Influence of out-of-distribution examples on the quality of semantic segmentation in remote sensing

Semantic segmentation for land cover maps follows the closed world assumption, where each pixel must be classified into a set of predefined classes. In order to fulfill this assumption, an additional class is usually introduced to describe all areas not covered by the main classes, called "clutter" or "other". Consequently, this class is extremely heterogeneous, and the classification is usually subpar. Using a common approach for uncertainty assessment of land cover classification, we analyze the influence of the clutter class being present or absent during training on the semantic segmentation. We assess the model uncertainties of two different deep learning models, U-Net and DeepLab V3+, and different training configurations by using a Monte-Carlo dropout based uncertainty metric. The corresponding uncertainty maps and histograms show a correlation between clutter class and the uncertainty metric.