Modernizing Medicinal Plant Recognition: A Deep Learning Perspective with Data Augmentation and Hybrid Learning

Abstract

This study proposes a deep learning-based solution to address the longstanding challenge of accurately identifying Indian medicinal plants, which are vital to Ayurvedic pharmaceutics but often misidentified due to their morphological similarities. The objective is to develop a reliable, automated classification system using image processing and advanced neural network architectures. A dataset of 5,945 images representing 40 distinct medicinal plant species was sourced from Kaggle and augmented to 11,890 images using techniques such as flipping, rotation, and scaling to enhance diversity. The models tested include a baseline Convolutional Neural Network (CNN), transfer learning with DenseNet121, DenseNet169, and DenseNet201, a voting ensemble of these DenseNet variants, and a hybrid DenseNet201-LSTM architecture. Experimental results show that the CNN model achieved the lowest accuracy at 69.58%, while the hybrid DenseNet201-LSTM model reached the highest validation accuracy of 93.38%, with a precision of 94.74%, recall of 93.38%, and F1-score of 93.42%. These findings confirm the hybrid model’s superior ability to capture spatial and sequential dependencies in leaf features. The novelty of this work lies in the integration of DenseNet201 with LSTM for medicinal plant classification, which has not been widely explored in this domain. The study also acknowledges dataset scalability as a limitation and proposes future work involving dataset expansion through botanical collaborations, integration of environmental metadata, and deployment of a mobile application using TensorFlow Lite for real-time, low-resource implementation. Overall, the research contributes a robust and scalable framework for medicinal plant identification, promoting trust in traditional medicine, supporting conservation efforts, and enabling practical field-level applications in both rural and clinical settings.