10.1109/JSTARS.2024.3518306

This article presents a method to improve the usability of lake ice cover (LIC) maps generated from moderate resolution imaging spectroradiometer (MODIS) top-of-atmosphere reflectance data by providing estimates of aleatoric and epistemic uncertainty. We used a random forest (RF) classifier, which has been shown to have superior performance in classifying lake ice, open water, and clouds, to generate daily LIC maps with inherent (aleatoric) and model (epistemic) uncertainties. RF allows for the learning of different hypotheses (trees), producing diverse predictions that can be utilized to quantify aleatoric and epistemic uncertainty. We use a decomposition of Shannon entropy to quantify these uncertainties and apply pixel-based uncertainty estimation. Our results show that using uncertainty values to reject the classification of uncertain pixels significantly improves recall and precision. The method presented herein is under consideration for integration into the processing chain implemented for the production of daily LIC maps as part of the European Space Agency's Climate Change Initiative (CCI+) Lakes project.