Language-Guided Change Detection for high-resolution remote sensing imagery with limited labelled data

Deep learning has been extensively applied in the field of remote sensing for tasks such as change detection (CD). However, since CD is a pixel-level task, the high cost of data annotation and often limited availability of labelled data significantly restrict the performance of existing deep learning-based CD methods. To mitigate this problem, a novel Language-Guided Change Detection (LGCD) framework is introduced. Within this LGCD network, text information is leveraged to precisely locate changed areas, addressing the shortcomings associated with insufficient labelled image data. Also, augmentation semi-supervised learning techniques are employed to generate high-quality pseudo-labels, further reducing the reliance on labelled samples. Additionally, the utilisation of Fusion UNet (FUNet) and Transformer capitalises on their sensitivity to local and global features respectively, offering a comprehensive examination of change features in high-resolution bi-temporal remote sensing imagery. For evaluation purposes, three publicly available CD datasets are exploited. Experimental results demonstrate that the proposed LGCD framework achieves exceptional detection performance in both fully supervised and semi-supervised settings, despite the constraints of limited labelled data.