To establish the value for nowcasts, the concept of the time value of a satellite-derived precipitation estimate is introduced: The more precise precipitation estimates (e.g,from Tomorrow.io radars), propagated by the nowcast, will exceed the baseline skill of real-time geostationary-derived precipitation estimates for a longer time period than a less precise estimate (e.g., from a sounder). These time value metrics drive the combined latency and revisit requirements of the Tomorrow.io constellation.
To provide a common basis for establishing the skill of these sensors, precipitation retrievals from multi-channel geostationary satellite data, the TROPICS Pathfinder microwave sounder (a proxy for the Tomorrow.io microwave sounder), and the Tomorrow-R1 and Tomorrow-R2 were assessed using probabilistic machine learning techniques, including convolutional neural networks (CNNs) to take advantage of spatial information from the sounders and geostationary imagery. Although the hierarchy of precipitation accuracy was as expected (radars best, followed by sounders, then geostationary), the improvement offered by CNNs is noteworthy. In many cases, the data-driven geostationary retrievals outperform products that incorporate passive microwave retrievals such as IMERG-Early and CMORPH, in agreement with other recent studies. These retrievals have been integrated into the Tomorrow.io platform and disseminated through the API, and will be further enhanced by data from the satellite constellation as it comes online.

