Quality-Aware Multi-Source Feature Fusion for Sentinel-2-Based Forestry Land-Cover Classification in Hunan Province

Accurate identification of vegetation types such as forest land, grassland, and orchard is essential for forestry resource monitoring and ecological management. This study proposes a quality-aware multi-source feature fusion method for provincial-scale forestry point-sample classification using Sentinel-2 growing-season composite imagery in Hunan Province. Multispectral bands, vegetation indices, observation quality features, topographic factors, and point-level phenological features are integrated to reduce category confusion caused by spectral similarity, scale inconsistency, and limited phenological representation in single-period imagery. A QGeo-Fusion model is developed to extract spectral-spatial features from Sentinel-2 patches, model point-level phenological information through an independent branch, and perform multi-source feature fusion for classification. Experimental results show that QGeo-Fusion achieves an overall accuracy of 0.9616, a Kappa coefficient of 0.9367, a Macro-F1 score of 0.9446, and a Weighted-F1 score of 0.9627 in the three-class task, outperforming Plain-CNN, ViT-fusion, and Swin-fusion. Ablation analysis indicates that point-level phenological features contribute most to performance improvement. The results demonstrate that integrating observation quality, topography, and phenological information can effectively improve forestry point-sample classification under single-period Sentinel-2 imagery.