Deep Learning for Smart Agriculture: A Comprehensive Review of CNN Architectures, Multispectral Imaging, Explainable AI and Transfer Learning for Crop Disease Detection

Abstract

Global food security faces unprecedented pressure from crop diseases, which are responsible for annual yield losses estimated at 20–40% of all food production. The application of deep learning methodologies, particularly convolutional neural network (CNN) architectures, to the automated detection and classification of crop diseases has emerged as a transformative paradigm within the domain of smart agriculture. A structured literature search was conducted using the academic databases Web of Science, Scopus, Google Scholar, and PubMed, covering the publication period from January 1996 to March 2026. The search strategy employed a combination of controlled vocabulary and free text search strings, including but not limited to the following terms and their Boolean combinations. The review examining the theoretical underpinnings and empirical performance of diverse CNN architectures including VGGNet, ResNet, Inception, DenseNet, EfficientNet, and Vision Transformers as applied to plant pathology. Special attention is directed towards the role of multispectral and hyperspectral imaging modalities, which extend disease detection capabilities beyond the visible spectrum and enable the identification of latent biochemical stress signatures before visible symptom onset. The review further explores the critical contribution of transfer learning in addressing the perennial challenge of limited annotated agricultural datasets, demonstrating how pre trained models can be fine tuned to achieve high diagnostic accuracy across diverse crop pathogen combinations. A dedicated section examines the rapidly maturing field of Explainable AI (XAI), with particular focus on gradient weighted class activation mapping (Grad CAM), integrated gradients, and SHAP based methods, which are essential for building agronomist trust and regulatory acceptability. The synthesis identifies persistent challenges including domain shift, class imbalance, computational constraints in field deployable systems, and the scarcity of standardised benchmark datasets. The review concludes with a forward looking perspective on federated learning, multimodal fusion architectures, and the integration of UAV based sensing with edge computing as the frontier of next generation agricultural AI systems.