Precipitation data are vital for understanding the water cycle, managing resources, and mitigating flood and drought risks. Yet current precipitation observations and monitoring face critical gaps in accuracy, coverage, and timeliness. To address these gaps, we present a framework based on explainable spatiotemporal vision transformer (XTPrecip) that learns radar, satellite, reanalysis, topography, and gauge observations to generate hourly precipitation fields as high as 1-km resolution. XTPrecip shows higher accuracy, resolutions, and lower latency relative to state-of-the-art satellite, radar, and merged precipitation products and substantially improves extreme events, as tested in the flood-prone Appalachian Mountains region. Through transfer learning with retraining using only a few gauges, it addresses large gaps in radar precipitation in the western United States, accurately estimates extreme rainfall events during 2025 Texas flood while reducing the latency of calibrated satellite estimates from weeks to hours, and extended these gains to coastal regions of eastern Brazil. Further, XTPrecip reveals the sources of contributions to accurate precipitation monitoring. These results suggest a major advance towards accurate and efficient high-resolution precipitation observations and monitoring over the globe.