Low-annotation apple flower counting: A color-SAM enhanced and uncertainty-guided semi-supervised framework.

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Accurate flower-load assessment is crucial for orchard management, but traditional methods are labor-intensive.
• UAV-based deep learning offers efficiency but requires extensive annotated data, which is scarce and costly.
• Semi-supervised learning (SSL) reduces annotation needs but struggles with complex orchard imagery due to background noise and small targets, impacting pseudo-label reliability.

Purpose of the Study:

• To develop a robust two-stage framework for accurate, stage-specific apple flower counting with limited labeled data.
• To address the challenges of complex backgrounds, small target sizes, and unreliable pseudo-labels in UAV-based orchard imagery.
• To provide methodological support for rapid and precise apple flower counting in low-annotation agricultural scenarios.

Main Methods:

• A color-SAM flower extractor (CSAM-FE) was developed, combining color thresholding and the Segment Anything Model (SAM) for background suppression and high-quality flower cluster extraction.
• An uncertainty-guided semi-supervised flower counting network (USCount-Net) was proposed, featuring adaptive pseudo-label filtering (PLF) using frequent forward uncertainty estimation (FFUE) to mitigate error propagation.
• A noise-sensitive adaptive gated fusion (AGF) module was introduced to effectively fuse cross-scale features, addressing scale variations from different phenological stages and observation angles.

Main Results:

• The proposed CSAM-FE effectively preprocesses images, extracting purified flower clusters for the counting network.
• USCount-Net demonstrated superior performance compared to state-of-the-art methods across 10%, 30%, and 50% labeling ratios, achieving lower Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).
• The framework proved effective in handling scale variations and complex backgrounds inherent in UAV orchard imagery.

Conclusions:

• The developed two-stage framework offers a significant advancement in automated flower counting for precision agriculture.
• The methodology provides a viable solution for accurate apple flower counting in scenarios with limited annotated data.
• This research supports the development of more efficient and data-driven thinning strategies in orchard management.