Foundation Model–Assisted Coffee Leaf Rust Severity Estimation
This repository accompanies the manuscript:
Foundation model–assisted segmentation enables robust field-based severity estimation of coffee leaf rust
This project presents a fully reproducible computer vision pipeline for quantitative estimation of coffee leaf rust (Hemileia vastatrix) severity under heterogeneous field conditions. The framework integrates object detection, lesion segmentation, pixel-based severity quantification, and concordance analysis grounded in phytopathometry principles.
The study compares classical image processing, supervised deep learning, and foundation segmentation models for lesion detection, and evaluates agreement with gold-standard pixel-level annotations using Lin’s Concordance Correlation Coefficient (LCCC).
Data Card Author(s)
Mary Paz Romero Benavides, Universidade Federal de Viçosa: Owner / Manager
Emerson M. Del Ponte, Universidade Federal de Viçosa: Contributor
Waldênia de Melo Moura, EPAMIG: Contributor
🌱 Project Overview
The methodological workflow consists of:
Leaf Detection – YOLOv8 trained using model-assisted annotations
Leaf Extraction – Detection-guided segmentation
Lesion Segmentation – Comparison of five approaches:
ImageJ thresholding
pliman (R package)
DeepLabV3+
Fine-tuned SAM2 (SAM_CLR)
Zero-shot SAM3
Severity Estimation – Pixel-based calculation:
S (%) = Diseased Area / Leaf Area × 100
Agreement Analysis – Lin’s Concordance Correlation Coefficient between predicted and reference severity
📊 Dataset Summary
The full dataset comprises:
1,285 field-acquired coffee leaf images
606 curated pixel-level rust lesion masks
100 independent evaluation masks
Roboflow dataset links:
📂 Repository Structure
📁 01_models
Contains documentation describing the trained models used in this study.
⚠️ Due to GitHub file size limitations, model weights are hosted on Hugging Face.
Models include:
YOLOv8 leaf detector
Fine-tuned SAM2 (SAM_CLR)
DeepLabV3+
Configuration used for zero-shot SAM3 inference
📁 02_binary_images
Contains validation binary masks (PNG format) corresponding to segmentation outputs from each evaluated method.
These masks were used to compute:
Intersection over Union (IoU)
Dice coefficient
Pixel accuracy
Precision
Recall
Disease severity (%)
Lin’s Concordance Correlation Coefficient (LCCC)
Binary mask format:
0 → background
255 → rust lesion
This folder enables independent verification of segmentation performance and severity calculations.
📁 03_analysis
Contains R scripts used to:
Compute severity metrics
Perform agreement and concordance analysis
Generate all figures included in the manuscript
Main R dependencies:
tidyverse
epiR
lme4
ggplot2
This folder reproduces the statistical analysis pipeline described in the paper.
🔬 Reproducibility
This repository provides:
Validation segmentation outputs
Statistical analysis scripts
Model documentation
External links to trained weights
Together, these components allow full reproducibility of segmentation metrics and severity agreement results reported in the manuscript.
🤖 Model Hosting
All trained model weights are hosted on Hugging Face:
This ensures accessibility without exceeding GitHub file size limitations.
📜 License
Code: MIT License
Binary masks and annotations: CC-BY 4.0
For questions or collaboration inquiries, please open an issue or contact the corresponding author.
