MFCIS

MFCIS integrates persistent homology and convolutional neural networks to extract multiscale topological and high-level features from leaf images for cultivar identification and morphological analysis.

Key Features:

• Integration of Persistent Homology and Deep Learning: Combines persistent homology with a convolutional neural network (CNN) to extract topological signatures of leaf shape, texture, and venation across multiple scales.
• Advanced CNN Architecture: Uses a fine-tuned Xception network to extract high-level image features that help discriminate subtle differences among cultivars.
• Application Across Species and Growth Stages: Applied to fruit species (sweet cherry, Prunus avium L.) and annual crops (soybean, Glycine max L. Merr.), accommodating datasets with thousands of leaf images from numerous cultivars or breeding lines.
• Score-Level Fusion Strategy: Implements score-level fusion by combining outputs from independently trained models for each growth stage to improve classification accuracy.

Scientific Applications:

• Plant Breeding and Germplasm Innovation: Facilitates cultivar identification and germplasm protection to support breeding programs and intellectual property management.
• Research and Development: Enables analysis of genetic diversity and morphological variation to inform crop improvement and variety characterization.

Methodology:

Processes leaf images using persistent homology to extract topological features, applies a CNN (Xception) for high-level feature extraction, employs score-level fusion across growth-stage models, and benchmarks performance on large datasets (mean accuracies reported as 83.52% for sweet cherry and 91.4% for soybean).