Comparative Analysis of Decision-Level Ensemble Voting Strategies for Tomato Leaf Disease Classification Using EfficientNet-B0 and MobileNetV2

Abstract

Abstract. Tomato leaf diseases are a major threat to agricultural productivity, making accurate and timely disease classification essential for supporting precision agriculture. Recent advances in convolutional neural networks have significantly improved automated plant disease recognition, while ensemble learning has further enhanced classification performance by combining predictions from multiple models. Among various ensemble approaches, decision-level voting strategies, including hard voting, soft voting, and weighted soft voting, are widely adopted because of their simplicity and effectiveness. However, systematic comparisons of these voting strategies, particularly when combining lightweight CNN architectures for tomato leaf disease classification, remain limited. Therefore, this study investigates the comparative performance of different decision-level voting strategies to provide empirical evidence for selecting the most appropriate ensemble approach. Purpose: This study aims to systematically compare hard voting, soft voting, and weighted soft voting for combining EfficientNet-B0 and MobileNetV2 to identify the most effective decision-level ensemble strategy for tomato leaf disease classification. Methods/Study design/approach: EfficientNet-B0 and MobileNetV2 were trained independently and combined using hard voting with confidence-based tie breaking, soft voting, and weighted soft voting, with voting weights optimized through grid search. Performance was evaluated using accuracy, macro-F1 score, confusion matrices, and McNemar's test. Result/Findings: EfficientNet-B0 outperformed MobileNetV2 as the best single model (94.05% accuracy, 0.9266 macro-F1 vs. 90.64% accuracy, 0.8781 macro-F1). All three ensemble strategies further improved classification performance, with weighted soft voting achieving the highest performance (95.15% accuracy, 0.9417 macro-F1). McNemar's test showed that only soft voting and weighted soft voting significantly outperformed the best single model, whereas hard voting did not. No statistically significant differences were observed among the three ensemble voting strategies. Novelty: This study systematically compares hard voting, soft voting, and weighted soft voting for decision-level ensemble using EfficientNet-B0 and MobileNetV2 in tomato leaf disease classification. In addition to evaluating predictive performance, the study incorporates McNemar's statistical test to determine whether the observed performance differences among the ensemble strategies are statistically significant, providing empirical guidance for selecting appropriate voting strategies.

Keywords

• Tomato Leaf Disease
• Decision-Level Voting
• Ensemble Learning
• EfficientNet-B0
• MobileNetV2