DROP: Diffusion-based Rainfall down-scaling with Observational and Prior guidance across scales

High-resolution sub-hourly precipitation data are essential for studying short-duration rainfall extremes, flash-flood risk, and hydrological impacts, yet existing products often trade spatial coverage for observational accuracy. Satellite and reanalysis products provide broad coverage but remain subject to substantial uncertainty, whereas radar and rain-gauge observations are more accurate but spatially sparse. To tackle these challenges, we propose DROP, a Diffusion-based Rainfall down-scaling framework with Observation and Prior guidance. To the best of our knowledge, DROP is the first latent diffusion framework for sub-hourly precipitation downscaling and value correction. DROP learns high-resolution precipitation patterns from reanalysis data and incorporates meteorological predictors, including digital elevation model (DEM) and synoptic weather type (SWT), to guide the generation process. The learned prior can be directly transferred to spatially continuous precipitation products, such as satellite observations, while sparse radar and rain-gauge measurements are incorporated to correct precipitation values. Evaluation against unseen AGCD gauge observations, which are entirely excluded from both training and generation, demonstrates that DROP consistently improves precipitation reconstruction and observational consistency, validating its effectiveness in integrating heterogeneous precipitation products and observations.