A Robust Co-training Framework to Handle Noisy Labels for Remote Sensing Image Segmentation

Effective learning with noisy labels is crucial in remote sensing data analysis, not only for instance-wise prediction but also for pixel-wise prediction. The recent surge in demand for robust machine learning models emphasizes the necessity of constructing large datasets and managing associated costs carefully. This trend highlights the potential of utilizing coarsely annotated datasets, which commonly entail the risk of producing noisy labels. To address label noise, methods of learning with noisy labels have been proposed for instance-wise classification assessed in computer vision. However, previous studies have not provided a detailed discussion on how to detect human error at the pixel level and how to adapt the resulting error maps within a single end-to-end learning framework in remote sensing. To address these issues, we propose a new co-training framework for segmentation modeling with automatic filtering to reduce the negative effects caused by noisy labels in the training dataset. Our method has three key contributions: (1) a trainable filter module to identify uncertain pixels at each iteration, (2) a weighted multi-loss strategy to effectively filter out uncertain training loss, and (3) a collaborative filter update based on outputs from co-trained networks. To assess the effectiveness of our method, we conducted experiments on the DeepGlobe landcover classification dataset, a self-constructed mangrove dataset, and the Massachusetts building dataset. We compared the performance of our method to several state-o-fthe-art noise-robust methods. The experimental results reveal that our proposed method outperforms the alternatives and demonstrates greater effectiveness on multi-class segmentation tasks.