Plant Ensemble Model: A Multi-Architecture Framework for Plant Disease Classification Using Convolutional Neural Networks and Transformer Models

Abstract

Crop diseases remain a critical concern in agriculture, threatening food supply and productivity, and therefore require effective detection techniques. In the past few years, there have been significant advances in using deep learning to automatically classify plant diseases. However, most previous models were built using a single model, which does not generalize very well to different environments. For multi-class plant disease classification, this study suggests an ensemble model, a feature-level heterogeneous ensemble of four ImageNet pre-trained deep learning backbones: Swin-Tiny (hierarchical vision transformer), EfficientNetV2-S (progressive convolutional neural network), DenseNet-169 (densely connected network), and ConvNext-Tiny (modernized large-kernel convolutional neural network). Each of the four backbones independently produces a global feature vector from the input image; each of these four vectors will be concatenated into one large 4480-dimensional joint feature representation and then classified by a three-layer fusion head that uses batch normalization and Kaiming-initialized weights. Two popular datasets, PlantVillage, which represents controlled laboratory conditions, and PlantDoc, which represents actual agricultural environments, were used to assess the model. Results from the experiments reveal that the ensemble model achieves near-perfect performance on PlantVillage (99.81%) and has excellent generalization to the more difficult PlantDoc with 77.54% accuracy. Overall, these outcomes reveal the effectiveness and the current limitations of using ensemble deep neural networks for practical plant disease diagnosis, which will be useful for the deployment of precision agriculture technology.