Assessing hydrologic prediction uncertainty resulting from soft land cover classification

For predictions in ungauged basins (PUB), environmental data is generally not available and needs to be inferred by indirect means. Existing technologies such as remote sensing are valuable tools for estimating the lacking data, as these technologies become more widely available and have a high areal coverage. However, indirect estimates of the environmental characteristics are prone to uncertainty. Hence, an improved understanding of the quality of the estimates and the development of methods for dealing with their associated uncertainty are essential to evolve towards accurate PUB. In this study, the impact of the uncertainty associated with the classification of land cover based on multi-temporal SPOT imagery, resulting from the use of the Random Forests classifier, on the predictions of the hydrologic model TOPLATS is investigated through a Monte Carlo simulation. The results show that the predictions of evapotranspiration, runoff and baseflow are hardly affected by the classification uncertainty when area-averaged predictions are intended, implying that uncertainty propagation is only advisable in case a spatial distribution of the predictions is relevant for decision making or is coupled to other spatially distributed models. Based on the resulting uncertainty map, guidelines for additional data collection are formulated in order to reduce the uncertainty for future model applications. Because a Monte Carlo-based uncertainty analysis is computationally very demanding, especially when complex models are involved, we developed a fast indicative uncertainty assessment method that allows for generating proxies of the Monte Carlo-based result in terms of the mean prediction and its associated uncertainty based on a single model evaluation. These proxies are shown to perform well and provide a good indication of the impact of classification uncertainty on the prediction result.