Monitoring of stormwater basins with Sentinel-2 image time series over the province of Walloon Brabant (Belgium)

Scientific and public attention for flood risk management are constantly growing in the recent years because flash floods have the propensity to affect populations and societies, sometimes leading to dramatic humanitarian crises. It is therefore necessary to ensure an integrated and coordinated management of flooding mitigation measures. However, the inventory and monitoring methods of stormwater basins (SB) that are currently used by the Walloon administration do not provide a comprehensive understanding of the location and the operational state of SB at both regional and provincial levels. Remote sensing constitutes in that regard an unexploited and promising method for acquiring that information at those geographical scales. In fact, the recent launch of ESA’s Sentinel-2 satellites has brought new perspectives for any development of an improved regional and provincial scale SB monitoring. The following research study builds on that renewed perspective while at the same time offers a continuity to the earlier initial project of data centralization launched by the Walloon Brabant Province. The study aimed at assessing the feasibility of monitoring the hydrological functioning of SB using Sentinel-2 images time series over the province of Walloon Brabant. This goal has been met through three steps. The first one consisted in the delineation of SB in order to constitute the SB dataset and the characterization of their potential detectability according to the adequacy between its characteristics and the limitations of Sentinel-2 images. Through the second step, a Random Forest (RF) classification model was developed to detect the presence of water inside the observable SB based on the spectral information provided by three Sentinel-2 calibration/validation images. The last step was dedicated to the application of the validated classification model on the time series of Sentinel-2 images in order to extract the temporal evolution of water inside potentially observable SB. This study demonstrated that a majority (140/156) of SB, initially recorded in the provincial platform of Walloon Brabant, were observable and delineated by means of orthophotoplans. The assessment of SB detectability suggested that a majority (85/140) of SB was potentially observable from Sentinel-2 images. These results have further strengthened the confidence in the fact that the Sentinel-2 MSI sensor seems to be highly adequate for SB monitoring. Furthermore, the implemented methodology makes possible to determine a priori whether Sentinel-2 would be able to observe the SB and thus monitor its water surface. This study concluded that Sentinel-2 image time series provides a cost-free mean for acquiring information about the elementary state and the temporal evolution of water presence in the observable SB over the province of Walloon Brabant. In fact, the successfully developed RF classification model was able to represent the temporal evolution of water with an accuracy of 72%. The 50 cloud-free observations spread over the years 2017 and 2018 provided a competitive observation frequency in comparison with visual inspection techniques. Future research should explore the possibility of integrating Sentinel-1 images in the proposed approach in order to address the cloud obscuration and the artifacts caused by heavy precipitation and wet snow.