Cotton boll extraction and single-boll weight estimation based on UAV multispectral imagery

Area of Science:

• Agricultural Science
• Remote Sensing
• Machine Learning

Background:

• Estimating single-boll weight (SBW) in cotton is challenging post-defoliation due to spectral interference from lint, soil, and senescent leaves.
• Existing methods struggle with background noise, limiting accurate spatial estimation of SBW.

Purpose of the Study:

• To develop and validate a Unmanned Aerial Vehicle (UAV) multispectral workflow for improved single-boll weight (SBW) mapping in cotton.
• To reduce background interference and enhance the transferability of SBW estimation models.

Main Methods:

• A workflow integrating object-based boll extraction (using maximum likelihood classification), spectral feature selection (Pearson correlation and SHAP), and machine learning regression (ridge, random forest, neural network) was employed.
• Data were collected from a two-year cotton experiment with varying varieties and planting densities.
• Maximum likelihood classification achieved >97% accuracy for boll extraction.

Main Results:

• The optimal model, combining maximum likelihood boll extraction with neural network regression, achieved a coefficient of determination of 0.80 and RMSE of 0.31 g.
• Vegetation indices from red, red-edge, and near-infrared bands, especially those mitigating soil effects, showed strong SBW correlation.
• Relative errors were consistently below 15% across different experimental conditions.

Conclusions:

• The proposed UAV multispectral workflow effectively reduces background interference for accurate SBW spatial estimation.
• This approach enables transferable SBW mapping crucial for cotton breeding programs, density optimization, and harvest management.