An efficient dual-attention guided deep learning model with interpretability for identifying medicinal plants

Abstract

Medicinal plants are important because of their diverse benefits. However, the accurate identification of these plants poses a significant challenge to the healthcare, agriculture, and pharmaceutical industries. Visual similarities between species and environmental variations complicate this process. Although traditional deep learning (DL) and machine learning (ML) approaches have demonstrated promising results in classifying medicinal plants, the question remains as to whether a model can perform more effectively and multidimensionally, incorporating features such as a plain and real image background and lightweight design. This study introduced a dual-attention convolutional neural network based on the DenseNet121 model named ”DenseDANet,”. The dual attention mechanisms enhance classification accuracy and effectiveness. The model employs Local Interpretable Model-Agnostic Explanations (LIME) to improve transparency, thereby enabling reliable and explainable identification of medicinal plants. Furthermore, this model outperformed transformer-based models, including Swin-T, MaxVit-T, FastVit-MA36, Vit-B16, and deep learning convolutional neural networks (CNNs), such as VGG19, ResNet50, ConvNextV2-T, and DenseNet161. DenseDANet was trained and evaluated on two public datasets: DS1 (Bangladeshi Medicinal Plant Dataset) and DS2 (BDMediLeaves), collectively comprising original 7029 images from 20 classes. A 70:20:10 split was used for training, validation, and testing, respectively, achieving the highest test accuracy of 99.50%. The proposed model offers a lightweight, interpretable, and efficient method for identifying medicinal plants. It significantly benefits traditional medicine, pharmaceutical research, and biodiversity conservation through its accurate specifications, making it ideal for real-time applications and reducing computational costs.