SEMI-SUPERVISED REMOTE SENSING IMAGE CHANGE DETECTION USING MEAN TEACHER MODEL FOR CONSTRUCTING PSEUDO-LABELS

In recent years, deep learning has ushered in great developments in remote sensing image change detection. Practically, it is labor-intensive and time-consuming to label images for co-registration. In this paper, we propose a semi-supervised training that uses the mean teacher model to construct pseudo-labels to increase the generalizability of the model trained with a handful of data supervision. More specifically, we first supervise the training of a student change detection model with a few labeled data, while migrating the model parameters from each training round to a teacher model with the same structure through Exponential Moving Average (EMA). The weakly augmented output produced by the teacher model was preferable to the strongly augmented prediction produced by the student to penalize the latter. We explore the paradigm of Strong-to-Weak Consistency in change detection. Experiments on the LEVIR-CD and WHU-CD have been extensively conducted and state-of-the-art performance has been achieved.