高解析度降雨降尺度對於地形複雜地區的水文氣象應用至關重要，特 別是在如台灣這樣的山區環境中。本研究提出 DiffRainNet，一種新穎的深 度學習架構，用於將粗解析度的大尺度環流資料轉換為細緻的降雨分布。該 模型結合三大關鍵模組：（1）基於 FiLM 的地形感知調製機制，可根據地形 資訊動態調整特徵表徵；（2）多層級注意力機制，包括 SEBlock、CBAM 與 Self-Attention 模組，有效提取具代表性的空間特徵；（3）條件式擴散模 型，透過學習殘差修正進行降尺度，以還原真實的降雨結構與強度。 實驗以台灣東北部冬季季風為案例，結果顯示 DiffRainNet 在數值誤 差與空間結構保留方面皆明顯優於基準模型。相較於標準 U-Net， DiffRainNet 使 RMSE 降低 47%、MAE 降低 53%，並在中雨與大雨事件中 獲得最高的 SSIM 分數。在主要的 10–30 mm/6 小時降雨區間（涵蓋多數 超過 10 mm 的降雨事件）中，DiffRainNet 較 U-Net 基準模型的 MAE 和 RMSE 降低了 48%，突顯其對最常見降雨強度的優化能力。視覺化比較進一 步驗證，DiffRainNet 能夠重建與地形一致的降雨梯度、保留精細空間結 構，並有效抑制乾燥區域的虛假信號。這些結果凸顯了 DiffRainNet 作為 一種物理一致且資料驅動的高解析度降雨模擬解決方案，在複雜地形環境 中的應用潛力。;High-resolution rainfall downscaling is essential for hydrometeorological applications in regions with complex terrain such as Taiwan. This study proposes DiffRainNet, a novel deep learning framework for reconstructing fine-scale rainfall fields from coarse-resolution atmospheric circulation data. DiffRainNet consists of three key components: (1) FiLM-based terrain-aware conditioning, which dynamically modulates feature representations according to topographic context; (2) multi-level attention mechanisms---including SEBlock, CBAM, and Self-Attention---that selectively emphasize informative spatial features across scales; and (3) a conditional diffusion model that refines coarse predictions by learning residual corrections, thereby restoring realistic rainfall structure and intensity. Experiments over northeastern Taiwan under winter monsoon conditions demonstrate that DiffRainNet substantially improves both numerical accuracy and spatial coherence. Compared to a standard U-Net, it reduces RMSE by 47% and MAE by 53%, while also achieving the highest SSIM scores in both moderate and heavy rainfall scenarios, indicating superior preservation of spatial rainfall structures. In the dominant 10-30 mm/6 hr rainfall regime covering most occurrences above 10 mm, DiffRainNet lowers MAE and RMSE by about 48% compared to the U-Net baseline, highlighting its capability to refine the most frequently occurring rainfall intensities. Visual comparisons further confirm that DiffRainNet reconstructs terrain-aligned rainfall gradients and preserves finescale spatial structures while suppressing spurious signals in dry regions. These results underscore its potential as a physically consistent, data-driven solution for high-resolution rainfall modeling in complex orographic environments.
