Scale Aware Deep Learning for Precipitation Prediction with Hybrid Loss Swin Transformer U-Net

We present a deep learning framework for high-resolution precipitation nowcasting that explicitly addresses the multiscale and nonlinear nature of convective rainfall. Our model, PrecNet, combines a Swin Transformer-based U-Net with spatial box filtering, enabling systematic control and analysis of spatiotemporal forecast skill. By coupling hybrid loss functions—including Balanced Mean Square Error(BMSE), SoftCSI, and Wasserstein distances—with differential prediction, PrecNet captures both intensity fidelity and event-based extremes. Using temporal correlation and the integral time scale (ITS), we reveal that coarser filtering suppresses small-scale noise, yielding improved RMSE and cross-correlation, while critical success index (CSI) gains are confined within the ITS. Beyond this limit, predictability declines sharply. Benchmarks against Pysteps, DEEPRANE, and an operational Korean NWP model demonstrate competitive or superior skill, particularly for extreme events, while maintaining real-time inference on a single GPU. Our findings underscore the importance of hybrid learning, multi-metric evaluation, and physical interpretability in AI-based nowcasting, offering insight into the fundamental limits of predictability and new pathways for operational forecasting in resource-constrained settings.