Benchmarking EfficientNet, MobileNet and ResNet Architectures for Cassava Disease Detection in Agricultural Robotics

Abstract

Aim: To demonstrate the feasibility of lightweight deep learning models for cassava disease detection, emphasizing their suitability for deployment on agricultural robots and other edge devices where computational efficiency is critical. Study Design: Experimental benchmarking of lightweight versus heavier convolutional neural networks on a standardized cassava disease dataset. Place and Duration of Study: Conducted as independent research by Michael Aboh at Santa Clara University, during spring 2022. Methodology: The 2020 Cassava Leaf Disease dataset, comprising approximately 26,000 labeled images across five categories: Cassava Mosaic Disease, Cassava Brown Streak Disease, Cassava Green Mottle, Bacterial Blight, and Healthy, was utilized. The dataset was divided into training (70%), validation (15%), and test (15%) subsets. Three convolutional neural networks were trained and evaluated: EfficientNet-Lite0 (lightweight), MobileNetV3-Large (mid-range), and ResNet18 (heavier baseline). Evaluation metrics included classification accuracy, macro-averaged F1 score, model size (parameters, M), and inference speed (frames per second, FPS). Model interpretability was examined using Grad-CAM to visualize disease-relevant activation regions. Results: ResNet18 achieved the highest test accuracy (85.3%) and macro-F1 score (0.757). EfficientNet-Lite0 achieved comparable performance (84.3% accuracy, 0.744 macro-F1) with one-third the parameters and faster inference (123 FPS). Conclusion: The findings show that lightweight convolutional neural networks can deliver competitive accuracy while offering significant efficiency advantages for real-time, low-power inference on agricultural robots and edge devices. These results highlight critical trade-offs between model complexity, accuracy, and speed for scalable, field-deployable AI systems.