Mobile-assisted deep learning framework for identification of insect pests and diseases of maize from field images

Maize ( Zea mays L. ) production is severely affected by diseases and insect pests, leading to significant yield losses when timely diagnosis and management interventions are not implemented. Although automated image-based diagnostic systems have shown promising results, most existing studies address diseases or pests independently, rely on controlled datasets, and offer limited robustness under real field conditions. To address these limitations, this study proposes a unified deep learning–based framework for integrated identification of maize diseases and insect pests under natural field environments by combining object detection and image classification within a mobile-assisted diagnostic system. Four economically important diseases and insect pests were investigated: Maydis Leaf Blight (MLB), Turcicum Leaf Blight (TLB), Common Rust, and Fall Armyworm (FAW). MLB and TLB were addressed using YOLO-based object detection architectures, while Common Rust and FAW were treated as image-level classification tasks using lightweight deep learning models optimised for mobile inference. A self-collected dataset comprising 10,343 images across four classes was acquired under real field conditions to capture variability in background complexity, illumination, phenological stages, and symptom expression. Experimental results on an independent test set comprising original images demonstrate that MobileViT achieved the highest classification accuracy (99%) for image-level disease and pest recognition, whereas YOLOv11n outperformed other detection models, achieving the best performance for MLB and TLB lesion detection with [email protected] of 0.875. Grad-CAM–based visual explanation analysis confirmed that the classification models focused on disease lesions and pest-infested regions, supporting interpretability. The framework was successfully deployed via a mobile application, enabling image acquisition, automated validation, diagnosis, and the generation of management recommendations. The results highlight the accuracy, robustness, and operational feasibility of the proposed system for in-field diagnosis of maize diseases and insect pests, supporting early detection and sustainable crop protection.