Comparative deep learning study for downy mildew detection in vegetables

Abstract

Several vegetable crops are affected by downy mildew, a major foliar disease resulting in notable reductions in yield. For sustainable agriculture and disease prevention, early and precise detection is crucial. To be able to detect downy mildew in five varied vegetables—bitter gourd, bottle gourd, cauliflower, cucumber (Rashid), and cucumber (Sultana)—this study evaluates three deep learning architectures: VGG19, DenseNet201, and MobileNetV2. This work focuses on imbalanced datasets collected from several sources, in opposition to prior work that depended on balanced laboratory datasets. Accuracy, precision, recall, and F1-score metrics were used to evaluate the models shortly after they were trained using transfer learning, data augmentation, and 5-fold cross-validation. Model focus regions were assessed by using gradient-weighted class activation mapping (Grad-CAM) visualizations, and statistical reliability was assessed based on paired t-tests and Wilcoxon signed-rank tests. By achieving mean accuracies above 98% and statistically significant results (p <0.05) on cucumber datasets, DenseNet201 accomplished superior performance. Despite attaining slightly lower accuracy (89.6–100%), MobileNetV2 offered the smallest model size (12.9 MB) and minimum inference time (85 ms). The proposed approach demonstrated a transparent, generalizable, and computationally efficient deep learning pipeline for precision agriculture’s real-time downy mildew detection.