Physiology-informed UAV multispectral estimation of tobacco leaf nitrogen nutrition index for variable nitrogen management

Introduction Spatial variability in crop nitrogen (N) status limits the effectiveness of uniform fertilization in tobacco, yet remotely sensed spectral signals cannot be directly translated into crop N demand. This study developed a physiology-informed unmanned aerial vehicle (UAV) multispectral framework linking leaf dry matter (LDM), leaf N concentration (LNC), the leaf nitrogen nutrition index (NNIleaf), and variable N management. Methods Field experiments involving seven N application rates (0–150 kg N ha −1 ) were conducted in Yunnan, China, during 2023–2024. A critical leaf-N relationship was established using ln(LDM) and LNC. Multispectral vegetation indices were screened using Pearson correlation analysis and recursive feature elimination. Partial least squares regression (PLSR), random forest (RF), and extreme gradient boosting (XGBoost) were compared for estimating ln(LDM) and LNC, from which NNIleaf was subsequently derived. Results The four seasonal critical points obtained in 2023 were described by N c,leaf = −3.156[ln(LDM)] 2 + 32.800ln(LDM) − 25.20 (R 2 = 0.992; RMSE = 0.978 mg g −1 ). XGBoost provided the best overall performance. In the independent 2024 validation, R² values for ln(LDM), LNC, and NNIleaf were 0.826, 0.783, and 0.832, respectively, with an NRMSE of 5.51% for NNIleaf. The fitted N-response relationship identified 82.5 kg N ha −1 as the NNIleaf-defined physiological N baseline (Nphys), corresponding to the lowest total N rate at which fitted NNIleaf reached 1. In a production-scale cross-year demonstration, the NNIleaf-based strategy recommended an average of 97.12 kg N ha −1 and was associated with slightly lower N input, 3.18% higher yield, and a 2.77-percentage-point reduction in yield coefficient of variation relative to conventional management. Discussion The framework provides a physiologically interpretable link between multispectral traits, leaf N sufficiency, and spatial N recommendations. Wider validation of the critical-N relationship and same-year fertilizer comparisons remain necessary.