MFCIS: an automatic leaf-based identification pipeline for plant cultivars using deep learning and persistent...

Area of Science:

• Plant Science
• Computer Vision
• Bioinformatics

Background:

• Reliable plant cultivar recognition is crucial for plant breeding, germplasm resource innovation, and intellectual property protection.
• Existing leaf image-based identification methods are insufficient for distinguishing highly similar cultivars.
• There is a need for advanced computational approaches to automate and improve cultivar identification accuracy.

Purpose of the Study:

• To develop and evaluate an automatic leaf image-based cultivar identification pipeline (MFCIS).
• To combine topological signatures from persistent homology with high-level features from Convolutional Neural Networks (CNNs).
• To benchmark the MFCIS pipeline on fruit (sweet cherry) and annual crop (soybean) species.

Main Methods:

• Utilized persistent homology to extract topological features of leaf shape, texture, and venation.
• Employed a fine-tuned Xception network (a CNN) for high-level leaf image feature extraction.
• Implemented a score-level fusion strategy for combining identification models across different soybean growth periods.

Main Results:

• Achieved 83.52% mean accuracy for sweet cherry cultivar identification using over 5000 leaf images from 88 varieties.
• Reached 91.4% classification accuracy for soybean cultivars by fusing models from five growth periods, significantly outperforming single-period models.
• Demonstrated the pipeline's effectiveness across different plant types and complexities.

Conclusions:

• The MFCIS pipeline effectively integrates multi-feature analysis for robust plant cultivar identification.
• The proposed method significantly enhances identification accuracy compared to traditional approaches, especially for closely related cultivars.
• A user-friendly web service (http://www.mfcis.online) is available to facilitate the adoption of this technology in plant breeding.