PCLCD: Patch-Based Contrastive Learning for Remote Sensing Image Change Detection

Abstract

Current remote sensing change detection (CD) remains a challenging task due to severe pseudo-change interference and insufficient local feature modeling in complex scenarios. To address these issues, we propose a novel patch-based contrastive learning CD framework (PCLCD), which systematically enhances discriminative feature representation through bitemporal image pair analysis. The core innovation involves two modules: 1) a patch-based contrastive learning module that explicitly captures inter-patch similarity relationships by generating dense feature-domain patches. This module strengthens feature consistency within homogeneous regions via positive sample pairs while suppressing semantic confusion in nonchange areas through negative pair contrastion; and 2) a dynamic contextual aggregation module that adaptively fuses multiscale local patterns with global contextual information, enabling robust change discrimination across varying spatial scales. Extensive experiments on benchmark datasets demonstrate that PCLCD achieves state-of-the-art performance, with significant improvements on LEVIR-CD, WHU-CD, and SYSU-CD.