Uncertainty quantification of aboveground biomass predictions produced by fine-tuned geospatial foundation models

Abstract

Portugal, among other European Union states, is required to build eco-system monitoring systems to deliver environmental monitoring based products periodically for its territory. In this work, we present a methodology to estimate uncertainty in Portugal’s above-ground biomass (AGB) maps. The AGB predictions are generated using Prithvi — a geospatial foundation model — fine-tuned on sparse space-borne data (i.e., GEDI measurements) collected across different eco-regions. We focus on identifying aleatoric uncertainty arising due to the challenges of using satellite imagery. Using test-time augmentation (TTA) the stochasticity in satellite imagery is captured by different data-transformations. While TTA has been used for uncertainty quantification (UQ) in medical image segmentation and classification confidence, our application extends the approach for pixel-wise regression tasks. UQ provides a distribution of AGB predictions instead of a point prediction making it more robust against the uncertainties in the data, in turn increasing the trustworthy usability of Prithvi at scale. With our approach we share practical UQ solutions to challenges faced in using satellite imagery with sparse labels to predict AGB values in support of a broad range of applications in forest carbon accounting and ecosystem stewardship.