Multi-Model Deep Learning with EDAS-Based Multi-Criteria Decision-Making for Robust Potato Leaf Disease Classification

Ahmad, Niaz, Alam, Fakhre, Ullah, Khalil
Potato Research · Springer Nature · 2026-10-30 · DOI: 10.1007/s11540-026-10166-0

Аннотация

Potato is a valuable crop both nutritionally and economically. However, its productivity is adversely affected by foliar diseases. Accurate and timely disease detection is critical. Deep learning models have achieved significant progress; however, limited work exists on specialized datasets. Most existing frameworks are trained on one or two deep learning models and use standard metrics evaluation, lacking multi-criteria decision-making for comprehensive model evaluation. To bridge the gaps in the available research, this study offers three key contributions. The creation of a well-labeled, balanced dataset consists of 5244 potato leaf images, representing six classes. The images were captured under different environmental conditions to accelerate model generalization. The evaluation of seven fine-tuned pre-trained convolutional neural networks (CNNs) on the benchmark dataset. The application of the evaluation based on distance from average solution (EDAS) method to multi-criteria model ranking, both overall and per-disease class, enables effective model selection. Experimental results demonstrate that InceptionV3 outperforms other models with a classification accuracy of 90.2% and an EDAS score of 0.8805, followed by Xception (accuracy 89.6%, EDAS 0.7675), DenseNet-201 (accuracy 89.3%, EDAS 0.7238), ResNet-50 (accuracy 89.1%, EDAS 0.6223), MobileNetV2 (accuracy 87.9%, EDAS 0.5000), GoogLeNet (accuracy 85.9%, EDAS 0.5000), and Darknet-53 (accuracy 85.2%, EDAS 0.500). EDAS-based disease-specific model ranking demonstrated that MobileNetV2 achieved the highest overall performance with an Average Satisfaction Performance (ASP) of 0.6460. The model achieved excellent results for Late Blight and Black Spot (both 1.0000), performed well on most other classes, though it was negligible on Early Blight. DenseNet-201 (ASP 0.6124) and Xception (ASP 0.5525) were close, with high scores on Late Blight, Healthy, and Black Spot. It is worth noting that all models detected Late Blight perfectly but consistently failed at detecting Early Blight. The ablation study shows the importance of backbone fine-tuning and L2 regularization for InceptionV3. The model achieved stable classification performance. Robustness was tested on blur, brightness shifts, domain shifts, and noise. Perturbation demonstrates resilience to general field distortions but performs poorly in extreme noise conditions. These findings show a trade-off between general accuracy and disease-specific strengths, demonstrating that EDAS is effective at selecting robust, targeted models.

Читать у издателя →

Похожие статьи