Biologically interpretable multi-scale feature engineering with two-stage pruning for tomato leaf disease classification

Automated tomato leaf disease classification remains challenging when different diseases exhibit similar color, texture, and necrotic symptoms. To improve pathological interpretability while maintaining competitive classification performance, this study proposes a multi-scale feature engineering framework that explicitly represents disease symptoms at both global and local scales. A composite Excess Green (ExG)-Triangle segmentation strategy is first used to preserve green, chlorotic, and diseased leaf tissues, achieving a Dice coefficient of 92.45% and a Recall of 99.40% on manually annotated leaf masks. Color, multi-scale Local Binary Pattern (LBP), shape, and lesion descriptors are then integrated, with a Lesion Spatial Logic (LSL) descriptor introduced to characterize the relative allocation of necrotic tissue between the marginal and interior regions of the leaf. A two-stage pruning strategy based on permutation importance and Pearson correlation reduces the feature dimensionality from 296 to 164, corresponding to a 44.6% reduction. The pruned Support Vector Machine (SVM) achieves 96.44% accuracy on the independent PlantVillage test set, while Early Blight Recall increases from 88.50% to 92.00%. An exact McNemar test shows no statistically significant difference between the full and pruned SVM models ( \(\:p=0.644\) ), and five-fold cross-validation shows limited performance variability across training partitions. Modern pretrained CNNs achieve higher absolute accuracy on PlantVillage, whereas the proposed framework provides explicit pathological feature representation, a compact classifier, and lower memory and model-storage requirements. These findings support pathology-guided feature engineering as an interpretable approach for tomato disease classification under controlled conditions, while further validation under greenhouse and field conditions remains necessary.