A Mutual Teaching Framework with Momentum Correction for Unsupervised Hyperspectral Image Change Detection

Abstract

Deep-learning methods rely on massive labeled data, which has become one of the main impediments in hyperspectral image change detection (HSI-CD). To resolve this problem, pseudo-labels generated by traditional methods are widely used to drive model learning. In this paper, we propose a mutual teaching approach with momentum correction for unsupervised HSI-CD to cope with noise in pseudo-labels, which is harmful for model training. First, we adopt two structurally identical models simultaneously, allowing them to select high-confidence samples for each other to suppress self-confidence bias, and continuously update pseudo-labels during iterations to fine-tune the models. Furthermore, a new group confidence-based sample filtering method is designed to obtain reliable training samples for HSI. This method considers both the quality and diversity of the selected samples by determining the confidence of each group instead of single instances. Finally, to better extract the spatial–temporal spectral features of bitemporal HSIs, a 3D convolutional neural network (3DCNN) is designed as an HSI-CD classifier and the basic network of our framework. Due to mutual teaching and dynamic label learning, pseudo-labels can be continuously updated and refined in iterations, and thus, the proposed method can achieve a better performance compared with those with fixed pseudo-labels. Experimental results on several HSI datasets demonstrate the effectiveness of our method.

AIO2: Online Correction of Object Labels for Deep Learning With Incomplete Annotation in Remote Sensing Image Segmentation

“…They also combine it with regularization constraining on original noisy labels, a popular strategy aiming to reduce the negative effects of mistaken corrected labels [18]. Besides, Sun et al utilized mutual teaching with two structurally identical models to update noisy pseudo labels for hyperspectral image change detection [67]. The aforementioned approaches for RS image segmentation with noisy labels primarily rely on pixel-wise correction, with some adjustments to enhance adaptability to RS images.…”