A data-driven reconstruction of spatially contiguous daily small catchment runoff for flood and drought monitoring in Switzerland

Runoff observations are critical for the monitoring and understanding of droughts and floods. Traditional methods for estimating runoff rely on physically-based hydrological models, which, while detailed, are often complex and computationally intensive. In contrast, recent advancements in deep learning have shown potential for more efficient and accurate runoff modeling. This study explores the efficacy of temporal neural networks for daily catchment-level runoff reconstruction in Switzerland from 1962 to 2023.

Our model, based on the long short-term memory (LSTM) architecture, is optimized on 87 catchments minimally affected by human activities. It is evaluated in an 8-fold cross validation setup and demonstrates similar performance compared to PREVAH, a distributed hydrological model that is used operationally in Switzerland. Notably, our model requires only precipitation and temperature as meteorological inputs, allowing for an extended reconstruction period back to 1962, unlike PREVAH's 1980 limitation due to its dependency on additional atmospheric forcings. In terms of Kling-Gupta efficiency, our model matches PREVAH's performance, despite its reduced data needs. We evaluate the quality of our reconstruction in terms of extreme events and trends based on the available observations and in comparison to the PREVAH simulations on the national level.

A key advantage of our neural network approach is its computational efficiency, enabling the reconstruction of daily runoff for 307 catchments that cover the entirety of Switzerland in under a minute on a high-performance GPU. This would facilitate real-time droughts and floods monitoring and support environmental scenario simulations. The findings underscore the potential of data-driven models in environmental monitoring and point towards future research in refining these models for broader applications in climate change impact assessments.