Benchmarking mobileNetV3 and efficientNet-B0 for corn leaf disease classification with imbalanced dataset using stratified cross-validation

Abstract

Corn leaf diseases pose a serious threat to crop productivity, yet most publicly available datasets for this task exhibit severe class imbalance that can mislead conventional accuracy-based evaluation. This study benchmarks two lightweight transfer learning architectures, MobileNetV3-Large and EfficientNet-B0, for multi-class corn leaf disease classification on the Seasonal Corn Leaf Disease Dataset from Mendeley Data 2025 containing 2,943 images across five imbalanced classes. Evaluation was conducted using Stratified 5-Fold Cross-Validation with Macro-F1 as the primary metric, complemented by per-class analysis through aggregated out-of-fold predictions. Class weights were applied to the CrossEntropyLoss function as a fixed experimental control for class imbalance, with the primary objective being the benchmarking of the two architectures rather than the comparison of imbalance-handling strategies. The experimental results revealed that EfficientNet-B0 consistently outperformed MobileNetV3, achieving a Macro-F1 of 0.9778 and an accuracy of 0.9796 with lower variance across folds. Error analysis through the OOF confusion matrix and a misclassification gallery confirmed that persistent errors predominantly occurred between Gray Leaf Spot and Healthy classes, particularly on early-symptom images captured under inconsistent lighting conditions.