Bayesian Gaussian Process for Correcting Artifacts from Atmospheric Correction and Sensor Noise - A Performance Evaluation

Artifacts in hyperspectral images are physically unrealistic values in reflectance spectra, caused either by incorrect estimation of parameters in atmospheric correction or by sensor noise. Atmospheric correction artifacts often appear as excessively high or low values where water has partial absorption. Sensor often appear as single band spikes in spectra. We previously showed that a Gaussian Process trained on library spectra can effectively detect and correct these artifacts. In this paper, we show that correcting for these artifacts can improve processing of the imagery for anomaly detection. Python code for this research shared openly.