CNN Feature Characterization via the Mahalanobis Taguchi System for Recognizing Optimal Date-Fruit Harvest: A Systematic Literature Review

Abstract

This study investigates a hybrid approach for determining optimal date-fruit harvest timing by integrating deep visual features from Convolutional Neural Networks (CNNs) with the Mahalanobis Taguchi System (MTS). High- resolution images of date fruits spanning canonical maturity stages (Kimri, Khalal, Rutab, Tamr) are processed through a pretrained CNN to extract discriminative embeddings emphasizing color, texture, and size cues. Candidate features are organized and screened using MTS orthogonal arrays and signal-to-noise (S/N) analysis to identify a parsimonious subset that maximizes class separability while minimizing redundancy. The Mahalanobis Distance (MD) then calibrated on a “healthy/optimal” reference space to yield decision thresholds for harvest readiness. The pipeline includes stratified cross-validation, ablation of feature groups, and comparisons against conventional machine-learning baselines using the same inputs. Results indicate that MTS-guided feature selection consistently improves generalization and interpretability, producing stable MD thresholds that align with horticultural expectations and reducing computational overhead relative to full-feature CNN embeddings. Beyond classification, the framework yields actionable diagnostic feature contribution ranks and MD control charts, that support field decisions and quality control. The proposed CNN-MTS methodology provides a transparent, statistically grounded route to operationalize computer vision for harvest scheduling and offers a reusable template for similar post-harvest applications where explainability and small-sample robustness are essential.