Classification of broad-leaved forest tree species based on HSFC-DeepLabV3+ in UAV multispectral images

This study focuses on the automatic classification of broad-leaved forest tree species in the Millennium Forest of Xiongan New Area using UAV-based multispectral imagery. To address the limitations of existing semantic segmentation models such as DeepLabV3+, we propose HSFC-DeepLabV3+, an enhanced model that significantly improves classification accuracy and efficiency. The model introduces four key enhancements: (1) EnhancedMobileNetV2 as the backbone with SEBlock integration for improved feature extraction; (2) SeparableASPP to reduce computational complexity while preserving semantic context; (3) an FPNFusion module for effective multi-scale feature fusion; and (4) a deeper classification head with multi-layer convolutions and Dropout to boost generalization. Additionally, a hybrid data augmentation strategy combining bilinear interpolation, geometric transformation, and Mixup is employed to enhance model robustness. Extensive experiments, including ablation studies and comparisons with state-of-the-art models (e.g., HRNet, PSPNet, and variants of DeepLabV3+), demonstrate the superiority of HSFC-DeepLabV3+. The model achieves improvements of 2.55% in overall accuracy (OA), 4.72% in mean pixel accuracy (MPA), and 9.32% in mean IoU (mIoU), while reducing total loss, validation loss, and training time by 0.065%, 0.076%, and 27.58 hours, respectively. Compared to DeepLabV3+ with MobileNetV2, our model also shows notable performance gains. These results highlight HSFC-DeepLabV3+ as a promising tool for high-precision, efficient remote sensing classification in forestry applications.