.png){kind=icon}
.png){kind=icon}
Return
Structure-Aware and Alignment-Optimized Remote Sensing Image Segmentation
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2026.3670335.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
High-resolution remote sensing image segmentation remains challenging due to spatial misalignment, intra-class variability, and blurred object boundaries. These issues are particularly severe for small-scale or elongated ground objects, and are further exacerbated by the limited availability of pixel-level annotations, which increases the difficulty of model training. To address these challenges, we propose HSGNet, a structure-aware and alignment-guided segmentation network. The framework follows a progressive pipeline: multi-scale encoding to extract semantic priors, pixel-wise displacement estimation for geometric correction, edge-aware enhancement combined with gradient consistency constraints to refine boundary representation, and confidence-filtered pseudo-labeling with boundary-sensitive regularization for stable semi-supervised learning. Experiments on the ISPRS Potsdam dataset, along with cross domain validation on the Vaihingen dataset, demonstrate that HSGNet consistently outperforms state-of-the-art methods in terms of pixel accuracy (Pixel Acc), mean Intersection-over-Union (mIoU), F1-score, and boundary F1-score (Boundary-F1).
Keywords:
Remote sensing image semantic segmentation
structure-aware modeling
spatial alignment optimization
weakly supervised learning
boundary refinement
consistency constraints