A comprehensive evaluation of lightweight deep learning models for tomato disease classification on edge computing environments

To achieve agricultural automation, deep learning applications for early and accurate disease detection in tomato plants have been extensively developed. However, there is a fundamental trade-off between computational efficiency and diagnostic accuracy in resource-constrained agricultural edge environments. This paper proposes an evaluation framework for seven architectures that represent standard, efficient, and hybrid CNN structures to assess their implementation potential. Through evaluations of explainability, computational efficiency, and diagnostic performance, seven lightweight architectures (ShuffleNetV2, MobileNetV3-Small, SqueezeNet, MobilePlantViT, DenseNet121, ResNet50, and VGG16) are thoroughly examined. Three significant findings are derived from experiments conducted on a subset of tomato diseases in the PlantVillage dataset. First, the MobilePlantViT architecture accurately strikes the ideal balance between efficiency and performance. Second, in order to quantitatively assess the explainability of XAI models (Grad-CAM, SHAP, and LIME) and identify the best option for edge devices, we propose the perturbation stability score (PSS) metric. Third, we test CPU inference measurements to better reflect the actual scenario and find that the hybrid design effectively leverages parallel computing. According to these findings, MobilePlantViT is the ideal architecture for applications that require operation on edge devices with limited resources and achieve high diagnosis accuracy (above 99.5%).