Dual-encoder Variational Autoencoder for Detection and Classification of Plant Leaf Diseases

Abstract

Background: Plant disease detection remains a major challenge in agriculture, with direct implications for improving crop productivity and ensuring food security. Seasonal variation significantly influences plant characteristics, making the classification of plant leaves by season-specifically summer and winter-important for optimizing disease detection and management strategies. Methods: In this study, a plant leaf disease detection dataset was developed and categorized based on seasonal conditions. The dataset includes 47 classes representing summer crops and 16 classes for winter crops. To classify plant leaf diseases effectively, we propose a novel dual-encoder Variational Autoencoder (VAE) model that integrates ResNet and VGGNet as parallel encoders. These encoders extract complementary feature maps from the seasonal datasets, which are then concatenated to improve classification accuracy. Result: Experimental evaluation demonstrates the robustness and accuracy of the proposed approach. The dual-encoder VAE achieved a classification accuracy of 98.86% on the summer dataset and 97.53% on the winter dataset, highlighting the model’s ability to generalize effectively across seasonal variations in plant leaf disease detection.

Code and data availability

The paper's plant leaf disease image inputs are publicly available Kaggle datasets explicitly cited as the sources for the summer and winter subsets (Sankalana plant-diseases-training-dataset, Gadde yellow vein mosaic, Kapadnis watermelon, Mir pumpkin). No authors' analysis code, trained model checkpoints, or paper-der