A Hierarchical Feature Assembly and Adaptive Enhancement Network for Remote Sensing Image Change Detection

Abstract

Change detection (CD) in remote sensing images plays a critical role in urban monitoring, land-use analysis, and disaster assessment. However, existing methods still struggle with insufficient local structure representation, inconsistent deep semantics, and ineffective cross-layer feature fusion, leading to false detections in complex scenes. To address these issues, we propose AssemCD, a hierarchical feature assembly network with adaptive enhancement for remote sensing image CD. Specifically, a Hierarchical Context Assembly Module (HCAM) is designed to integrate local convolutional representation with linear context modeling, enabling unified local-global feature learning. An Adaptive Dual-path Module (ADM) is further introduced to enhance directional consistency in shallow layers and improve multiscale local structure representation. In addition, a Semantic Recalibration Fusion Module (SRFM) is employed to adaptively recalibrate deep features via lightweight global semantic guidance, thereby improving semantic consistency during cross-layer fusion. Experiments on the LEVIR-CD, WHU-CD, and SYSU-CD datasets demonstrate that AssemCD achieves competitive CD performance. On LEVIR-CD, AssemCD obtains a changed-class IoU of 70.82% and an F1 score of 82.92%, achieving the highest IoU, F1 score, and Kappa coefficient among the compared methods. On WHU-CD, it achieves a changed-class IoU of 63.79% and an F1 score of 77.89%, demonstrating competitive foreground discrimination performance. Under the fixed 25% SYSU-CD sampling protocol, AssemCD achieves a changed-class IoU of 62.73%, an F1 score of 77.09%, and a Kappa coefficient of 70.29%. Ablation experiments further verified the effectiveness and complementarity of HCAM, ADM, and SRFM in hierarchical feature modeling. The complexity analysis indicates that AssemCD is not designed to minimize the parameter count; instead, it achieves competitive detection accuracy under a moderate computational budget.