Boundary-Aware Residual Attention U-Net with Spectral Index Fusion for Land Use/Land Cover Segmentation of Sentinel-2 Imagery

Accurate land use and land cover (LULC) mapping from medium-resolution satellite imagery remains challenging due to spectral similarity among classes and the difficulty of delineating complex class boundaries. This study presents a Boundary-Aware Residual Attention U-Net framework for pixel-wise semantic segmentation of 30 m Sentinel-2 imagery. The proposed model integrates residual learning and attention-gated skip connections with two enhancements: on-the-fly spectral index fusion using the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI), and a morphological boundary-weighted focal loss designed to improve classification performance along class transition zones.

The framework was developed and evaluated over Sylhet District, Bangladesh, using Sentinel-2 Surface Reflectance imagery and ESRI Global Land Cover data for 2023. Seven land cover classes were considered: Water, Trees, Flooded Vegetation, Crops, Built Area, Bare Ground, and Rangeland. To address class imbalance, inverse-frequency class weighting was incorporated into the loss function. A patch-based training strategy with 50% overlap was employed to increase sample diversity, and five ablation configurations were evaluated to quantify the contributions of spectral index fusion and boundary-aware supervision.

The best-performing model (V5, λ = 2.0) achieved an overall accuracy of 91.69%, a weighted IoU of 85.40%, and a mean F1-score of 84.86%. Boundary-aware evaluation yielded a BF Score of 0.6571 and a Boundary IoU of 48.93%. Compared with the baseline model, the proposed approach improved mIoU by 4.84%, Boundary IoU by 4.17 percentage points, and BF Score by 0.0387. The largest class-level improvements were observed for Rangeland and Flooded Vegetation. These results demonstrate that integrating spectral index fusion with boundary-aware optimization enhances both classification accuracy and boundary delineation quality in medium-resolution satellite image segmentation.