A GAN-Augmented Multi-Scale CNN for Blight Detection and Classification in Solanaceous Crops

Plant disease is one of the most concerning problems in agriculture as it presents a very serious threat to crop production and global food security. The most common type of disease in solanaceous crops such as potato and tomato is the leaf blight disease which appears in different visual forms ranging from small brown spots to large lesions. Automated disease detection is still challenging because of lack of balanced, high-quality dataset and disease variation across plant species. To address these limitations, this research proposes Tri-ScaleNet, a unified framework that integrates a Conditional Generative Adversarial Network (cGAN) for data augmentation with a Multi-Scale Multi-Head Convolution Neural Network (MSMH-CNN) for disease classification. The cGAN generates realistic synthetic images by maintaining disease specific textures, lesion boundaries and colour variation. Then MSMH-CNN is used to extract features using parallel convolution kernels of different sizes, enabling the model to detect small scale spots as well as large decaying areas. Experimental finding reveals that our proposed method yields a strong performance with a validation accuracy of 98.62% across all classes. Additionally, we have used Grad-CAM visualization to demonstrate that model concentrates on disease-affected areas, which increases transparency and user trust by visualizing the decision processes, enhancing its effectiveness and potential usefulness in precision agriculture.