In recent years, early detection and management of plant diseases at an early stage has become very important for sustaining crop health, increasing yields, and protecting food security in agricultural systems. Untreated diseases caused by fungi, bacteria, viruses, and pests significantly reduce agriculture and pose a risk to global food production. However, existing models face some challenges in agriculture because they focus only on local features, fail to focus on noisy backgrounds, lighting, overlapping leaves, and difficulty in detecting different parts of the leaf. To overcome this problem, this paper proposes an attention-fused residual lightweight (RTR-Lite-MobileViT) model, which is an enhanced version of the original MobileVit model designed for efficient deployment on resource-constrained devices. Different attention techniques, such as Squeeze-and-Excitation Networks (SENet), Efficient Channel Attention (ECA) and Triplet Attention, are added to reduce the computational footprint of the model while boosting its ability to capture complicated disease patterns. Experiments were validated using publicly available datasets, such as the Plant Disease and PlantDoc datasets. Next, the performance of proposed RTR-Lite-MobileViT, obtained high results of accuracy (99.95%), precision (98.97%) in plant disease dataset and accuracy (99.96%), precision (97.98%) in Plantdoc dataset compared to existing model like MobileNetV2.
Layth Hussein (Wed,) studied this question.