Advancements and outlooks in utilizing Convolutional Neural Networks for plant disease severity assessment: A comprehensive review

In addition to causing production losses, plant diseases lead to environmental harm due to the indiscriminate use of pesticides. Once limited to disease detection, Convolutional Neural Networks (CNNs) applications now exhibit automatic severity calculation and classification solutions. Research in this domain is evolving with novel approaches and methods, prompting the necessity for reviews that can delineate the current landscape and guide researchers in developing new methodologies. This paper conducts a systematic literature review to examine the scientific aspects, current evolution of methods, their outcomes, and potential gaps for future research. Following searches in six different databases and prominent conferences and the subsequent application of inclusion and exclusion criteria, this work analyzed 64 articles, addressing five specific review questions. The findings indicate a rising interest in using CNNs for plant disease severity assessment. Both classification and segmentation CNN-based methods have been employed, with a trend toward improved and hybrid frameworks constituting approximately 50% of the studies. Notably, lightweight neural networks, Long Short-Term Memory (CNN-LTSM), attention mechanisms, transfer learning, and combinations with fuzzy logic and Generative Adversarial Networks (GANs) are gaining traction. Bacterial and fungal diseases in crops such as tomato, wheat, and rice are the most studied, accounting for 42% of the research. Results indicate that future research should focus on expanding datasets in natural environments, diversifying the range of crops studied, and enhancing model robustness and generalization. Additionally, there is a need to explore improved methods further, including transfer learning, few-shot learning, and self-supervised deep learning, while also developing lighter models suitable for real-time assessment in field conditions.