Deep learning to improve plant biosecurity and agriculture sustainability: A systematic literature review

Abstract Plant stress significantly threatens agricultural productivity and quality, necessitating advanced solutions for effective plant biosecurity. Manual inspection is time-consuming and requires expert knowledge, making automated monitoring crucial for crop protection and food security. While deep learning has shown significant potential in plant stress monitoring and improving plant biosecurity, the existing literature remains unsystematic and highlights numerous challenges that hinder its real-time application. Therefore, a systematic literature review is essential to consolidate current knowledge, offer a comprehensive overview of deep learning-based plant biotic stress monitoring, and identify research gaps, trends, and challenges. In this view, this article employs the PRISMA framework to analyze 745 articles from six recognized electronic databases published between 2016 and April 2024, addressing eight research questions. Findings from this study indicate that convolutional neural networks dominate the field, appearing in 85.23% of articles, with limited consideration for hybrid models. Further, the reliance on supervised learning and high-quality annotated datasets highlights significant challenges related to data availability, diversity, and model generalization. Besides, the limited focus on efficiency metrics for model evaluation further compounds these issues, hindering accurate assessment of model performance in real-world applications. Additionally, only 8.19% of the articles offered directly applicable solutions to farmers, highlighting the necessity for practical and scalable applications. Moreover, while 63.42% of studies originated from the USA and China, research output from Africa, Central Asia, South America, and Southeast Asia remains low. These findings highlight the need for equitable research and localized solutions to address challenges in underrepresented regions. Future research should explore unsupervised and semi-supervised learning to reduce reliance on annotated datasets, integrate IoT, edge computing, and cloud-computing for real-time deployment, and incorporate efficiency metrics to ensure robust real-world performance. This review offers a comprehensive overview of the latest advancements and outlines potential future directions for computer science, agriculture, and ecology.