Using preprocessed datasets to construct and interpret multiclass identification models.

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Data Science

Background:

• Image and near-infrared (NIR) spectroscopy are vital for precision agriculture analytical models.
• Direct use of raw data presents challenges in model interpretability and robustness due to data ambiguity and imbalanced datasets.

Purpose of the Study:

• To develop interpretable and robust multiclass identification models using preprocessed agricultural data.
• To overcome limitations of raw image and NIR spectral data in analytical modeling.

Main Methods:

• Utilized preprocessed data: morphological features from images and chemical component concentrations from NIR spectra.
• Employed combined kernel Support Vector Machine (SVM) models for classification.
• Optimized model parameters using Particle Swarm Optimization (PSO) for self-adaptability.
• Conducted feature importance and contribution analysis with Shapley Additive Explanations (SHAP).

Main Results:

• Achieved high classification accuracy: 97.9% for rice variety, 97.4% for tobacco cultivation region (cross-validation).
• Validated model performance on an independent tobacco dataset with 97.7% accuracy.
• Identified key predictive variables and quantified their contributions to model outcomes.

Conclusions:

• The proposed methodology effectively enhances the interpretability and robustness of analytical models in precision agriculture.
• This approach expands the utility of image and NIR spectroscopic data for agricultural quality control and improvement.
• Provides researchers with a powerful tool for investigating critical factors in agricultural product quality.