Uncertainty Quantification of Satellite-Based Essential Climate Variables Derived from Deep Learning

Abstract Accurate uncertainty information associated with essential climate variables (ECVs) is crucial for reliable climate modeling and understanding the spatiotemporal evolution of the Earth system. Recent developments in deep learning have remarkably advanced the estimation of ECVs with improved accuracy. However, the quantification of uncertainties associated with outputs of such deep learning models has yet to be widely adopted. This survey explores the types of uncertainties associated with ECVs derived from deep learning methods, including aleatoric (data) and epistemic (model) uncertainty, and the techniques to quantify them. The focus is on highlighting the importance of considering uncertainty associated with inputs in the deep learning models to account for the dynamic and multifaceted nature of satellite observations. The survey starts by clarifying the definitions of aleatoric and epistemic uncertainties and their roles in a typical satellite observation processing workflow, followed by bridging the gap between conventional statistical and deep learning views on uncertainties. Then, we comprehensively review the existing uncertainty quantification methods for deep learning algorithms and discuss their strengths and limitations. A comprehensive literature review about quantifying uncertainties in the deep learning estimates of ECVs follows the theoretical survey, covering a wide range of ECVs. The specific need for modification to fit the requirements from both the Earth observation side and the deep learning side in such interdisciplinary tasks is highlighted. We further demonstrate our findings with two selected ECV examples, snow cover and terrestrial water storage, to provide clear insights into different methods by promoting quantitative comparison. In the end, we summarize our findings and provide perspectives for future research.