Agricultural sustainability plays a vital role in ensuring global food security, economic development, and environmental protection. With rapid technological advancements, smart agriculture has become a transformative domain driven by deep learning, the Internet of Things (IoT), and big data analytics. Traditional farming and classical machine learning methods often struggle to handle the complexity, variability, and scale of agricultural data, especially in dynamic field environments. In order to overcome these shortcomings, this review examines 25 recent studies that are dedicated to deep learning-based plant disease detection and crop classification. Some of the models used in the studies are convolutional neural networks (CNNs), deep neural networks (DNNs), transfer learning, attention mechanisms, lightweight models (MobileNet), and hybrid networks (CNN-SVM, Inception-Xception, and vision transformers). These methods have been implemented on crops such as rice, wheat, corn, and potato where the high accuracy has been realized in the controlled and semi-field environments. The most recent trends identified in these studies include lightweight edge-compatible models, ensemble & hybrid learning, and explainable AI for greater interpretability. However, the issues of limited dataset diversity, model generalization and IoT deployment challenges remain prevalent. Overall, this study consolidates some methodologies and findings to pave the way toward the development of scalable, interpretable and viable deep learning architectures for sustainable agriculture.