Cotton Boll Extraction and Single-Boll Weight Estimation Based on UAV Multispectral Imagery

Maoguang Chen¹Caixia Yin¹Na Su¹TAO Lin²Xiuliang Jin³Fengquan Wu¹Pingan Jiang¹Qiuxiang Tang^1*

• ¹Engineering Research Centre of Cotton, Ministry of Education / College of Agriculture,, Xinjiang Agricultural University,, Urumqi, China
• ²Institute of Cash Crops, Xinjiang Academy Sciences, Key Laboratory of Crop Physiology, Ecology and Cultivation in Desert Oasis, Ministry of Agriculture and Rural Affairs, Urumqi 830091, China, Urumqi, China
• ³Key Laboratory of Crop Physiology and Ecology, Institute of Crop Sciences, Chinese Academy of Agricultural Sciences, Ministry of Agriculture, Beijing 100081, China, Urumqi, China

Single-boll weight (SBW) is difficult to estimate after defoliant application because canopy spectra include numerous mixed pixels from lint, soil, and senescent leaves, leading to strong background interference. Here we propose a UAV multispectral workflow that combines object-based boll extraction, spectral feature selection, and machine-learning regression to improve SBW mapping. Data were collected from a two-year drip-irrigated cotton experiment in Xinjiang, China involving four varieties evaluated under five planting densities treatments. Boll extraction was treated as a supervised object-based classification problem, and maximum likelihood, mahalanobis distance, and parallelepiped classifiers were compared. Fifteen vegetation indices were computed from the extracted boll pixels; informative features were identified using Pearson correlation and SHapley Additive exPlanations importance ranking. SBW was then estimated with ridge regression, random forest regression, and neural network regression using an independent validation dataset. Maximum likelihood consistently achieved overall accuracy above 97% with Kappa values above 0.93, outperforming the other classifiers. Indices derived from the red, red-edge, and near-infrared bands, particularly those designed to reduce soil background effects, showed the strongest relationships with SBW and ranked highest in SHAP. The best-performing model, which integrated maximum likelihood-based boll extraction with neural network regression, achieved a coefficient of determination of 0.80 and a root mean square error of 0.31 g on the validation set. Relative errors remained below 15% across different years, varieties, and planting densities. This workflow reduces background interference and enables transferable SBW spatial estimation for breeding evaluation and density and harvest management.