Using a Convolutional Siamese Network for Image-Based Plant Species Identification with Small Datasets.

Area of Science:

• Computer Vision
• Machine Learning
• Botany

Background:

• Deep learning models often require large datasets, limiting their application in areas with scarce data.
• Few-shot learning, one-shot learning, and Siamese networks offer solutions for data-limited scenarios.
• Accurate plant species identification is crucial for biodiversity monitoring and ecological studies.

Purpose of the Study:

• To propose and evaluate a convolutional Siamese network (CSN) for plant species identification using leaf images, particularly in low-data conditions.
• To compare the performance of the CSN against a traditional convolutional neural network (CNN).
• To investigate the impact of dataset size on the accuracy of both CSN and CNN models.

Main Methods:

• A convolutional Siamese network (CSN) was developed to learn a similarity metric for discriminating plant species from leaf images.
• The CSN was trained using pairs of similar and dissimilar leaf images, minimizing Euclidean distance for similar pairs and maximizing it for dissimilar pairs.
• Experiments involved varying dataset sizes (5-30 images/species) from the FLAVIA dataset and testing on unseen data, including species not present in the training set.

Main Results:

• The CSN demonstrated strong performance in classifying new plant species with limited available images.
• Accuracy comparisons between CSN and CNN were conducted across different dataset sizes to identify performance thresholds.
• The CSN showed potential for generalizing its learned similarity function to entirely new species beyond the training set.

Conclusions:

• Convolutional Siamese networks offer a viable solution for plant species identification when deep learning datasets are small.
• The CSN's ability to generalize to novel species highlights its effectiveness in few-shot learning for botanical image analysis.
• Further research can explore optimal dataset size thresholds for CSN versus CNN performance in plant identification tasks.