Answers are grounded in this paper and cite their sources.
Validation study demonstrates that a canopy-corrected drone framework accurately estimates cotton boll density under foliage occlusion, highlighting practical organ-scale crop phenotyping.
Accurate estimation of yield-contributing opened cotton boll density from UAV imagery is limited by canopy occlusion, because nadir-view detection models can only identify opened bolls exposed on the canopy surface. To address this problem, this study developed an interpretable correction framework that converts visible opened boll density into yield-contributing opened boll density by incorporating Plant Area Index (PAI) as a canopy-structure proxy. The framework integrates a position-specific lightweight YOLOv8n detector, a sliding-window and deduplication-based plot-scale counting pipeline, and a PAI-based correction model. Field data collected from 2023 to 2025 were used, with the 2024 dataset serving for detector training, counting-pipeline development, and correction-model calibration, while the 2023 and 2025 datasets were reserved for independent cross-year validation. The improved detector achieved an mAP@0.50:0.95 of 0.789, with 2.15 million parameters and an inference time of 2.90 ms. In the 2024 calibration dataset, the PAI-integrated correction model reduced the estimation error from 8.26% using visible opened boll density alone to 3.48%, with an RMSE of 5.09 bolls·m −2 . In independent validation across 102 plots from 2023 and 2025, the framework achieved an overall MAPE of 2.24%, with treatment-level errors ranging from 0.20% to 8.58%. The resolution sensitivity analysis indicated that the framework remained stable when the spatial resolution was finer than 0.37 cm·pixel −1 . These results demonstrate that explicitly modeling canopy occlusion with PAI enables a robust and interpretable conversion from visible opened boll counts to yield-contributing opened boll density estimates, providing a practical pathway for UAV-based organ-scale cotton phenotyping and yield-component assessment.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: