Poster

IAMAS

JMP02 - Machine Learning in atmospheric, ocean and earth-system prediction: forecasting, simulation and scientific analysis

HYDRO: Hybrid Deterministic-Residual Diffusion Framework for Precipitation Nowcasting

Accurate precipitation nowcasting is essential for disaster mitigation, agricultural management, infrastructure protection, urban planning, and socioeconomic stability. Traditional numerical weather prediction models effectively forecast large-scale weather patterns but often lack the necessary temporal and spatial resolution required for precise short-term rainfall prediction. Recent developments in diffusion-based generative deep learning models have demonstrated promising capabilities in capturing complex precipitation dynamics, providing high-quality probabilistic forecasts and addressing existing limitations. This study proposes a unified precipitation nowcasting framework that integrates residual diffusion modeling with systematically optimized deterministic forecasting components. While previous approaches have explored convolutional neural networks (CNNs), recurrent neural networks (RNNs), and Transformer-based architectures individually, comprehensive comparisons of these deterministic components within a diffusion-based precipitation forecasting framework remain limited. Additionally, emerging architectures such as state space models (SSMs), known for efficiently modeling complex temporal dynamics, have not yet been extensively investigated in precipitation nowcasting. The primary objectives of this research are to rigorously evaluate CNN-based, Transformer-based, and SSM-based deterministic components to determine optimal architectures for predictive accuracy, geographic generalization, and computational efficiency within the unified residual diffusion framework. Additionally, this work critically examines the necessity of explicit recurrence mechanisms within deterministic components, comparing recurrent approaches against recurrent-free strategies based on convolution or attention. Clarifying the effectiveness of recurrence will guide informed architectural decisions for future precipitation nowcasting systems. Key contributions include (1) comprehensive experimentation of deterministic architectures (CNN, Transformer, and SSM) within a unified deterministic-residual diffusion framework, (2) systematic evaluation of recurrent versus recurrent-free deterministic strategies, (3) extensive validation across geographically diverse datasets (SEVIR, MeteoNet, GK2A Radar), and (4) the integration of advanced precipitation-specific evaluation metrics such as Critical Success Index (CSI), Heidke Skill Score (HSS), Fraction Skill Score (FSS), Threat Score (TS), and Probability of Detection (POD), alongside conventional error-based measures (MSE, MAE, RMSE) and perceptual metrics (SSIM, PSNR, LPIPS). Collectively, these contributions aim to significantly enhance precipitation nowcasting methods, influencing operational forecasting practices and ultimately strengthening global resilience against extreme weather events.