DYffCast: Regional Precipitation Nowcasting Using IMERG Satellite Data. A case study over South America

TL;DR

This work extends the DYffusion framework to precipitation nowcasting using IMERG satellite data, targeting 4-hour forecasts over South American regions. A novel loss, LCB, combines perceptual and classical pixel-wise terms to improve rainfall feature fidelity, while key training and sampling modifications tailor the model to rainfall dynamics. Empirical results show DYffusion-LCB achieving higher CSI across weak, moderate, and heavy rain thresholds and maintaining perceptual sharpness (LPIPS < 0.2) as lead time grows, outperforming ConvLSTM and STEPS baselines. The approach offers a memory-efficient probabilistic forecasting pathway and demonstrates potential for operational nowcasting with further conditioning on atmospheric context.

Abstract

Climate change is increasing the frequency of extreme precipitation events, making weather disasters such as flooding and landslides more likely. The ability to accurately nowcast precipitation is therefore becoming more critical for safeguarding society by providing immediate, accurate information to decision makers. Motivated by the recent success of generative models at precipitation nowcasting, this paper: extends the DYffusion framework to this task and evaluates its performance at forecasting IMERG satellite precipitation data up to a 4-hour horizon; modifies the DYffusion framework to improve its ability to model rainfall data; and introduces a novel loss function that combines MSE, MAE and the LPIPS perceptual score. In a quantitative evaluation of forecasts up to a 4-hour horizon, the modified DYffusion framework trained with the novel loss outperforms four competitor models. It has the highest CSI scores for weak, moderate, and heavy rain thresholds and retains an LPIPS score $<$ 0.2 for the entire roll-out, degrading the least as lead-time increases. The proposed nowcasting model demonstrates visually stable and sharp forecasts up to a 2-hour horizon on a heavy rain case study. Code is available at https://github.com/Dseal95/DYffcast.

Figures (2)

• Figure 1: An illustration of DYffusion sampling: DYffusion forecasts a sequence of $N$ (or $h$) timesteps $x_{t+i_{1}}$ , $x_{t+i_{2}}$ , ... , $x_{t+h}$ given the initial condition $x_{0}$. $x_{t+h}$ is forecasted at each step using $x_{0}$ and the nearest intermediate value $x_{t+i_{n}}$.
• Figure 2: The evolution of Cyclone Yaku over 4 hours beginning at 03:00 UTC on March 9, 2023. The hourly forecast values (undersampled) are shown for each of the five evaluated models.