<![CDATA[Maize (Zea mays) seed quality assessment is essential for supporting agricultural productivity and sustainable seed management. This study proposes a non-destructive machine learning framework for predicting maize seed quality using portable Visible/Near-Infrared (Vis/NIR) spectroscopy. The framework integrates NIPPY-based spectral preprocessing, metaheuristic wavelength selection, and fuzzy ensemble learning to handle spectral noise, multicollinearity, and nonlinear relationships in small-sample spectral data. Informative wavelengths were selected using Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). Two fuzzy ensemble models were developed: a Fuzzy Residual-Corrected Ensemble that refines predictions through residual-based fuzzy correction, and an RF+XGB Fuzzy Ensemble that combines Random Forest and XGBoost outputs using confidence-based fuzzy weighting. The models were evaluated for Moisture Content (MC), Germination Rate (GR), and Electrical Conductivity (EC) using repeated cross-validation, variability measures, and statistical validation. The proposed fuzzy ensemble models achieved R² values ranging from 0.8249 to 0.8689 and showed performance comparable to the strongest Random Forest baseline. Statistical comparison indicated that the main contribution of the fuzzy ensemble framework lies not in large gains in mean accuracy, but in prediction stability, residual correction, and uncertainty-aware modeling. SHAP-based explainability further identified physiologically meaningful wavelength regions, including visible pigment-related bands and near-infrared moisture-related bands. The dataset consists of 800 maize seed samples from four varieties under laboratory conditions, which limits generalization to field environments. Future work will focus on multi-location validation, domain adaptation, and real-time implementation. Overall, the proposed framework provides a statistically validated and interpretable approach for portable Vis/NIR-based maize seed quality prediction.]]>
Copyrights © 2026
