Which strategy to improve the performances of an LSTM-based model for extreme stream temperature values?

Abstract. Deep-learning models have demonstrated strong performances in reproducing stream temperature dynamics, which is promising for the reconstruction of missing stream temperature records at ungauged locations. However, model accuracy for the range of high, summer stream temperature has been usually overlooked, raising the question of the suitability of using deep-learning methods during this crucial season. In this study, we investigated strategies to improve the performances of a stream-temperature model based on LSTM (Long Short-Term Memory) cells over the highest 10 % observed values at 21 stations located in the Garonne river catchment. We quantified the gain in model performance thanks to regional multi-catchment training with static attributes, exploiting hydrologically relevant variables, and further penalizing the errors at extreme temperature values using custom loss functions. Our key results are: (1) Regional multi-catchment training is the best strategy to improve the performances of LSTM models not only over the top 10 % values but also over the whole range of observations. (2) The gain in performances was mainly brought by the use of static, catchment and reach attributes. (3) Customizing the loss function to emphasize the model errors on extreme temperature values did not lead to significant gains in test performances. This study further confirms the suitability of well-trained LSTM models for extreme stream temperature values, offering significant advantages for water management at data-sparse regions during summer periods.