Rainfall Estimation From TEMPEST-D CubeSat Observations: A Machine-Learning Approach

Abstract

In this study, a machine-learning model was used to produce surface rainfall estimates from Temporal Experiment for Storms and Tropical Systems – Demonstration (TEMPEST-D) microwave radiance observations from a CubeSat. The machine-learning model is based on an artificial neural network (ANN). The space-borne TEMPEST-D sensor performed brightness temperature (TB) observations at five frequencies (i.e., 87, 164, 174, 178, and 181 GHz) during its nearly three-year mission. The TEMPEST-D TBs were used as inputs, and the multiradar/multisensor system (MRMS) radar-only quantitative precipitation estimation product at the surface was used as the ground truth to train the ANN model. A total of 19 storms were identified that were simultaneously observed by TEMPEST-D and ground weather radar over the contiguous United States. The training dataset used 14 of the 19 storm cases. The other five storm cases, consisting of three continental storms and two land-falling hurricanes, were used for independent testing. A spatial alignment algorithm was developed to align the TEMPEST-D observed storm with the ground radar measurement of the storm. This study showed that the TEMPEST-D TBs captured storm features as well as current-generation satellite sensors, such as the global precipitation mission microwave imager. The results of this study demonstrated that the rainfall estimated from TEMPEST-D matches well with the MRMS surface rainfall product in terms of rainfall intensity, area, and precipitation system pattern. The average structural similarity index measure score for the five independent test cases is 0.78.