Combining Kriging and Neural Networks for Station‐Based Precipitation Forecasting in Radar‐Sparse Alpine Regions

ABSTRACT

Mountainous terrain presents particular challenges for precipitation forecasting due to complex topography, strong spatial variability, and limited radar coverage. This study evaluates a lightweight station‐based forecasting approach that combines kriging‐based preprocessing and neural‐network models for short‐term precipitation prediction in the Swiss canton of Valais. Using 10‐min observations from the SwissMetNet network (2019–2023), we first assess the role of kriging for completing station time series and then train a two‐stage forecasting model composed of a binary classifier for precipitation occurrence and a regression model for non‐zero precipitation amounts. Several lightweight neural architectures are compared, including multilayer perceptrons, Long Short‐Term Memory networks, and Kolmogorov–Arnold Networks. Additional benchmark analyses are used to assess the contribution of the interpolation step and to compare the forecasting models against simpler alternatives. A recursive forecasting strategy combines occurrence and intensity predictions to extend forecasts up to 24 h ahead. The results show measurable predictive skill relative to persistence, while performance remains more limited for precipitation intensity and upper‐percentile wet periods. Overall, the study provides a reproducible and computationally efficient assessment of station‐based precipitation forecasting in a radar‐sparse alpine environment, while highlighting both the potential and limitations of such lightweight approaches.