Agrivision: Hybrid Deep Learning Models for Enhanced Plant Disease Detection

Plant diseases threaten agricultural sustainability, especially in low-resource environments where expert diagnosis is scarce. This study introduces AgriVision, a novel hybrid deep learning model that addresses critical limitations in existing plant disease detection systems through innovative architectural integration and advanced training strategies. Using the PlantDoc dataset comprising 2,598 images from non-laboratory conditions across 13 species and 17 disease classes, AgriVision uniquely combines SEResNet50 with modern training strategies including label smoothing, AdamW optimization, cosine annealing, and Stochastic Weight Averaging (SWA) to improve generalization while maintaining computational efficiency. The key innovation lies in our systematic integration of multiple complementary techniques with proper validation protocols-a combination rarely addressed comprehensively in existing literature. Unlike prior models that focus on individual architectural improvements, AgriVision incorporates a dedicated validation set to guide early stopping and optimize performance, addressing a critical gap in current research. AgriVision achieved a best test accuracy of 78.23% and a weighted F1 score of 78.17%, significantly outperforming baseline CNNs while maintaining computational efficiency with only 190 minutes training time. The model maintains an optimal balance between high accuracy and low resource consumption, making it suitable for field-deployable agricultural tools.