Land-Cover Classification of Sentinel-2 Imagery Using a Hybrid Swin-Style Transformer and SegFormer-Inspired Network: A Case Study of Hawija, Iraq

Mapping land cover helps with the management of environmental, agricultural and urban areas; however, it is challenging in semi-arid regions, where several classes have similar spectral responses. The research offers a combination of new SWIN with SegFormer network for classifying five land-cover classes occurring in Hawija District, Iraq, using ten Sentinel-2 Level-2A spectral bands and eight spectral indices. As many as 111,740 labeled samples were extracted as 11 × 11-pixel patches with 18 input channels, and were divided by an 80:20 stratified random split into 80,452 training and 22,348 test samples. The data was divided into 80,452 training samples (72%), 8940 validation samples (8%), and 22,348 test samples (20%). The validation subset was used to monitor the training, early stopping, and checkpoint selection, while the test subset was only used for a performance evaluation at the end. The three-way partitioning protocol was adopted for the HSSNet Swin-style Transformer and SegFormer-like baseline. HSSNet is a combination of Swin-inspired attention with SegFormer-inspired feature processing, gated cross-attention fusion and squeeze-and-excitation channel recalibration. HSSNet obtained an overall accuracy (OA) of 96.18%, an average accuracy (AA) of 95.47% and a kappa coefficient of 0.9434, compared with 90.80%, 83.39% and 0.8614 for the Swin Transformer baseline and 93.68%, 90.73% and 0.9062 for the SegFormer-like baseline. The results indicate that hybrid feature fusion can improve land-cover discrimination in spectrally heterogeneous semi-arid landscapes, although spatially independent validation is still required to confirm the generalization of these findings.