<p>In agricultural technology, accurate identification and classification of plant diseases are essential to maintaining crop health and updating agrarian productivity. Traditional methods often require extensive human expertise and are time-consuming. Deep learning models have shown significant success in plant disease classification studies, yet many research datasets are constructed under controlled conditions. This limitation highlights the challenge of deploying models that perform reliably across diverse and less controlled environments. The study proposes multi-level features and enhancement operations with the attention network (MFANet) for plant disease classification to address this issue. The proposed model uses advanced attention mechanisms to enhance the discriminatory capability of feature maps and improve classification performance. MFANet integrates residual channel and spatial attention within its architecture to focus on salient features and conceal irrelevant variations in complex plant image data. The effectiveness of MFANet is explained across four separate plant datasets, each presenting unique classification challenges. Our model achieves better performance metrics, markedly outperforming existing approaches. On the bean dataset, MFANet achieved an accuracy of 98.75% and a loss of 0.027. For the cassava dataset, it recorded 92.40% accuracy and 0.034 loss. In the grapevine dataset, the model reached 96.05% accuracy with a 0.021 loss; for the rice dataset, it achieved an impressive accuracy of 99.30% and a loss of 0.018. These results highlight the robustness and potential of MFANet in significantly advancing the field of plant disease diagnosis, providing a scalable solution for deployment across various agricultural settings, thereby facilitating rapid and accurate disease management.</p>