Learning from Noisy Labels in Remote Sensing
Deep learning (DL) based methods have recently seen a rise in popularity in the context of remote sensing (RS) image classification. Most DL models require huge amounts of annotated images during training to optimize all parameters and reach a high-performance during evaluation. The availability and quality of such data determine the feasibility of many DL models. However, annotating RS images with multi-labels at large-scale to drive DL studies is time consuming, complex, and costly in operational scenarios. To address this issue, existing thematic products (e.g., Corine Land-Cover map) can be used, however the land-use and land-cover labels through these products can be incomplete and noisy. Handling data with incomplete and noisy labels may result in uncertainty in the classification model and thus may lead to a reduced performance on label prediction.
We have recently started to research and develop noise robust DL models to reduce the negative impact of noisy land-use and land-cover annotations at the Remote Sensing Image Analysis (RSiM) group, TU Berlin.