Automated Quality Control Scheme for GPM Satellite Precipitation Products.

The constellation approach underpinning precipitation products such as the Integrated Multi‐satellitE Retrievals for GPM (IMERG) is key to achieving high resolution, but the use of data from multiple sources can unintentionally incorporate instrumental artifacts. Here, we introduce a machine learning–based anomaly detection scheme called SPEEDe, which processes a two‐dimensional precipitation field into a re‐estimated precipitation field that can be compared with the input. Large differences identify IMERG fields with bad orbit data, separating most of the bad cases from the good cases. When modified to process the passive microwave inputs, SPEEDe can pick out orbits with bad data, enabling quality control on these IMERG inputs. SPEEDe works by producing a locally realistic‐looking precipitation field when given unphysical data, which results in a larger‐than‐normal difference between the input and the output. SPEEDe is implemented as an automated quality control for GPM precipitation products. Plain Language Summary: A machine learning–based scheme, SPEEDe, has been developed to identify bad data in satellite precipitation products. This scheme is able to pick out precipitation maps that have unrealistic patterns better than a conventional approach based on the distribution of values. It can be run as the precipitation data are being produced, thus allowing for automated quality control. Avoiding manual intervention is important in computing quality‐controlled precipitation data in near‐real time for applications such as flood monitoring that require highly reliable data with minimal delay. Key Points: A machine learning scheme called SPEEDe can detect artifacts in satellite precipitation fields based on an aggregate grid‐wide scoreSPEEDe identifies bad data by modifying anomalous inputs substantially to produce realistic‐looking outputs, leading to a large scoreSPEEDe is being applied operationally in GPM data production for quality control [ABSTRACT FROM AUTHOR]