Texas A&M drone AI predicts cotton nitrogen needs with 88% accuracy

A team led by Vaishali Swaminathan at Texas A&M published a new study on arXiv (2608.07801) demonstrating a machine-learning pipeline for precision nitrogen management in cotton. Using three years of UAV multispectral imagery collected between early vegetative growth and flowering, the researchers combined spectral vegetation indices with morphological features—plant height and fractional canopy cover—to train several regression models. The goal was to estimate dry biomass weight, plant nitrogen uptake, plant nitrogen concentration, critical nitrogen dilution, and the nitrogen nutrition index (NNI), all critical for deciding when and how much fertilizer to apply.

The strongest results came from decision-tree ensembles. Random forest regression achieved the best overall accuracy in trial-held-out validation, with R²=0.88 for dry biomass weight (MAPE 23.14%) and R²=0.84 for nitrogen uptake. Extreme gradient boosting (XGBoost) performed comparably for biomass (R²=0.87) and produced the most accurate nitrogen nutrition index for identifying nitrogen-deficient plots and categorizing stress levels. Models were validated using both trial-held-out and leave-one-year-out schemes, ensuring robustness across different growing seasons and field conditions. The study highlights how fusing spectral and morphological data with scalable ML models can bring affordable precision agriculture to cotton farming, reducing excess fertilizer use while maintaining yield.