Fine-tuning with attention mechanisms for LULC classification: a case study of Tianjin, China

Land Use and Land Cover (LULC) classification based on remote sensing data remains a highly challenging task, particularly given the increasing availability of launched and in-orbit satellites. Higher spatial and spectral resolutions, together with a larger number of spectral channels, increase the potential of remote sensing data. Nevertheless, exploiting these multispectral characteristics to enhance image classification performance under small-sample conditions remains a considerable challenge. This study proposes a novel image classification network named Tuning-SeNet, which utilises ResNet50 as its backbone and integrates transfer learning with a channel attention mechanism. By leveraging open datasets and fusing multi-channel spectral information, the model extracts high-quality, semantically rich features. Applied to LULC classification tasks in Tianjin, China, the model demonstrates outstanding performance. Comprehensive evaluation shows that Tuning-SeNet outperforms traditional deep learning models in both classification accuracy and convergence speed. The incorporation of transfer learning and the attention mechanism enables the proposed model to achieve Overall Accuracy (OA) of 97.20%, F1-score of 97.10% and Kappa coefficient of 96.73% on the Tianjin LULC dataset. Comparative analysis with related studies using Sentinel-2 data validates the classification efficacy of the proposed model. These results underscore the significant practical potential of Tuning-SeNet for remote sensing-based LULC classification.