Maize, a critical staple crop in Zambia, faces persistent threats from foliar diseases such as Gray Leaf Spot, Northern Corn Leaf Blight, and Maize Streak Virus, significantly affecting smallholder productivity. Limited access to expert diagnostics, coupled with complex field conditions including occlusions and variable lighting, necessitates accessible, real-time disease detection systems tailored to local environments. To address this gap, this study first developed a novel field-captured dataset of Zambian maize leaf images, annotated with bounding boxes for disease lesions and labeled by disease type and severity to reflect real-world agri-ecological variability. Building on this dataset, we propose ZamYOLO-Maize, a multi-stage automated diagnostic framework integrating lesion detection, hierarchical disease classification, and severity assessment. A comparative evaluation was conducted using four state-of-the-art object detection models: YOLOv5n, YOLOv8s, YOLOv10s, and YOLOv8n, with performance assessed using precision, recall, F1-score, and inference speed. Experimental results demonstrate that YOLOv10s achieved the highest predictive performance (Precision = 0.997, Recall = 0.999, F1-score = 0.999), while YOLOv8n provided the optimal trade-off for edge deployment, achieving the fastest inference speed (4.65 ms/image) with a competitive F1-score of 0.995. The framework exhibited strong robustness under field variability, confirming its practical applicability. By integrating a locally representative dataset with an efficient deep learning pipeline, this study establishes a scalable foundation for mobile-based maize disease diagnostics, contributing to precision agriculture and supporting food security initiatives in Zambia and comparable agricultural regions.
Kalunga et al. (Wed,) studied this question.