Performance assessment of groundwater level forecasting with deep learning: a case study of Lower Saxony, Germany.

Groundwater level forecasting with machine learning has been widely studied due to its generally accurate results and little input data requirements. Furthermore, machine learning models for this purpose are set up and trained in a short time when compared to the effort required for process-based, numerical models. Despite the high performance of models obtained at specific locations, applying the same model architecture to multiple sites across a regional area might lead to contrasting accuracies. Likely causalities of this discrepancy in model performance have been barely examined in previous studies. Here, we investigate the link between model performance and the effects of geospatial site characteristics and time series features. Using precipitation and temperature as predictors, we model groundwater levels at approximately 500 observation wells in Lower Saxony, Germany, using a 1-D convolutional neural network with a fixed architecture and hyperparameters tuned for each time series individually. The performances are evaluated against geospatial and time series features using correlation coefficients. Model performance is negatively influenced at sites near waterworks and densely vegetated areas. Besides, the more complex the time series, the higher the metrics, but autocorrelation reduces the model performance. The new insights evidence that further information is required at certain locations to improve model accuracy due to external impacts.