DeepSpecN: A new hybrid method combining PROSPECT-PRO and Conv-Transformer to estimate leaf nitrogen content from...

Area of Science:

• Agricultural Science
• Remote Sensing
• Spectroscopy

Background:

• Accurate leaf nitrogen content (LNC) estimation is vital for crop monitoring.
• Existing methods (empirical, physical, hybrid) face challenges like extensive data needs, ill-posed inversion, and domain shift.

Purpose of the Study:

• To develop a novel hybrid method (DeepSpecN) for non-destructive maize LNC estimation using hyperspectral bidirectional reflectance.
• To overcome limitations of traditional methods, particularly the domain shift issue in hybrid approaches.

Main Methods:

• DeepSpecN integrates continuous wavelet transform (CWT), PROSPECT-PRO simulation, an improved Transformer model, and spectral similarity-based sample selection.
• Validated against physically-based, non-parametric hybrid, and vegetation index (VI)-based methods using 1724 maize leaf samples.

Main Results:

• DeepSpecN achieved superior LNC estimation accuracy (RMSE = 0.247 g/m², R² = 0.665) compared to other methods.
• The sample selection strategy effectively addressed domain shift by identifying representative simulated samples.
• Chlorophyll-based empirical formulas outperformed protein-based ones for LNC estimation.

Conclusions:

• DeepSpecN demonstrates high accuracy and generalizability across different crop species for LNC estimation.
• Addressing domain shift in hybrid methods using bidirectional reflectance is key to improving LNC estimation.
• The study highlights the potential of advanced hybrid methods for precision agriculture.