Generalising Across Different Crop Diseases With InceptionCNN

Abstract: This Plant diseases have a significant impact on the agriculture of the world and lead to decreased crop productivity, food quality, and economic productivity as a result. Following is an automated pathogen detection system for plant diseases using an Inception-type convolutional neural network (InceptionCNN) in PyTorch. The model was trained on an original Kaggle-based New Plant Dataset containing 38 classes of healthy and diseased plant leaf images. We performed the appropriate preprocessing (duplicate deletion, stratified data splitting, normalization, and augmentations) to improve generalization and robustness of our models. We offer a multi-scale architecture of multi- scale convolutional branches utilizing 1×1, 3×3, and 5×5 filters among others to select a collection of lesion patterns and disease structures. Adam optimization, scheduling of the learning rate, dropout regularization, and class imbalance reduction were utilized as the training method. Experimental results demonstrate optimal performance with 99.24% validation accuracy and minimal validation loss. These evaluation metrics involved measured performance accuracy, recall, F1 score, confusion matrix analysis, and Grad-CAM visualizations which helped enable consistent classification of the disease under the categories. Our proposed framework represents a scalable, accurate and reproducible approach to the intelligent plant disease detection and lays the groundwork to future AI-related precision agriculture applications.

Keywords: Plant disease detection; Convolutional Neural Network; Deep Learning; Precision Agriculture; Plant Protection