Intercomparison of Machine Learning Models for Spatial Downscaling of Daily Mean Temperature in Complex Terrain

• Published in: Atmosphere 2024-09, Vol.15 (9), p.1085
• Main Authors: Bhakare, Sudheer, Dal Gesso, Sara, Venturini, Marco, Zardi, Dino, Trentini, Laura, Matiu, Michael, Petitta, Marcello
• Format: Article
• Language: English
• Subjects: Accuracy; Artificial neural networks; Climate change; Climate models; Comparative analysis; convolutional neural network; Daily temperatures; downscaling; Elevation; Geospatial data; Intercomparison; Learning algorithms; Machine learning; Mean daily temperatures; Mean temperatures; Neural networks; random forest; Regression analysis; Seasons; Summer; T2M; Temperature; Terrain; Topography; Valleys; Variables; Weather; Winter
• ISSN: 2073-4433; 2073-4433
• DOI: 10.3390/atmos15091085

We compare three machine learning models—artificial neural network (ANN), random forest (RF), and convolutional neural network (CNN)—for spatial downscaling of temperature at 2 m above ground (T2M) from a 9 km ERA5-Land reanalysis to 1 km in a complex terrain area, including the Non Valley and the Adige Valley in the Italian Alps. The results suggest that CNN performs better than the other methods across all seasons. RF performs similar to CNN, particularly in spring and summer, but its performance is reduced in winter and autumn. The best performance was observed in summer for CNN (R2 = 0.94, RMSE = 1 °C, MAE = 0.78 °C) and the lowest in winter for ANN (R2 = 0.79, RMSE = 1.6 °C, MAE = 1.3 °C). Elevation is an important predictor for ANN and RF, whereas it does not play a significant role for CNN. Additionally, CNN outperforms others even without elevation as an additional feature. Furthermore, MAE increases with higher elevation for ANN across all seasons. Conversely, MAE decreases with increased elevation for RF and CNN, particularly for summer, and remains mostly stable for other seasons.