By combining plant radiative transfer theory with deep learning and hyperspectral reflectance data, the approach significantly improves the reliability and transferability of nitrogen assessment across multiple crop species.
Nitrogen is a fundamental nutrient for plants, forming the backbone of proteins, chlorophyll, and nucleic acids. Its concentration in leaves directly reflects photosynthetic capacity and growth potential. Conventional nitrogen measurements depend on destructive sampling and laboratory chemical analysis, which are costly and time-consuming. Hyperspectral sensing provides a non-destructive alternative by linking nitrogen-related biochemical properties to spectral absorption features. However, existing approaches face trade-offs. Empirical models require extensive field data and often fail when applied to new environments. Physically based models are more transferable but struggle with ill-posed inversions. Hybrid methods combine both strategies, yet they commonly suffer from “domain shift,” where simulated spectra used for training differ from real-world measurements.
A study (DOI: 10.1016/j.plaphe.2025.100125) published in Plant Phenomics on 10 October 2025 by Daoliang Li’s & Kang Yu’s team, China Agricultural University & Precision Agriculture Lab, offers a promising pathway toward faster, non-destructive monitoring of crop nitrogen status.
Using a combination of spectral transformation, parametric and non-parametric modeling, and cross-crop validation, this study systematically evaluated methods for leaf nitrogen content (LNC) estimation from hyperspectral reflectance. First, simulated directional–hemispherical reflectance (DHRF) spectra and measured bidirectional reflectance factor (BRF) spectra were processed using continuous wavelet transform (CWT) and first derivatives (FD) to reduce discrepancies caused by specular reflection and domain shift. The transformed spectra showed markedly reduced differences between simulated and measured data, particularly in the visible and near-infrared regions, and enhanced key absorption features, indicating improved spectral comparability. Based on these transformed spectra, parametric regression models using 30 vegetation indices (VIs) were tested. When trained on the full simulated dataset, several VIs derived from the nitrogen allocation model (e.g., GARI, GNDVI, GRVI, CI800,550) achieved moderate accuracy, whereas all VIs performed poorly under the protein-to-nitrogen conversion formulation. When regression models were recalibrated using a representative subset of simulated samples (the T100 dataset), LNC estimation accuracy improved substantially, with indices such as SR708,775 reaching RMSE=0.303 g m⁻² and R² = 0.494, demonstrating that sample representativeness outweighed sheer sample size for parametric approaches. Non-parametric hybrid methods were then evaluated by combining machine learning or deep learning models with spectral transformations. Across the full simulated dataset, deep learning models generally outperformed traditional machine learning models, with the Conv-Transformer achieving the best performance among hybrid methods and surpassing physically based inversions. When trained on the T100 dataset, the Conv-Transformer further improved accuracy (RMSE = 0.247 g m⁻², R² = 0.665), exceeding results obtained using the full simulation database. Finally, ablation and cross-crop validation demonstrated that both the spectral similarity-based sample selection strategy and the modified Transformer architecture contributed synergistically to performance gains. Consistent improvements were observed for maize, wheat, rice, and sorghum, confirming that the DeepSpecN framework enhances LNC prediction accuracy and robustness across different crops by effectively mitigating domain shift.
The results demonstrate that accurate leaf nitrogen estimation is possible even in data-scarce conditions, without requiring costly field calibration. This capability is particularly valuable for precision agriculture, where timely nitrogen diagnostics can support optimized fertilization, reduce environmental pollution, and improve crop yields. Because the approach relies on leaf-scale bidirectional reflectance—more practical than integrating-sphere measurements—it is well suited for routine agricultural monitoring and technology transfer.
###
References
DOI
10.1016/j.plaphe.2025.100125
Original Source URl
https://doi.org/10.1016/j.plaphe.2025.100125
Funding information
This work was supported by the ‘AmAIzed’ project funded by the AgroMissionHub, the National Natural Science Foundation of China (grant numbers 32373186), and the CAU-TUM joint PhD training program.
About Plant Phenomics
Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.