J4C.1 Application of a U-Net to Microwave Integrated Retrieval System (MiRS) Retrieved Vertical Profiles for Improved Precipitation Estimation

Monday, 29 January 2024: 4:30 PM
338 (The Baltimore Convention Center)
Shuyan Liu, CIRA, Fort Collins, CO; and C. Grassotti and Q. Liu

The NOAA Microwave Integrated Retrieval System (MiRS) retrieves vertical profiles of water vapor, cloud liquid water, graupel water, rain water, and temperature. These profiles include important information related to atmospheric stability, vertical distribution of hydrometeors, as well as precipitation occurrence and intensity. MiRS uses a post-processing algorithm which uses vertically integrated values of the hydrometers to determine surface precipitation rates and does not consider atmospheric stability. In this study, we applied a U-Net convolutional neural network architecture on MiRS retrieved vertical profiles to adjust retrieved precipitation rates from NOAA-20 Advanced Technology Microwave Sounder (ATMS) observations. The training input features are aforementioned vertical profiles, surface precipitation rate, and satellite observation zenith angle. The training target data are hourly precipitation rates from the operational Multi-Radar/Multi-Sensor System (MRMS) blended radar-rain gauge analysis over the Continental U.S. (CONUS). The MRMS analysis contains information on precipitation type (e.g. stratiform, convective, frozen, liquid, etc.) which can be used to further enhance precipitation accuracy for various cloud systems. The U-Net was trained using one year of collocated MiRS and MRMS data over the CONUS during 2021. Independent validation of the U-Net was performed using data from 2022 and 2023. The input image-like data dimensions are 96 (FOVs) × 96 (vertical pressure levels) with five and four convolutional layers at contraction and expansion paths, respectively. The filter’s size is 5×5 with maximum pooling layer configured with same padding and the stride is 2 pixels. Our previous study using a similar U-Net architecture to correct MiRS retrieved precipitation showed that U-Net improved precipitation performance in terms of histogram distributions, bulk error statistics, and categorical scores. Input data to that model were NOAA-20 ATMS retrievals of total precipitable water and precipitation rate, along with geolocation information. The training dataset is the same as the current study and the validation is 2022. The U-Net consistently and significantly improved the light rain frequency and geospatial distribution, greatly improved estimates of spatially averaged daily precipitation, and noticeably improved the MiRS annual total precipitation spatial distribution. Building on the previous study, the addition of vertical profile information with a focus on a variety of cloud and precipitation systems aims to further improve MiRS precipitation retrievals.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner