A light-weight enhanced multi-level attention network for plant disease identification

Plant diseases are the main factor responsible for global crop losses, significantly impacting the world economy. Researchers are developing intelligent agriculture solutions integrating the Internet of Things and Machine learning (ML) techniques to detect and control diseases early on. Many of these systems employ ML techniques that rely on visual data to detect and diagnose diseases in real time. Deep learning techniques have revolutionized plant disease detection and identification, developing novel approaches that utilize convolutional neural networks (CNN). This study presents a lightweight, enhanced multi-level feature extraction network that can find plant diseases efficiently. The suggested network contains a local network, a global network, and a transformer encoder. These modules extract the most useful and distinctive characteristics from images of plant leaves. Finally, we integrate the features gathered at each level, send them to the dense layers, and then move them to the softmax layer to compute the probabilities. The proposed model attained an accuracy of 98.55% on the cassava and 99.48% on the rice leaf dataset, greatly outperforming current techniques with an improvement range of 3.35- \(-\) $-$ 8.55% for cassava and 2.58- \(-\) $-$ 6.68% for rice leaf. Our ablation investigation underlined the importance of each component in our model. Specifically, the global feature block, local feature block, and transformer encoder all contributed significantly to the model’s performance. When combined, these components led to the highest accuracy. We validate the superiority of our proposed method by comparing it against ten state-of-the-art approaches across two benchmark datasets.