A tertiary systematic literature review and experimental evaluation of deep learning models for plant disease detection

Abstract

Abstract Objective Minimizing crop losses through the early detection of plant diseases is vital for enhancing global agricultural efficiency. While deep learning has emerged as a promising solution, a significant gap exists between laboratory performance and practical, in-field utility. This study evaluates this discrepancy through a dual-methodological approach. Methods First, a tertiary systematic literature review was conducted, synthesizing 22 secondary reviews encompassing over 750 unique primary studies to establish the current state of the art. Second, an empirical validation was performed using a VGG16 transfer learning model trained on three distinct dataset types, which vary in scale (small vs. large), environment (laboratory vs. in-field), and condition (raw vs. pre-processed). Results The tertiary review identifies Convolutional Neural Networks, particularly VGG architectures, as the leading model but highlights a critical reliance on private and unrealistic datasets. Furthermore, the analysis reveals that Accuracy, the most common metric, is often insufficient for evaluating the imbalanced datasets typical of the field. Empirical results corroborate these findings, demonstrating that VGG16 performance is highly dependent on dataset characteristics; models perform significantly better on large, pre-processed laboratory data than on realistic in-field datasets. Conclusion These findings suggest that many current models remain inapplicable to real-world agricultural scenarios. To bridge this reality gap, future research must prioritize the development of open-source, standardized, and validated in-field datasets to ensure the reliability and scalability of automated disease detection systems.

Plant phenotyping relevance

植物病害を画像から検出する深層学習手法を体系的にレビューし、VGG16を異なるデータセット条件で実証評価しており、病害状態の取得・推定方法が中心である。

Code and data availability

The supplied blocks contain only the article's abstract, citation metadata, and access links (DOI and edepot full-text). No public phenotype/trait datasets, plant images, author analysis code, or trained model checkpoints are mentioned with any availability statement or URL. The VGG16 model and three dataset types are,

No evidence-backed public reproduction asset is currently recorded.