Core Concepts
A novel change detection network, LRNet, is proposed based on a localization-then-refinement strategy to accurately discriminate change areas and their boundaries in high-resolution remote sensing imagery.
Abstract
The paper presents a novel change detection network, LRNet, which consists of two stages: localization and refinement.
Localization Stage:
LRNet employs a three-branch encoder to simultaneously extract features from the original bi-temporal images and their differences.
Learnable Optimal Pooling (LOP) is proposed to replace the widely used max-pooling, reducing information loss during feature extraction.
Change Alignment Attention (C2A) and Hierarchical Change Alignment (HCA) modules are designed to effectively interact features from different branches and accurately localize change areas of various sizes.
Refinement Stage:
The Edge-Area Alignment (E2A) module is proposed to constrain and refine the change areas and edges obtained from the localization stage.
The decoder in the refinement stage combines the C2A-enhanced differential features to progressively refine the change areas and edges from deep to shallow.
A hybrid loss function of Binary Cross-Entropy (BCE) and Intersection over Union (IOU) is adopted for supervised training optimization.
The proposed LRNet outperforms 13 other state-of-the-art methods in terms of comprehensive evaluation metrics and provides the most accurate boundary discrimination results on the LEVIR-CD and WHU-CD datasets.
Stats
The LEVIR-CD dataset consists of 637 pairs of high-resolution remote sensing images, each with a size of 10241024 and a spatial resolution of 0.5 m/pixel. The WHU-CD dataset comprises one pair of high-resolution remote sensing images, each with a size of 32,50715,354 and a spatial resolution of 0.3 m/pixel.