Spatially contiguous reconstruction of water temperature and discharge in Switzerland using deep learning

Water temperature and discharge are critical variables for Switzerland’s ecosystem health and energy production, particularly for nuclear cooling and hydroelectric production. These variables are physically coupled; low-flow conditions often exacerbate high water temperatures, creating "compound events" that can threaten biodiversity and constrain energy availability. 

Pressures on the hydrological system are intensified by anthropogenic climate change, e.g. through rising water temperature and decreasing summer discharge.

 

Switzerland's monitoring network of discharge gauges and water temperature sensors provides valuable insights into past and present conditions, but gaps in spatial coverage and record length still limit robust assessment of country-scale trends.

We develop a joint reconstruction of daily catchment-level water temperature and discharge from 1962 to 2023 using a Long Short-Term Memory (LSTM) network.

Our network is trained on 226 catchments and requires precipitation and air temperature as meteorological inputs, besides static land properties.

We assess the potential of this data-driven reconstruction through an exhaustive spatio-temporal cross-validation and evaluation at different temporal scales.

In addition, we explore the potential of the multi-output architecture to decipher physical coupling of water temperature and discharge and thereby improve representation of compound events.

 

Our network achieves a median Kling-Gupta efficiency (KGE) of 0.83 for water temperature and 0.71 for discharge. The computational efficiency of our model enables an extended reconstruction with spatially contiguous predictions at 1193 locations along the Swiss river network. This first joint reconstruction of water temperature and discharge for Switzerland opens avenues for process understanding and assessments of national trends. The results are promising and highlight potential for refining the model and expanding its applications, such as coupling with climate projections.