Abstract
Multiclass change detection in remote sensing images plays a vital role in remote sensing applications. However, the existing methods still have the problem of subtle changes missed. In this paper, we propose a model named PM-MCD, which consists of a VMamba-based pyramid feature extraction encoder for remote sensing images and a multi-scale information aggregation decoder based on MLP and MSSC module, enabling efficient multiclass change detection in remote sensing images. In addition, we propose a multi-scale attention fusion module, MSSC, to enhance the model's ability to recognize small-scale change regions. Experimental results show that, on the WHU-CD, Landsat-SCD, and CNAM-CD datasets, our model outperforms existing CNN- and Transformer-based methods, achieving 99.4/96.77/90.86% overall accuracy (OA), 90.18/82.27/68.50% mean intersection over union (mIoU), and 91.44/89.88/79.86% F1 scores.