In this Master’s thesis, we modelled the spatio-temporal dynamics of flooding in Lake Cerknica, a highly dynamic karst hydrological system. The analysis encompassed both the modelling of past system states and the development of a machine learning model for short-term forecasting of the future spatial distribution of flooded areas. For the period 2023–2025, we generated a time series of 127 water masks from cloud-free PlanetScope optical satellite imagery. Due to radiometric inconsistencies between images acquired by individual nanosatellites, a simple threshold-based approach using a selected spectral index was not sufficient. We therefore developed a Random Forest model for water surface classification, which achieved F1 = 0.948, IoU = 0.901, and AUC = 0.996 in five-fold cross-validation. To address the limitations of optical imagery in detecting water beneath dense vegetation, we enhanced the classification by integrating topographic data and water level information. As no reliable digital terrain model was available for the lake area, we derived our own digital terrain model (DTM) from the available digital surface model (DSM). Using satellite, topographic, hydrological, meteorological, and vegetation data, we then developed a spatio-temporal LightGBM model to forecast flooding three days ahead. The model, trained on data from 2023–2024 and independently evaluated on data from 2025, achieved AUC = 0.985, F1 = 0.912, and IoU = 0.838, outperformed the persistence model on 33 of 45 test dates, and reduced the mean absolute error in predicted flooded area from 2.41 to 0.99 km2 . SHAP (SHapley Additive exPlanations) analysis showed that predictions were most strongly influenced by topography and the previous state of the system. We also developed an interactive web application that uses up-to-date data to generate three-day-ahead spatial forecasts of flooding in Lake Cerknica.
|
