Satellite Precipitation Algorithms and AI

Abstract:

Satellite precipitation measurement must be viewed as two step problem. The first is the traditional estimation problem, where either physical or statistical methods are used to infer surface precipitation from a number of passive or active spaceborne measurements. Where the information content is high, such as with spaceborne radars, methods gravitate towards physical inversions. This avoids having to deal with the lack of globally applicable training data – particularly over oceans and data sparse regions where satellite estimates are most needed. When information content is lower, such as with passive microwave or VIS/IR sensors, space or ground based radars become viable training datasets. This initially led to regression techniques or physical methods with constraints imposed from this training data. More recently, these techniques have begun to give way to AI methods – Bayesian, Random Forests, Quartile Regression Neural Networks and other techniques designed to derive surface rainfall from a set of spaceborne observations. Applicable training data becomes crucial in these cases. A quick examination of some of the current techniques further makes clear that when the information content is low, ancillary data is often needed to constrain the algorithm, thereby reducing the spread of validation results from one precipitation regime to another. Finding the meteorological data that defines these regimes and minimizes the validation statistics between different regions and seasons is the second problem that needs to be addressed and one that is also suited to AI approaches. This talk, aside from illustrating various approaches for constructing the algorithm itself, will also delve into our exploration of Meta-learning approaches which appear suitable to solve the second problem.