SSRF-Net: A Stagewise Scheduled Rainfall Forecasting Network With an Asymmetric Architecture

Deterministic deep learning models for precipitation nowcasting often face several limitations, including cumulative error in long-sequence predictions, over-smoothing, and a reduced ability to capture rare, high-impact rainfall due to data imbalance. To address these challenges, we propose the stagewise scheduled rainfall forecasting network (SSRF-Net), a convolutional framework for continuous multistep rainfall prediction that achieves lower floating-point operations (FLOPs) than competitive baselines under a standardized evaluation. Our framework introduces a multistage, sliding-window prediction mechanism trained with teacher forcing and scheduled sampling to mitigate error accumulation and stabilize training. We design an asymmetric encoder–decoder (E–D) architecture featuring a differential selective encoder (DSE) for selective feature compression and an additive fusion decoder (AFD) that progressively reconstructs details and alleviates over-smoothing. We further introduce an intensity-weighted Gaussian KL divergence loss that aligns sequence-level Gaussian summaries (means and variances) of predictions and ground truth via a KL term, prioritizing heavy-rain events without assuming pixelwise Gaussianity. Extensive experiments on the KNMI and SEVIR datasets show that SSRF-Net outperforms strong baselines, particularly for moderate to severe precipitation; on KNMI, it yields up to 41.8% higher per-frame critical success index (CSI) at the 30-mm/h threshold, with consistent gains on SEVIR.