Comment on amt-2022-15

Abstract

The Global Precipitation Measurement (GPM) mission aims to provide global measurements of precipitation with a temporal resolution of three hours in order to allow close monitoring of the global hydrological cycle. To achieve global coverage at such high temporal resolution, GPM combines observations from a constellation of passive microwave (PMW) sensors. The Goddard Profiling Algorithm (GPROF) is the operational retrieval for precipitation and hydrometeor profiles from these sensors. Aiming to investigate the effect of the retrieval algorithm on the accuracy of the PMW precipitation retrieval, we present two neural network based, probabilistic implementations of GPROF: GPROF-NN 1D, which processes individual pixels just as the current GPROF algorithm, and GPROF-NN 3D, which employs a convolutional neural network to incorporate structural information into the retrieval. To isolate the impact of the retrieval method from any issues in the training data, the GPROF and GPROF-NN algorithms are evaluated on a separate test set that has the same statistical properties as the data used for the training of the GPROF-NN algorithms and the development of the current GPROF algorithm. Comparison of GPROF and the GPROF-NN 1D algorithm shows that by replacing GPROF with an identical neural network based retrieval the accuracy of the retrieved surface precipitation from the GPM Microwave Imager (GMI) can be improved by 10 to 25 % in terms of absolute error, root mean squared error and symmetric mean absolute percentage error. Comparable improvements are observed for the retrieved hydrometeor profiles and their column integrals. The improvements are consistent spatially as well as with respect to different surface types. The effective resolution in along track direction of the retrieved surface precipitation fields is increased from 23 km to 14 km. Similar, additional improvements are found for the GPROF-NN 3D retrieval over the performance of the GPROF-NN 1D retrieval, showing the added benefits of incorporating structural information into the retrieval. The effective resolution in along-track direction of the GPROF-NN 3D algorithm is reduced to 13.5 km, which is the upper limit imposed by the along track separation of consecutive scan lines. Comparable improvements are found also when the algorithms are applied to synthetic observations from the cross track scanning Microwave Humidity Sounder (MHS) sensor. Application of the retrieval algorithm to real observations from the GMI and MHS sensors of Hurricane Harvey suggest that these improvements can be expect to carry over to operational application. The novel GPROF-NN algorithms presented here were designed to be functionally equivalent to the current implementation, which enables the neural network approach to replace the current Bayesian scheme in a future update. Despite their superior retrieval accuracy, the single CPU core runtime required for the operational processing of an orbit of observations is lower than that of GPROF. The GPROF-NN algorithms thus promise to be a simple and cost efficient way to improve the accuracy of the PMW precipitation retrievals of the GPM constellation and thus help improve the monitoring of the global hydrological cycle.