DSA-UNet: Attention-Integrated and High-Frequency Feature-Preserved U-Net for Fuding White Tea Bud Segmentation

Tea image segmentation in tea gardens is essential for intelligent tea harvesting, but white tea bud segmentation remains challenging because buds are densely distributed, vary in scale, and resemble mature leaves. This study proposes DSA-UNet, an improved U-Net for Fuding white tea bud segmentation. The model integrates discrete wavelet transform (DWT), MaxPool-SE, and attention gate (AG) modules to preserve edge and high-frequency texture information, enhance tea bud-related channel responses, and suppress redundant shallow background features. A custom Fuding white tea image dataset was constructed. Images from 2024 autumn tea were used for training and basic evaluation, whereas 2025 autumn tea and 2026 spring tea images were used for cross-temporal generalization testing. On the 2024 autumn tea test set, DSA-UNet achieved a precision of 79.7%, recall of 82.0%, and Dice coefficient of 80.6%, outperforming the baseline U-Net by 4.6, 7.7, and 6.1 percentage points, respectively. It achieved Dice coefficient gains of 11.5 and 4.4 percentage points over the baseline U-Net on the 2025 autumn tea and 2026 spring tea cross-temporal generalization test sets, respectively. These results indicate that DSA-UNet provides more accurate and robust white tea bud segmentation under year-to-year and seasonal variations and can contribute to automated recognition, picking point localization, precision tea-garden management, and intelligent horticultural applications.