Integrating hyperspectral radiative transfer modeling and Machine learning for enhanced nitrogen sensing in almond leaves

Precisely quantifying crop nitrogen content is critical for adopting sustainable nutrient management practices. This study offers a comprehensive analysis of using hyperspectral data to accurately measure area-based nitrogen content (N) in almond trees at the leaf level. We collected spectral data ranging from 400 to 2500 nm of multiple leaves from 190 samples across two orchards spanning two years. Our methodology involves building a hybrid model that merges a physically based model (PROSPECT-PRO) and a data-driven model (multi-output Gaussian process regression), demonstrating exceptional performance in area-based nitrogen prediction, achieving R² values of 0.54 and an RMSE of 0.03 mg/cm² for area-based nitrogen sensing. The hybrid method incorporates synthetic spectra produced through principal component analysis (PCA) and labeled with biochemical traits retrieved by PROSPECT-PRO for training and validation, while the real data was kept unseen for testing. We compared the performance of physically based, hybrid, and data-driven models using R² and NRMSE as metrics. The Partial Least Squares Regression (PLSR) model showed a strong relationship between leaf N and spectral reflectance (R² = 0.75); however, PLSR is prone to bias from the training set and may perform poorly on unseen data. The findings also highlight the importance of the Short-Wave Infrared region in nitrogen determination, particularly the bands from 2100 to 2200 nm. Additionally, protein content was found to be a more reliable proxy for nitrogen than chlorophyll. By comparing the retrieved leaf traits with ground truth data, we realized that PROSPECT PRO consistently underestimates almond leaf traits such equivalent water thickness (EWT), carbon-based compounds (CBC), and overestimates Nitrogen. Therefore, adjustment factors were determined for these traits that are estimated with PROSPECT-PRO.