Characterization of Out-of-distribution Samples from Uncertainty Maps Using Supervised Machine Learning

The quality of land use maps often refers to the data quality, but distributional uncertainty between training and test data must also be considered. In order to address this uncertainty, we follow the strategy to detect out-of-distribution samples using uncertainty maps. Then, we use supervised machine learning to identify those samples. For the investigations, we use an uncertainty metric adapted from depth maps fusion and Monte-Carlo dropout based predicted probabilities. The results show a correlation between out-of-distribution samples, misclassifications and uncertainty. Thus, on the one hand, out-of-distribution samples are identifiable through uncertainty, on the other hand it is difficult to distinguish between misclassification, anomalies and out-of-distribution.