Scene Invariants for Quantifying Radiative Transfer Uncertainty

Remote imaging spectroscopy investigations typically use Radiative Transfer Models (RTMs) to predict the measured radiance at the sensor given a specific surface and atmospheric state. Discrepancies between RTM model assumptions and physical reality can lead to subtle and systematic errors in remote surface property estimates. We present a statistical approach to quantify these model errors without invoking ground truth measurements. Our approach exploits scene invariants - properties of the environment which are stable over space or time, which can be used to estimate RTM model discrepancies. We describe techniques for discovering these features opportunistically in flight data. We then demonstrate data-driven methods which allow investigators to estimate the aggregate errors due to model discrepancy without having to explicitly identify the mechanisms involved. The resulting distributions can improve the accuracy of retrievals and posterior uncertainty predictions.