CISCS: Classification of inter-class similarity based medicinal plant species groups with machine learning

Area of Science:

• Botany and Pharmacognosy
• Computer Science and Machine Learning
• Bioinformatics

Background:

• Reliable classification of medicinal plants is crucial for healthcare quality and safety.
• Existing methods struggle with visually similar species and imbalanced datasets.
• Deep learning models like ResNet18 and VGG16 show limitations due to overfitting.

Purpose of the Study:

• To develop a robust and computationally efficient model for classifying Indian medicinal plant species.
• To address challenges of high inter-class visual similarity and dataset imbalance.
• To improve the accuracy and reliability of plant species identification.

Main Methods:

• A novel multi-level feature fusion model combining 3D normalized color histograms, extended uniform Local Binary Patterns (LBP), Gabor filters, and Histogram of Oriented Gradients (HOG).
• SMOTE-based synthetic augmentation to address class imbalance.
• A soft-voting ensemble of machine learning classifiers with cosine similarity metrics for classification.

Main Results:

• The proposed model achieved 100% accuracy in Group 1 and 95.82% in Group 3 on Indian medicinal plant datasets.
• Consistently outperformed deep learning baselines, with over 90% accuracy in other groups.
• Demonstrated robustness in conditions of high inter-class similarity and dataset imbalance.

Conclusions:

• The multi-level feature fusion model provides a superior alternative to deep learning for medicinal plant classification.
• The approach is computationally efficient and scalable, supporting biodiversity and ecological studies.
• This method enhances the reliability of plant identification, crucial for pharmaceutical and conservation efforts.