Abstract
The estimation of hydrological components on a spatiotemporal scale poses a challenge for researchers as they develop data-driven tools that can be transferred to different regions with varying characteristics. In this study, we propose a hybrid architecture of a surrogate deep learning (DL) model based on the data obtained by the Wflow estimation. The choice of the target region is based on extensive lowlands and large variations in elevation, which makes it more challenging to improve the model’s accuracy. General tasks are addressed in this paper related to geodata frame structuring and preprocessing, DL model enhancement, and multiscale evaluation. Our contribution focuses on proposing a novel combination of LSTM, MLP, CNN, and CVAE to achieve robust outcomes on a finer scale, followed by an integration of FCM clustering for a comparative evaluation of the models’ performance. Training both LSTM and MLP using climate data and geophysical information of the catchment provides a performance comparable to the Wflow benchmark. In addition, thanks to the FCM clustering which classifies the basin in homogenous subregions, we can gain extra insights as to how the models perform in the region as a whole. Our findings show that MLP performs very well in the first subregion, whereas LSTM outperforms the rest of the catchment. The integration of spatial information provided from both models via CNN_Hyb and CVAE_Hyb significantly enhances the overall spatiotemporal prediction across the entire region. However, clustering-based evaluation reveals the influence of LSTM and MLP on CNN accuracy, indicating persistent biases in the third subregion. In contrast, the utilization of CVAE_Hyb effectively mitigates this bias, resulting in a performance increase from 0.85 to 0.93.