segWCD: A new segmentation-based weak supervision neural network for building change detection

Yunyang Wu, Xiaobo Zhang, Xiaole Zhao, Yimin Sun, Tianrui Li

Abstract: Manual annotation of changes in high-resolution remote sensing images is labor-intensive and limits advancements in change detection. We introduce the Segmentation-based Weakly Supervised Change Detection (segWCD) framework to mitigate this challenge. Our method leverages a semantic segmentation model to generate pseudo-labels, offering weak supervision for detecting changes. The Creator module further refines these labels, enhancing the model’s detection accuracy. Additionally, we address the issue of label noise by variably weighting the pseudo-labels based on their confidence, thus optimizing the training process. Experimental results show that segWCD achieves a Recall of 0.921, an F1 score of 0.627, and an MIOU of 0.708, performing comparably to fully supervised methods. This approach marks a significant step forward in weakly supervised learning, demonstrating the potential of refined pseudo-labeling techniques.