UAV-based multitier feature selection improves nitrogen content estimation in arid-region cotton.

Area of Science:

• Agricultural Science
• Remote Sensing
• Data Science

Background:

• Nitrogen is crucial for cotton yield and fiber quality.
• Estimating cotton plant nitrogen concentration (PNC) from UAV imagery is challenging due to complex, redundant remote-sensing data.
• This limits model precision and transferability for precision agriculture.

Purpose of the Study:

• To develop an accurate and transferable method for estimating cotton PNC using UAV remote-sensing data.
• To identify optimal machine-learning algorithms and critical spectral features for PNC estimation.
• To provide guidance for precision nitrogen management in cotton production.

Main Methods:

• A hierarchical feature-selection scheme combining Elastic Net and Boruta-SHAP was employed to reduce data dimensionality.
• Six machine-learning algorithms were evaluated for their performance in estimating cotton PNC.
• In-field observations were used to validate model outputs and assess nitrogen dynamics.

Main Results:

• Five critical remote-sensing features (Mean_B, Mean_R, NDRE_GOSAVI, NDVI, GRVI) significantly improved model performance.
• Random forest algorithm demonstrated superior performance with R² values of 0.97-0.98 and RMSE of 0.05-0.08.
• Cotton PNC decreased throughout development, with optimal irrigation and nitrogen fertilization sustaining higher levels.

Conclusions:

• The study successfully enhanced the accuracy and transferability of cotton PNC estimation from UAV imagery.
• The findings support the use of random forest and selected spectral indices for precision nitrogen management.
• This research offers practical guidance for optimizing cotton production in arid environments through data-driven nitrogen strategies.