A review of deep learning architectures for plant disease detection.

Abstract

Background/aim The rapid advancement of deep learning (DL) has revolutionized plant disease detection by enabling highly accurate, image-based diagnostic solutions. This review provides a comprehensive synthesis of DL-based methodologies for plant disease detection, systematically structured around the key stages of the modeling pipeline, encompassing data acquisition, preprocessing, augmentation, classification, detection, segmentation, and deployment. Materials and methods The review focuses on evaluating convolutional neural network (CNN) architectures such as VGG, ResNet, EfficientNet, and DenseNet across diverse experimental contexts. Classification strategies are categorized according to their integration of visualization techniques (e.g., saliency maps, Grad-CAM) to enhance model interpretability, emphasizing the pivotal role of explainable artificial intelligence (XAI) in plant pathology. Object detection models are systematically examined within both one-stage (YOLO, SSD) and two-stage (Faster R-CNN) paradigms. Furthermore, critical challenges-such as environmental variability, data imbalance, and computational constraints-along with potential solutions including transfer learning, synthetic data generation using generative adversarial networks (GANs) and diffusion models, and edge computing for real-time deployment, are comprehensively discussed. Results This review summarizes best practices for dataset selection and model optimization for mobile platforms, emphasizing their role in improving the efficiency and accuracy of plant disease detection systems. Conclusion Deep learning-based methods show strong potential to enhance precision and resilience in real-world plant disease detection and monitoring.

Plant phenotyping relevance

植物病害の画像ベース検出手法を体系的にレビューしており、病害状態という植物表現型の取得・推定方法が中心である。

Code and data availability

This is a review article synthesizing prior literature on deep learning for plant disease detection. The authors conducted no original phenotyping measurements or computational experiments. The many dataset URLs in the reference list (PlantVillage, LeafSnap, Kaggle datasets, Mendeley datasets, etc.) are third-party, un

No evidence-backed public reproduction asset is currently recorded.