Leaf diseases significantly affect the Maize productivity. Early and timely detection of these diseases help in reducing crop loss. It also ensures sustainable agricultural practices. Traditional examination of leaves by means of manual methods are protracted, subjective, and difficult to scale, especially when working on large agricultural fields. In the current work, a lightweight hybrid deep learning (DL) framework for maize leaf disease detection has been developed. The proposed framework is based on MobileNetV2 integrated with a spatial attention mechanism. The MobileNetV2 backbone allows effective feature extraction and the attention module helps in improving the model's ability to focus on the regions affected by disease. To enhance generalization and reduce overfitting, several training strategies are employed. These include MixUp augmentation, exponential moving average (EMA) and label smoothing. The evaluation of the model is performed on datasets with semi-realistic conditions with moderate background and illumination variability. A peak accuracy of approximately 97.45% was achieved by the proposed approach as confirmed by the evaluation results and exhibits stable convergence. These findings indicate that lightweight architectures, when synergistically combined with attention mechanisms and robust training strategies, can attain high accuracy. The architecture also preserves computational efficiency. This makes them suitable for real-world agricultural deployment.
Kumari et al. (Thu,) studied this question.