Improved uncertainty estimates for geophysical parameter retrieval

Abstract

Algorithms for retrieval of geophysical parameters from radiances measured by instruments onboard satellites play a large role in helping scientists monitor the state of the planet. Current retrieval algorithms based on neural networks are superior in accuracy and speed compared to physics-based algorithms like iterated minimum variance (IMV). However, they do not have any form of error estimation, unlike IMV. This thesis examines the suitability of several different approaches to adding in confidence intervals and other methods of error estimation to the retrieval algorithm, as well as alternative machine learning methods that can both retrieve the parameters desired and assign error bars. Test datasets included both current generation operational instruments like AIRS/AMSU, as well as a hypothetical future hyper- spectral microwave sounder. Mixture density networks (MDN) and Sparse Pseudo Input Gaussian processes (SPGP) were found to be the most accurate at variance prediction. Both of these are novel methods in the field of remote sensing. MDNs also had similar training and testing time to neural networks, while SPGPs often took three times as long to train in typical cases. As a baseline, neural networks trained to estimate variance were also tested, but found to be lacking in accuracy and reliability compared to the other methods.