Going Beyond One-Hot Encoding in Classification: Can Human Uncertainty Improve Model Performance in Earth Observation?

Abstract

Technological and computational advances continuously drive forward the field of deep learning in remote sensing. In recent years, the derivation of quantities describing the uncertainty in the prediction, which naturally accompanies the modeling process, has sparked interest in the remote sensing community. Often neglected in the machine learning setting is the human uncertainty that influences numerous labeling processes. As the core of this work, the task of local climate zone (LCZ) classification is studied by means of a dataset that contains multiple label votes by domain experts for each image. The inherent label uncertainty describes the ambiguity among the domain experts and is explicitly embedded into the training process via distributional labels. We show that incorporating the label uncertainty helps the model to generalize better to the test data and increases model performance. Similar to existing calibration methods, the distributional labels lead to better-calibrated probabilities, which in turn yield more certain and trustworthy predictions.