Machine Learning and Deep Learning for Plant Disease Detection: A Review of Techniques and Trends

Abstract

Plant diseases pose a significant threat to global agricultural productivity, making early and accurate detection critical for yield protection and food security. This study evaluates the evolution, effectiveness, and practical applicability of Machine Learning (ML) and Deep Learning (DL) models for plant disease detection while analyzing research trends to identify leading models, data limitations, and implementation challenges. A systematic literature review and bibliometric analysis were conducted using the PRISMA framework, examining 625 peer-reviewed articles published between 2017 and 2025 from major databases. The analysis highlights the most influential studies, commonly used datasets, and top-performing ML/DL models, assessed in terms of accuracy, methodology, dataset type, and real-time deployment potential. Results show that models such as YOLOv4, VGG19, ResNet50, and MobileNetV2 achieved accuracy levels between 98% and 99.99%, with most trained on the PlantVillage dataset or custom annotated datasets. Several studies demonstrated successful real-time deployment via mobile and edge-device applications. However, key challenges remain, including limited dataset diversity, poor model generalization across environments, and reduced performance under real-field conditions. This study provides a comprehensive overview of progress in AI-based plant disease detection, emphasizing the need for lightweight, adaptable, and field-ready models to support scalable real-world deployment.

Plant phenotyping relevance

植物病害状態の画像ベース検出手法を対象とした系統的レビューであり、モデル、データセット、精度、実運用性を方法論的に評価しているため、植物フェノタイピング手法レビューとして含める。

Code and data availability

This is a systematic literature review of plant disease detection using ML/DL. It presents no original phenotype datasets, images, code, or trained models of its own; the only URLs in the text are citations to prior works (references [8], [11], [21]) and are not paper-specific assets of this review.

No evidence-backed public reproduction asset is currently recorded.