Dynamic Muddy Flood Susceptibility Mapping in Wallonia

Details

Supervisors

Bielders, Charles ; Defourny, Pierre

Faculty

Faculté des bioingénieurs

Degree label

Master [120] en sciences agronomiques et industries du vivant, à finalité spécialisée: Geo-Information Science and Earth Observation for Environmental Modelling and Management

Abstract

Muddy floods are sediment-laden runoff events generated when rainfall interacts with vulnerable agricultural surfaces and transports eroded soil towards downstream infrastructure. In Wallonia, their occurrence is influenced by rainfall erosivity, land-cover conditions, erosion intensity, topography, and sediment connectivity, while available inventories are incomplete because reported events provide confirmed cases but unreported locations cannot be assumed to represent true absences. This thesis develops a regional, dynamic framework for prioritising muddy-flood risk at exposed infrastructure locations by combining physically informed environmental predictors with Positive-Unlabelled Random Forest (PU-RF) modelling. The framework integrates Sentinel-2-derived vegetation dynamics, daily rainfall erosivity, RUSLE-based erosion indicators, WaTEM/SEDEM sediment-output information, terrain and hydrological-connectivity variables, and representative watersheds delineated upstream of PICC-flowaxis pour points. Several model families were evaluated using a watershed-date case/background dataset and leave-one-date-out cross-validation, with additional spatial validation used to assess geographic transferability. The final rainfall-constrained PU-RF model identified mean rainfall erosivity, the 95th-percentile RUSLE erosion rate, and the 95th-percentile cover-management factor as the strongest predictor combination. Conventional discrimination performance was modest, with a ROC-AUC of 0.620 and PR-AUC of 0.024. However, the model demonstrated useful prioritisation skill: 15 of 36 reported muddy-flood cases were captured within the top 10% of ranked watersheds per event date, corresponding to 41.7% of cases and more than four times the capture expected from random ranking. Expanding the priority area to the top 20% captured only two additional cases. Spatial validation reduced top-10% capture to 25.0% under nested spatial-within-date validation and 16.7% under pure spatial block validation, indicating limited transferability to geographically unseen areas. The results demonstrate the potential of combining dynamic environmental indicators and Positive-Unlabelled learning for regional muddy-flood prioritisation. The resulting model is best interpreted as a relative, event-specific risk-ranking framework rather than a calibrated probability or deterministic prediction of muddy-flood occurrence.