Fusion of EfficientNet‑B0 and MobileNetV2 for tomato disease classification

Abstract

The study proposes a fusion architecture for tomato disease classification that combines two complementary backbones—EfficientNet‑B0 and MobileNetV2—via feature concatenation and a hierarchical fusion head. Each branch applies global pooling, dropout, and fully connected layers before feature aggregation; the concatenated representation is then processed by the fusion head to produce the final prediction. Evaluation on an independent test set (n = 1000) shows that the fusion model reduces loss and improves overall accuracy relative to individual backbones, effectively correcting many systematic errors made by MobileNetV2 while achieving performance comparable to EfficientNet‑B0. Paired statistical testing using McNemar's test indicates the reduction in misclassifications is practically significant, suggesting the improvements are not attributable to random variation. The concatenation-based aggregation preserves full channel-wise information from both backbones, enabling the fusion head to learn cross-channel interactions and selectively reweight complementary signals to mitigate the weaknesses of each backbone. These findings support the use of feature‑level fusion to enhance the robustness and accuracy of plant disease classification systems.