Multi-stage Superpixel-guided Mamba-based Network for Change Detection

Abstract

Remote sensing image change detection relies on accurate modeling of multi-temporal features. Existing Mamba-based methods, however, often suffer from insufficient local information and redundant computations during feature aggregation, leading to inaccurate change boundary predictions. To address these issues, we introduce a superpixel mechanism into the Mamba-based framework and propose a Multi-stage Superpixel-guided Mamba Network for Change Detection (MSPMCD). In the encoder, a superpixel aggregation module extracts structurally consistent region-level features to enhance local and global representations, while a superpixel–pixel association matrix is constructed at the pre-event stage and updated at the post-event stage to preserve bi-temporal correlations. During feature fusion, superpixel-based regional constraints and feature proxy mechanisms enable region-level bi-temporal interaction. In addition, a superpixel consistency loss enforces similarity among pixel predictions within the same superpixel, ensuring intra-superpixel consistency and improving change boundary accuracy. Experiments on multiple public datasets demonstrate that MSPMCD achieves high F1 and IoU scores.

Publication
2026 IEEE International Symposium on Circuits and Systems
刘婧
刘婧
副教授,博导