Uncertainty quantification of hyperspectral scene reconstruction methods for high consequence remote sensing applications

Abstract

Hyperspectral imaging (HSI) is a key capability for detecting chemical signatures of gases, liquids, or solids in remote sensing applications. Detection becomes challenging when targets are partially or fully occluded by objects in the scene. Three-dimensional scene understanding methods, including Neural Radiance Fields (NeRFs) and Gaussian Splatting (GS), provide a mechanism to cohesively synthesize multi-observation HSI data into a single scene representation, enabling the ability to effectively “look around” targets or obstructions. These approaches have shown promise for gas plume detection and identification; however, their limitations remain underexplored. We present an ensemble-based approach to quantify both spatial and spectral uncertainty. The proposed method achieves adequate uncertainty coverage and is model agnostic, making it applicable across a wide range of scene representation techniques. Importantly, we demonstrate cases where classic pixel-level metrics fail to indicate missing plume structure, giving a false sense of agreement. Our uncertainty quantification approach provides an efficient diagnostic cue that the model representation in these regions may be deficient, despite apparently reasonable performance metrics. For remote sensing applications, decision-makers require detection outputs accompanied by measures of confidence, particularly in complex environments where occlusions, atmospheric effects, or sparse measurements can degrade performance. By characterizing uncertainty, our method enables unreliable predictions to be flagged, supports prioritization of additional observations, and ultimately improves trust and robustness in HSI-based detection systems.