Abstract
This study examines the performance of three machine learning models—namely, Artificial Neural Network (ANN), Random Forest (RF), and Convolutional Neural Network (CNN)—for spatial downscaling of seasonal forecasts of daily minimum temperature from 12 km to 250 m horizontal resolution. Downscaling is carried out with a one-month lead time, with analysis split into short-term (1 to 8 days) and extended (9 to 28 days) forecast periods, allowing a detailed assessment of the performance of models over time. Results suggest that CNN outperforms ANN and RF, achieving lower Root Mean Square Error (ranging from 2.04 °C to 2.66 °C) and Mean Absolute Error (1.59 °C to 2.03 °C) along with higher correlation (0.75 to 0.88) and reduced bias (−0.38 °C to −0.68) across all seasons, for the short term. The CNN model also exhibits superior performance in frost prediction, with the highest F1 score (0.78) and lowest False Discovery Rate (0.30) in predicting frost events, particularly in early spring for the short-term forecast period over 2010–2018. However, errors increase in transitional months, like April, and in the extended forecast period, confirming the intrinsic challenges inherent to predicting frost events in these months. Despite the decreased skills for extended forecast periods, results suggest that the CNN model’s effectiveness for spatial downscaling of minimum temperature and frost forecasting over complex terrain provides a valuable tool for frost risk management.