Deep Learning for Daily Precipitation and Temperature Downscaling

Downscaling is a critical step to bridge the gap between large-scale climate information and local-scale impact assessment. This study presents novel deep learning approach: Super Resolution Deep Residual Network (SRDRN) for downscaling daily precipitation temperature. approach was constructed based on an advanced convolutional neural network with residual blocks batch normalizations. The data augmentation technique utilized address overfitting that due highly imbalanced nonprecipitation days sparse extremes. Synthetic experiments were designed downscale maximum/minimum temperature from coarse resolutions (25, 50, 100 km) high resolution (4 km). results showed that, during validation period, SRDRN not only captured spatial temporal patterns remarkably well, but also reproduced both extremes in different locations time at local scale. Through transfer learning, trained model one region directly applied another environment, notable improvement compared classic statistical methods. outstanding performance of stemmed its ability fully extract features without suffering degradation issues incorporations blocks, normalizations, augmentations. thus powerful tool can potentially be leveraged any hydrologic, climate, earth system data.