EMERGING DEEP LEARNING APPROACHES FOR PLANT DISEASE IDENTIFICATION: TRENDS AND INNOVATION

Abstract

Plant diseases remain a significant global agricultural productivity threat, necessitating the adoption of Artificial Intelligence (AI), and more so Deep Learning (DL), in precision agriculture. This thesis distills results from 25 peer-reviewed journal articles between 2020 and 2025 on novel DL methods for the identification of plant diseases, with a heavy focus on Convolutional Neural Networks (CNNs). More than half of the studies reviewed employed CNN-based models because of their established success in accurate classification and real-time diagnosis. Lightweight optimized CNN models like Shallow CNN, VGG-ICNN, and Optimized Custom CNN were often designed for mobile and resource-constrained environments. Some studies incorporated enhancement methods such as feature reduction, residual learning, and optimization algorithms (e.g., Beluga Whale Optimization) to enhance further model efficiency and accuracy. Hybrid models incorporating CNNs with other deep learning techniques—such as LSTM networks, autoencoders, and Vision Transformers (ViTs)— proved to be a notable trend. Architectures such as PlantXViT and MobilePlantViT exhibited promising performance in terms of both interpretability and performance. Data augmentation strategies like LeafGAN also helped enhance model generalization through the creation of synthetic disease images. The research also investigated practical applications, such as mobile apps and real-time detection software, with high accuracy rates (up to 99%). Common datasets such as PlantVillage and AgroPath were used as the foundation for training and testing these models. In general, the researched papers depict an increasing trend towards lightweight, hybrid, and explainable deep learning models, pushing the research area of automatic plant disease detection and enabling sustainable, technology- based agricultural practices.