Mid-FTIR and machine learning for predicting fig leaf macronutrients content

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Accurate prediction of leaf mineral composition is crucial for plant health monitoring and agricultural optimization.
• Traditional methods for nutrient analysis are often destructive and time-consuming.

Purpose of the Study:

• To develop and evaluate a rapid, non-destructive method for predicting macronutrient (N, P, K, Ca, Mg) levels in fig leaves (Ficus carica L.).
• To compare the performance of different machine learning models, including Random Forest (RF), Support Vector Regression (SVR), and Gradient Boosting (GB), when combined with FTIR-ATR spectroscopy.

Main Methods:

• Fourier-transform infrared spectroscopy with attenuated total reflectance (FTIR-ATR) was used to collect spectral data from 90 fig leaves.
• Spectra were preprocessed using baseline correction and second-derivative transformations.
• Three machine learning models (RF, SVR, GB) were trained and evaluated using fivefold cross-validation, assessing performance with RMSE, R², and RPD.

Main Results:

• The Gradient Boosting (GB) model demonstrated superior predictive performance across all macronutrients compared to RF and SVR.
• GB achieved the best validation results for magnesium (Mg) and phosphorus (P), with R² values of 0.7351 and 0.6873, respectively.
• While some overfitting was observed (training R² ≈ 0.999 vs. validation R² = 0.61-0.74), GB's reliability (RPD > 1.5) indicates its utility for nutrient screening.

Conclusions:

• FTIR-ATR coupled with machine learning, particularly the Gradient Boosting model, offers a viable, rapid, and non-destructive alternative to conventional plant tissue analysis.
• The model's performance provides actionable insights for precision nutrient management, contributing to sustainable agricultural practices.
• Distinct spectral features for Mg and P likely contributed to their higher prediction accuracy.