Benchmarking Lightweight Deep Learning Models for Field Based Sorghum Leaf Disease Detection

Abstract

Sorghum is a climate-resilient cereal crop central to food and livelihood security across Southern Africa, yet its productivity is frequently reduced by foliar diseases that smallholder farmers struggle to diagnose reliably in the field. While deep learning has demonstrated strong potential for automated plant disease identification, sorghum remains underrepresented in the literature, and no prior study has provided a comprehensive, deployment-oriented benchmark of modern architectures across its major leaf diseases. This study evaluates three widely used convolutional neural networks, MobileNetV3-Large, EfficientNet-B0, and ResNet-50, using the publicly available Mendeley Sorghum Disease Image Dataset, consisting of 7,167 field-derived RGB images across six disease classes, under a unified and standardized training pipeline. MobileNetV3-Large and ResNet-50 achieved the highest classification accuracy (99.54%), followed by EfficientNet-B0 (98.79%), while MobileNetV3-Large provided the most favorable balance between accuracy, model size, and inference latency. To assess robustness and generalization, additional experiments were conducted, including multi-seed stability, data augmentation sensitivity, robustness to image perturbations, and background ablation. Results show that performance is stable across training runs but sensitive to image degradation, particularly blur, and to the removal of contextual information, which led to a marked decline in accuracy. This indicates that models rely not only on localized disease features but also on broader contextual cues, raising important considerations for robustness and generalization in real-world deployment. This study establishes a reproducible, efficiency-aware benchmarking framework for sorghum disease classification and provides new insight into robustness, context dependence, and deployment constraints in field-based plant disease detection, supporting the development of lightweight diagnostic tools for resource-constrained environments.