GPROF-NN: a neural-network-based implementation of the Goddard Profiling Algorithm

Abstract. The Global Precipitation Measurement (GPM) mission measures global precipitation at a temporal resolution of few hours to enable close monitoring the hydrological cycle. GPM achieves this by combining observations from spaceborne radar, constellation passive microwave (PMW) sensors, and geostationary satellites. Goddard Profiling Algorithm (GPROF) is used operationally retrieve all PMW sensors constellation. Since resulting rates serve as input for many level 3 retrieval products, GPROF constitutes an essential component processing pipeline. This study investigates ways improve using modern machine learning methods. We present two neural-network-based, probabilistic implementations GPROF: GPROF-NN 1D, which (just like current implementation) processes pixels individually, 3D, employs convolutional neural network incorporate structural information into retrieval. accuracy retrievals evaluated test dataset consistent with data in development retrievals. allows assessing method isolated representativeness training data, remains major source uncertainty Despite same GPROF, 1D improves retrieved surface Microwave Imager (GMI) 0.079 0.059 mm h?1 terms mean absolute error (MAE), 76.1 % 69.5 symmetric percentage (SMAPE) 0.797 0.847 correlation. improvements Humidity Sounder (MHS) are 0.085 0.061 MAE, 81 70.1 SMAPE, 0.724 0.804 Comparable found hydrometeor profiles their column integrals, well detection precipitation. Moreover, ability resolve small-scale variability improved more than 40 GMI 29 MHS. 3D further MAE 0.043 h?1; SMAPE 48.67 %; correlation 0.897 h?1, 63.42 %, 0.83 Application Hurricane Harvey shows moderate when compared co-located GPM-combined ground-based radar measurements indicating that least partially carry over assessment against independent measurements. Similar MHS do not show equally clear improvements, leaving validation future investigation. Both algorithms make use output original algorithm thus may replace implementation update superior accuracy, single-core runtime required operational orbit lower GPROF. promise be simple cost-efficient way increase