Abstract
OBJECTIVE: The agricultural sector is important in the supply of food globally as well as in enhancing the economy especially in developing countries where it forms the backbone of the economy. Corn can be classified as such crops important to the world’s food system. Unfortunately, corn crop is vulnerable to a lot of diseases and this may result into heavy losses and disruption in the food supply system. Hence, it is important to detect and classify these diseases accurately and promptly to limit losses and achieve the highest possible productivity. This study intends to solve these problems by constructing a trustworthy and interpretable model based on deep learning approaches focused on accurate identification of corn leaf disease. MATERIAL: In this study, the cumulative dataset comprises 4188 images which are further divided into four classes of corn leaf disease, with 1146 images of blight leaf, 1306 images of common rust, 574 images of gray spot and 1162 images of healthy leaves. In order to train and validate the model 70% of data was used for training while 30% was used for testing. This division was appropriate because it allowed enough data to be used during model training and also enough for model evaluation on new data. METHODS: The research employs ResNet152, a well-known deep leaning structure in image classification, because uses residual connections that improve the training of deep networks. Furthermore, Grad-CAM (Gradient-weighted Class Activation Mapping) is employed to improve the explainability of the model. Grad-CAM produces human interpretable visual images in the form of heatmaps and it indicates the areas of corn leaves that have had the greatest impact in the model, which is fairly useful in understanding the model. The model processes and predicts the corn leaves into four classes: healthy (H), blight (B), gray spot (GS) and common rust (CR), with precision and explainability. RESULTS: The results of training the ResNet152 model were remarkable as it registered a 99.95% accuracy during training as well as 98.34% during testing. Also, applying Grad-CAM for interpretability purposes proved to be useful as it created heatmaps that indicated the most important parts of the leaf images for making the model predictions. This added to the understanding of the model and its predictions, which was especially important for users such as farmers who required accurate diagnoses of diseases. CONCLUSION: This study demonstrates the effectiveness of the ResNet152 model, enhanced with Grad-CAM for explainability, in classifying corn leaf diseases. Achieving good training and testing accuracy, the model provides transparent, human-readable explanations, fostering trust and reliability in automated disease diagnosis and aiding farmers in making better-informed decisions to improve crop yields.