Improved Semantic Segmentation Network Integrating ResNet-18 and FSCF Modules for Enhanced Transformer Encoder Efficiency

Semantic segmentation is vital in tasks like land cover classification and cultural heritage surveys. While deep learning advancements have popularized CNNs and Transformers for this purpose, challenges persist in balancing efficiency and accuracy in high-resolution remote sensing image processing. This study introduces an enhanced method combining ResNet-18, FSCF modules, and a GLTCB decoder to optimize feature extraction and fusion, improving overall performance. Experiments on UAVid and Vaihingen datasets show our method's superior accuracy, especially in complex scenes and object boundaries, outperforming traditional methods in metrics like Mean F1 and mIoU, proving its effectiveness in high-resolution remote sensing and urban environments.