Integrating UAV multi-sensor imagery and machine learning for non-destructive estimation of nitrogen nutrition index in tea plants

• UAV multispectral, RGB and LiDAR improves tea NNI estimation accuracy. • Boruta-selected feature subsets outperform those from RFE and ElasticNet. • GPR model using Boruta-selected MS+RGB features achieves the highest accuracy. Accurate diagnosis of nitrogen status is globally crucial for precise fertilization and optimal management in tea plantations. The Nitrogen Nutrition Index (NNI) provides a robust quantitative measure of N sufficiency, yet its practical application in tea plants has been limited due to the lack of established critical nitrogen dilution curves (CNDC) and validated non‑destructive monitoring methods. To address this, this study developed a novel UAV‑based multi‑source remote sensing approach for NNI estimation in tea plants, innovatively combining LiDAR with multispectral (MS) and visible‑light (RGB) data to capture complementary structural and spectral information. Field experiments involving three tea varieties under six N treatments were conducted, and NNI estimation models were built by integrating three feature selection methods (Boruta, RFE, ElasticNet) with four machine learning algorithms (SVM, PLS, RF, GPR). The results demonstrated a significant nitrogen fertilization effect, with NNI values ranging from 0.36 to 1.16, increasing with nitrogen application. Single-source remote sensing features selected by the three algorithms exhibited the performance ranking of MS > RGB > LiDAR. Multi‑sensor feature fusion significantly outperformed single‑sensor models. The optimal model, integrating Boruta-selected MS+RGB features with GPR, achieved the highest accuracy (R 2 =0.7346, RMSE=0.0878, RPD=1.6987). This model provides an accurate and non‑destructive tool for field‑scale monitoring of tea nitrogen status, supporting precision fertilization that can reduce fertilizer application in tea cultivation.