Improvement of NOAA SFR algorithm through machine learning approach

Abstract:

The NESDIS operational Snowfall Rate (SFR) product is retrieved from passive microwave measurements taken by radiometers aboard polar-orbiting satellites. Currently, the product is generated at near real-time from a suite of satellites operated by NOAA or its partner agencies. The SFR algorithm performs well in relatively warm regions, but its performance degrades under colder weather conditions due to diminished accuracy in snowfall detection. The reduced accuracy and lack of retrievals in colder weather are major obstacles to product applications in areas where snowfall observations are scarcest such as the western US and most of Alaska. The SFR algorithm consists of two main components: snowfall detection (SD) and snowfall rate estimation. The SD component of the SFR product is a logistic regression model trained from collocated satellite and ground observations. However, logistic regression is a relatively simple statistical model that limits the performance of the SD model under colder weather conditions. To improve the SD algorithm, we have built a collocated global precipitation dataset between ground measurement and S-NPP ATMS satellite measurements from years 2012-2018. A neural network classification algorithm will be trained based on the global precipitation dataset. We will present preliminary results of the SD model in terms of probability of detection (POD), false alarm rate (FAR) and Heidke Skill Score (HSS). SFR is derived from a 1DVAR-based algorithm to retrieve cloud properties and surface emissivity at six window and wave vapor sounding frequencies. However, the current 1DVAR algorithm utilizes a simple one-layer two-stream Radiative Transfer Model (RTM), and the single scattering properties of ice clouds were computed based on Mie theory (i.e. spherical particles). To improve the accuracy of the RTM, we have built a 25-layer full-stream RTM based on DISORT and adopted a non-spherical ice particle model for ice cloud. A fast yet accurate machine learning based forward model will be trained by a large simulation dataset with various combinations of surface, atmosphere, and cloud conditions. This approach is expected to significantly improve the computational efficiency of the 1DVAR algorithm. Preliminary results of the improved 1DVAR algorithm and case studies will be presented.