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

Nitrogen plays a pivotal role in determining cotton yield and fiber quality. Nevertheless, because high-dimensional remote-sensing data are inherently complex and redundant, accurately estimating cotton plant nitrogen concentration (PNC) from unmanned aerial vehicle (UAV) imagery remains problematic, which in turn constrains both model precision and transferability.

Accordingly, this study introduces a hierarchical feature-selection scheme combining Elastic Net and Boruta-SHAP to eliminate redundant remote-sensing variables and evaluates six machine-learning algorithms to pinpoint the optimal method for estimating cotton nitrogen status.

Our findings reveal that five critical features (Mean_B, Mean_R, NDRE_GOSAVI, NDVI, GRVI) markedly enhanced model performance. Among the tested algorithms, random forest achieved superior performance (R² = 0.97-0.98; RMSE = 0.05-0.08), exceeding all alternatives. Both in-field observations and model outputs demonstrate that cotton PNC consistently decreases throughout development, but optimal conditions of 450 mm irrigation and 300 kg N ha⁻¹ sustain relatively elevated nitrogen levels.

Collectively, the study provides robust guidance for precision nitrogen management in cotton production within arid regions.