A Comparison of Uncertainty Quantification Methods for Earth Observation Image Regression Data

Abstract

Over the past decade, neural networks (NNs) have been successfully applied to earth observation (EO) data and opened new research avenues. Despite the theoretical and practical advances of these techniques, NNs are still considered black box tools and by default are only designed to give point predictions. However, the vast majority of EO applications demand reliable uncertainty estimates that can support practitioners in decision making tasks. This work provides a theoretical and quantitative com- parison of popular uncertainty quantification methods for NNs with the focus on univariate image regression problems in the EO domain. More specifically, we consider the task of predicting tree-cover percentage from 4 channel satellite imagery. Given a base architecture consisting of a Ran- dom Convolutional Feature (RCF) extractor and a subse- quent Multi-layer Perceptron Network (MLP), we apply a wide range of uncertainty quantification (UQ) methods to compare and evaluate their performance under geospatial distribution shifts.