Uncertainty-driven semi-supervised learning for phenology-aware apple flower counting

• Dual-stage framework proposed for robust multi-phase apple flower counting. • A novel Uncertainty-driven SSL with dynamic pseudo-label filtering was proposed. • The first UAV-based multi-phase apple flower dataset constructed for validation. • Experiments demonstrated its superiority in multi-phase apple flower counting. UAV-based flower counting methods struggle with high-density orchard occlusion, costly pixel-level annotation demands, and phenological phase confusion. To address these limitations, this study proposes a dual-stage framework for apple flower counting and phenological phase identification. The first stage employs color-threshold segmentation to extract floral clusters from complex backgrounds. The second stage introduces an uncertainty-driven semi-supervised phenology-aware apple flower counting network (SPAF-Net), based on the mean-teacher framework, integrating three key innovations: an uncertainty-driven pseudo-label filtering (UPF) for noise suppression in unlabeled training, uncertainty-adaptive anti-overfitting masking (UAAM) for generalization enhancement in labeled training, and a phenological phase mapper (PPM) with channel mutual exclusion loss to enhance inter-phase discriminability. Evaluations on our custom floral cluster dataset (9012 images, 23,013 manual annotations) demonstrated that the proposed dual-stage framework significantly outperforms single-stage methods in both speed and accuracy, while SPAF-Net achieves lower counting error compared to state-of-the-art semi-supervised models. The results establish the proposed methodology as an effective semi-supervised solution for apple flower counting, providing methodological support for digital management of orchard flowering.