Answers are grounded in this paper and cite their sources.
Accurate delineation of individual tree crowns from UAV imagery is essential for fine-scale orchard monitoring but remains challenging when transferring models across orchards with differing illumination, canopy structures, and background conditions. Semi-supervised domain adaptation (SSDA) offers a label efficient solution, yet existing methods predominantly align global features while neglecting instance-level discrepancies critical for dense canopies, and conventional losses lack adequate boundary supervision. To address these limitations, we propose DFANet, a framework that integrates semi-supervised learning with dual level domain alignment and boundary-aware mask optimization. First, a Mean Teacher strategy is employed to exploit unlabeled target images through consistency regularization. Second, a Dual Level Adversarial Network aligns feature distributions at both the multi-scale feature level and the instance-level, reducing domain mismatch in dense canopy environments. Third, a Dual-Focus Dice Loss enhances mask accuracy by jointly optimizing interior regions and boundary bands, thereby improving crown separation in adhesion-prone areas. Experiments on three UAV datasets under six bidirectional transfer settings show that, with only 10% la beled target samples, DFANet achieves the strongest overall ro bustness among all compared supervised, SSL, UDA, and SSDA methods. Ablation studies confirm the complementary roles of dual-level alignment and boundary-focused optimization, while feature space visualizations further illustrate their contributions to cross-domain consistency. The results indicate that DFANet provides an effective and label-efficient solution for deploying crown segmentation models across heterogeneous orchard environments.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: