Content
Nitrogen is essential for plant growth, constituting a critical component of proteins, chlorophyll, and nucleic acids. Its concentration within leaves serves as an important indicator of photosynthetic capacity and overall growth potential. Traditional nitrogen measurement methods rely on destructive sampling combined with laboratory chemical analysis, which are time-consuming and expensive. In contrast, hyperspectral sensing offers a non-destructive alternative by linking biochemical properties related to nitrogen with specific spectral absorption features. However, existing methods face challenges: empirical models require extensive field data and often perform poorly outside their training environments, while physically based models, though more transferable, struggle with complex inversion problems. Hybrid methods attempt to merge these approaches but commonly suffer from “domain shift” issues where simulated spectra used for training do not accurately reflect real-world measurements.
A recent study published in Plant Phenomics on October 10, 2025, by Daoliang Li and Kang Yu’s team from China Agricultural University and the Precision Agriculture Lab presents a novel solution for rapid, non-destructive monitoring of leaf nitrogen content (LNC) across multiple crop species. Their approach integrates plant radiative transfer theory with deep learning and hyperspectral reflectance data, aiming to enhance both reliability and transferability of nitrogen assessment. The methodology involves processing simulated directional–hemispherical reflectance (DHRF) spectra and measured bidirectional reflectance factor (BRF) spectra using continuous wavelet transform (CWT) and first derivatives (FD). These spectral transformations reduce discrepancies caused by specular reflections and domain shifts, especially in the visible and near-infrared regions, enhancing spectral comparability and highlighting key nitrogen-related absorption features.
Using these transformed spectra, the research team first applied parametric regression models incorporating 30 vegetation indices (VIs). When trained on a comprehensive simulated dataset, several VIs derived from nitrogen allocation models (such as GARI, GNDVI, GRVI, and CI800,550) exhibited moderate accuracy, while others, particularly those relying on protein-to-nitrogen conversion formulations, performed poorly. The accuracy of leaf nitrogen estimation improved notably when the models were recalibrated using a more representative subset of simulated samples (the T100 dataset). For instance, the vegetation index SR708,775 achieved a root mean square error (RMSE) of 0.303 g/m² and an R² of 0.494, underscoring the importance of sample representativeness over sheer dataset size in parametric modeling approaches.
The study further explored non-parametric hybrid methods by combining machine learning and deep learning models with spectral transformations. Deep learning models, especially the Conv-Transformer architecture, outperformed traditional machine learning techniques and physically based inversions across the full simulated dataset. Training on the T100 subset further enhanced the Conv-Transformer’s accuracy, reducing RMSE to 0.247 g/m² and raising R² to 0.665. Ablation studies and cross-crop validations revealed that both the spectral similarity-based sample selection strategy and the modified Transformer architecture synergistically contributed to these performance gains.
Importantly, the framework demonstrated consistent improvements in LNC prediction accuracy and robustness across diverse crops, including maize, wheat, rice, and sorghum. This indicates the method’s ability to mitigate domain shift effectively, offering accurate nitrogen estimation even in data-scarce scenarios without requiring costly field calibrations. Given its reliance on leaf-scale bidirectional reflectance—more practical than integrating-sphere measurements—the approach is well suited for routine agricultural monitoring and facilitates technology transfer.
Overall, this integration of plant physics, deep learning, and hyperspectral data represents a significant advancement for precision agriculture. Timely and accurate nitrogen diagnostics can enable optimized fertilization strategies that enhance crop yields while minimizing environmental pollution. The study’s findings pave the way for scalable, cost-effective, and transferable nitrogen monitoring tools critical for sustainable agriculture.
