Search for: All records

Abstract Satellite precipitation retrieval is inherently an underdetermined inverse problem where additional physical constraints could substantially enhance accuracy. While previous studies have explored static (pixel‐based/spatial‐context‐based) environmental variables at discrete satellite observation times, their temporal dynamic information remains underutilized. Building on our earlier finding that retrieval errors depend on storm progression (event stage), we propose a new, physically interpretable mechanism for improving retrievals, namely, leveraging environmental variables' temporal dynamics as proxies for event stages. Using IMERG satellite product and GV‐MRMS as ground‐truth over CONUS (2018–2020), we first demonstrate robust coevolution patterns of environmental variables and satellite errors throughout events, and show that these variables' temporal gradients reliably infer event stages. We then demonstrate that incorporating these variables and their gradients into a machine‐learning post‐processing framework improves retrieval accuracy. This work inspires and guides more thorough utilization of spatiotemporal atmospheric fields encoding rich physical information within advanced machine‐learning frameworks for further algorithm improvement. 

Abstract Understanding the nature and origin of errors in satellite precipitation products is important for applications and product improvement. Here we propose a new error decomposition scheme incorporating precipitation event (continuous rainy periods) information to characterize satellite errors. Under this framework, the errors are attributed to the inaccuracies in event occurrence, timing (event start/end time), and intensity. The Integrated MultisatellitE Retrieval for Global Precipitation Measurement (IMERG) is used as our test product to apply the method over CONUS. The above‐listed factors contribute approximately 30%, 20%, and 50% to the total bias, respectively. Significant asymmetry exists in the temporal distribution of biases throughout events: early event endings cause threefold more precipitation amount bias than late event beginnings, while early event beginnings cause fourfold more bias than late event endings. Dominant contributors vary across seasons and regions. The proposed error decomposition provides insight into sources of error for improved retrievals.