Predicting Region-Dependent Biases in a GOES-16 Machine Learning Precipitation Retrieval

Abstract

Abstract Despite its long history, improving upon current precipitation estimation techniques remains an active area of research. While many methods exist to assess precipitation, the use of satellites has allowed for near-global observation. However, satellites do not directly sense precipitation, resulting in retrieval uncertainties. Analysis of these uncertainties is typically conducted through validation studies, which, while necessary, are sensitive to local conditions. As such, predicting retrieval uncertainties where there is no validation data remains a challenge. In this study, we propose a method by which validation statistics can be extended to other regions. Using a neural network–style retrieval, the Geostationary Operational Environmental Satellite–16 (GOES-16) Precipitation Estimator using Convolutional Neural Networks (GPE-CNN), we show that, by exploiting the information content of both the satellite and ancillary meteorological data, one can predict large-scale retrieval behaviors over other regions without the need for that region’s validation data. By developing classes using satellite information content, we demonstrate bias prediction improvement of up to 83% relative to a simple extension of mean bias. Including relative humidity information improves the overall prediction by up to 98% relative to the original mean bias. Although limited in scope, this method presents a pathway toward characterizing uncertainties on a broader scale.