Abstract
Artificial intelligence (AI) enables rapid and precise plant disease detection, offering transformative potential for crop protection. Maize downy mildew (MDM), a destructive disease, causes substantial yield losses, making early detection critical. In this study, we evaluated the performance of thirteen machine-learning (ML) and deep-learning (DL) algorithms for classifying healthy and infected maize leaves using a curated field dataset. Model performance was assessed using multiple metrics, including accuracy, precision, recall, F1-score, and AUC-ROC. Among the tested models, VGG16 achieved the highest performance, with 97% accuracy, 0.98 precision, 0.95 recall, 0.97 F1-score, and an AUC-ROC of 0.99. Training and validation curves indicated minimal overfitting, demonstrating robust generalization. Feature visualization using t-SNE revealed clear separability between healthy and diseased samples, while Grad-CAM analysis confirmed that VGG16 focused on biologically relevant symptomatic regions, such as chlorotic streaks and leaf discoloration. Confusion matrix analysis further validated near-perfect classification, with very few misclassifications. Furthermore, we developed a web-based application (https://maize-mdm.streamlit.app/) that not only classifies MDM but also provides farm-level advisory measures. Two-year field trials of DSS-guided fungicide applications effectively suppressed MDM, reducing disease severity (PDI 3.20-5.20; PROC 93-96%), increasing grain yield (75.6-80.2 q/ha; PIOC 195-289%), and improving economic returns (B:C ratio 3.36-3.57) compared to untreated controls. Overall, this study demonstrates that AI-driven models, integrated with web-based decision support, provide accurate, interpretable, and actionable solutions for precision management of maize diseases, contributing to improved yield, profitability, and sustainable agricultural practices.