Abstract
Accurate estimation of leaf nitrogen content (LNC) is critical for optimizing fertilization strategies in greenhouse tomato production. This study developed a robust hyperspectral-based framework for non-destructive LNC prediction by combining advanced spectral preprocessing, feature selection, and machine learning. Hyperspectral reflectance data were collected across five nitrogen and irrigation treatments over key growth stages. Signal quality was enhanced through Savitzky-Golay smoothing (SG) and Standard Normal Variate normalization (SNV). Key nitrogen-sensitive wavelengths-centered around 725 nm and 730 - 780 nm-were identified using Competitive Adaptive Reweighted Sampling (CARS) and Principal Component Analysis (PCA). Four predictive models were compared, among which a hybrid Stacked Autoencoder-Feedforward Neural Network (SAE-FNN) achieved the highest accuracy (test R² = 0.77, RPD = 2.06), effectively capturing nonlinear spectral-nitrogen interactions. In contrast, Support Vector Machine (SVM) exhibited overfitting and Partial Least Squares Method (PLSR) underperformed due to its linear constraints. These results underscore the potential of integrating hyperspectral sensing with deep learning for intelligent nitrogen monitoring in controlled-environment agriculture.