Land cover classification in high-resolution remote sensing: using Swin Transformer deep learning with texture features

The enhancement of spatial resolution has improved the spatial information conveyed by remote sensing images. Nevertheless, the impact of spectral-texture combined features on classification accuracy in deep learning remains understudied. This study develops five spectral texture features (Spe-GLCM, Spe-FRS, Spe-Gabor, Spe-CS-LMP, and Spe-FD) using GID15 and Vaihingen datasets. These features were compared with U-Net++, Deeplabv3+, and PSPNet models, incorporating fine-tuned model structures. The results indicate that integrating texture features enhances accuracy in both overall and sparsely sampled land cover classifications. This highlights the potential of deep learning combined with texture features to recognize land cover from high-resolution remote sensing imagery.