SAR) satellite imagery is particularly suited to observe floods due to the synoptic view, low cost and timely availability ofsatellite imagery and the all-weather imaging capabilities of SAR sensors. The resulting observations are crucial for various purposes, including emergency relief, post-disaster damage assessment, the calibration and validation of floodprediction models, and risk assessment.Despite the clear advantages of SAR imagery, several factors complicate the flood extent retrieval from this imagery type. These include surfaces or land dynamics characterized by a SAR backscatter similar to that of water/flooding,as well as the presence of urban features and vegetation. Moreover, existing approaches often lack the robustness and automation necessary for operational purposes. This thesis aims to contribute to the accuracy and automation of SAR-based flood mapping approaches, by elaborating on several of theremaining challenges. More specifically, the objectives of this thesis are:
1.to investigate the state of the art in SAR-based flood mapping andidentify the strengths and limitations of existing methods, as well as possible trends;
2.to assess the potential of C-band SAR for the delineation of floodedvegetation, and suggested an approach for doing so in an automated way;
3.to identify the main obstacles with respect to automated flood monitoring,and develop an approach that allows putting science into practice.
In the process of pursuing these objectives, special attention is given to automation, as this is key for objective and timely observations, and to optimally employing available data, as additional data can substantially improve flood observations but not handling these critically may be have adverse effects. Additionally, the potential of object-based image analysis (OBIA) techniques is investigated, as they have proven their added value using optical imagery but SAR-based applications remain limited. Sentinel-1imagery is the main datasource considered in this thesis, as this medium-resolution C-band imagery is freely available and provides consistent global coverage.First, the state of the art in SAR-based flood mapping is investigated. Distin-guishing between approaches for the retrieval of open water, flooded vegetationand urban flooding, deployed input data and classification techniques are discussed. As it is difficult to draw conclusions regarding the strengths and limitations of these classification techniques based on their scientific publications, an in-depth assessment and comparison of a selection of these is carried out. This selection includes thresholding, active contour modeling and theHSBA-Flood method, and both single scene and change detection-based maps are generated. To tackle the second objective of this thesis, the detectability of both woody and herbaceous vegetation using Sentinel-1 is investigated. Moreover, an automated, object-based clustering approach, making use of globally and freely available data only, is presented and applied on four study areas with varying characteristics. The resulting flood maps discriminate between dryland, permanent water, open flooding and flooded vegetation. Forests are indicated too, in order to underline the uncertainty related to these areas where flooding cannot or only to a limited extent be detected.In the last part of this thesis, an approach for operational flood monitoringin Flanders is presented. This approach was developed for and with input of the local water manager,i.e.the Flanders Environment Agency, and makesuse of high-resolution ancillary data available for the region of interest. By combining a pixel-based and an object-based approach, a discrimination is made between dry land, permanent water, open flooding, probable flooding, flooded vegetation and probably flooded forests. The approach is extensively tested on flood events of different sizes that occurred between 2016 and 2020. Both the detectability of these flood events and the accuracy of the developed algorithm, in the presence and absence of flooding, are assessed and discussed. |
Flood mapping from radar remote sensing using automated image classification techniques (2021)
