Robust multiclass classification of crop leaf diseases using hybrid deep learning and Grad-CAM interpretability

Abstract

The key objective of this study is to propose an effective and accurate deep learning (DL) framework to detect and classify diseases in banana, cherry, and tomato leaves. The performance of multiple pre-trained models is compared against a newly presented model.The experiments used a publicly released dataset of healthy and unhealthy leaves from banana, cherry, and tomato plants. This dataset was uniformly split into training, validation, and test sets to obtain consistent and unbiased model evaluations. The data pre-processing also involved pre-processing steps suitable for DL architectures to keep the input the same among all the models.We use several state-of-the-art pre-trained ConvNets models for the baselines, such as EfficientNetV2, ConvNeXt, Swin Transformer, and Vi-Transformer (ViT), to have an outlook on the performance. A new ConvNet-ViT hybrid model combines the ConvNet and ViT layers for local feature extraction and maintaining the global context. The classifier's performance was reinforced by a 5-fold cross-validation mechanism to avoid overfitting.The proposed Hybrid ConvNet-ViT model outperformed all the compared models evaluated, achieving a testing classification accuracy of 99.29%, which outperforms all the pre-trained models. This finding shows that combining ConvNets' local feature learning with the capability of global representation of the ViT is effective.The result shows that the Hybrid ConvNet-ViT model is an effective and accurate solution in detecting and classifying plant leaf diseases. Its outstanding performance of the state-of-the-art pre-trained top models positions itself as a solid model for practical agricultural use. Fusing the ConvNet and transformer frameworks jointly is beneficial for improving classification performance in image-based disease detection work.

Keywords: Classification; ConvNet; Deep learning; Hybrid ConvNet-ViT; Plant leaf disease; Vision transformer.

Fig. 1

Overall architecture of proposed study.

Fig. 2

Sample dataset images of all diseases.

Fig. 3

Sample EfficientNetV2 architecture.

Fig. 4

Sample ConvNeXt architecture.

Fig. 5

Model architecture of Swin transformer.

Fig. 6

Sample ViT model architecture.

Fig. 7

Proposed hybrid ConvNet-ViT.

Algorithm 1

CNN-transformer hybrid classification.

Fig. 8

Performance metric comparison of proposed and pre-trained models.

Fig. 9

Validation accuracy and loss graph of proposed model.

Fig. 10

Confusion matrix generated from genuine test predictions for all nine classes across banana, cherry, and tomato leaf diseases.

Fig. 11

Grad-CAM visualization of 3-plant leaf diseases (a) BU and CPM, (b) TSL and TSM, (c) TTS and TTM, (d) TTY.

Fig. 12

Accuracy comparison of proposed and other SOTA models.