Abstract Precipitation is a critical component of the hydrologic cycle, affecting water availability and flood risk. Its estimation can help protect lives and mitigate property damage. Machine learning (ML) models have become popular for estimating precipitation; however, when choosing a target dataset, certain characteristics must be considered. In situ data are ideal to train an ML model, but alternatives must be explored when such data are sparse. Thus, a U-Net was trained three times using the Climate Prediction Center (CPC) 4-km infrared data as input and different precipitation target datasets. These included the National Severe Storms Laboratory (NSSL) Multi-Radar Multi-Sensor (MRMS) system, NASA’s Integrated Multi-satellitE Retrievals for Global Precipitation Measurement Final Run (IMERG Final), and CPC Combined Passive Microwave Precipitation (MWCOMB). The results were evaluated against Stage IV. The MRMS model performed best with a correlation coefficient (CC) of 0.50 and a probability of detection (POD) of 0.90 at an hourly scale. The IMERG Final model followed with a CC of 0.47 and a POD of 0.85, making it a viable alternative given its temporal resolution and data availability. The MWCOMB model overestimated precipitation and had a CC and POD of 0.45 and 0.82. Across daily and monthly scales, MRMS consistently outperformed the other models, with IMERG Final ranking second. Due to MWCOMB’s persistent overestimation, different data preparation may be required for effective use. In conclusion, MRMS is the most suitable dataset for ML-based precipitation estimation due to its in situ nature and robust performance; however, IMERG Final is a viable alternative. Significance Statement Precipitation estimation is extremely important for hydrologic applications. However, limited access to in situ precipitation data is a challenge for machine learning (ML) models, as they require large amounts for training. This study explores how the selection of precipitation target datasets affects the performance of a U-Net model for precipitation estimation. Using infrared data as input, the results not only show superior performance of the National Severe Storm Laboratory (NSSL) Multi-Radar Multi-Sensor (MRMS) but also demonstrate that NASA’s IMERG Final offers a viable alternative when in situ data are unavailable. The findings of this study provide insight for researchers seeking to improve ML-based precipitation models, specifically in regions with sparse in situ data.
Arellano et al. (Wed,) studied this question.