GAN-Based Sub-Instance Augmentation for Open-Pit Mine Change Detection in Remote Sensing Images

Remote sensing change detection (CD) for open-pit mines plays a critical role in both mineral development and environmental conservation. The performance of supervised deep-learning-based CD for open-pit mines is often limited by the amount of available data, necessitating data augmentation. Current CD data augmentation methods mainly focus on image-level, region-level and instance-level transformations. However, mining area changes occur at a finer sub-instance level, leading to suboptimal performance of existing methods. Therefore, this study proposes a novel data augmentation method named generative adversarial network-based sub-instance augmentation (GSIA). This method enables the generation of realistic and diverse CD samples using unchanged data from mining areas to address the issue of data scarcity in open-pit mine CD. GSIA comprises three steps. In the first two steps, GSIA achieves sub-instance-level transformations by sequentially applying GAN-based local editing to the labels and images of the mining areas. In the third step, GSIA constructs a synthetic CD dataset by randomly combining the bitemporal data. The effectiveness of GSIA was evaluated by comparing it with fourteen other data augmentation methods on five CD models, and GSIA outperformed all of them. In addition, training solely on synthetic data generated by GSIA achieved an overall accuracy of 97.64% and an F1-score of 78.65%, which were comparable to training with all available real data. Furthermore, GSIA can address the insufficient correlation between the training and test sets in domain adaptation. GSIA plays a significant guiding role in the data augmentation for open-pit mine CD.