BarleyDiseaseSegmentation

A fully reproducible deep-learning pipeline for brown rust, ramularia, and multiclass segmentation.

This repository provides a unified, end-to-end system for automated segmentation and phenotyping of barley leaf diseases. It supports:

• A quick demo inference
• Fully automated figure reproduction from the associated manuscript
• A complete training pipeline with two-phase HPO, encoder selection, retraining, and inference

Reference Paper:

1. Environment Setup

1.1 Clone Repository

git clone https://github.com/grimmlab/barley-disease-segmentation.git
cd barley-disease-segmentation

1.2 Install Environment

Docker Container

docker build -t barley_disease_segmentation_image .
docker run -it \
    --gpus '"device=0"' \
    --volume $(pwd):/workspace \
    barley_disease_segmentation_image

Local Installation

pip install -e .

Optional: edit barley_disease_segmentation>config.py to customise paths

2. Dataset Setup

Download the following three folders from Mendeley Data, DOI: doi.org/10.17632/4ny92p2r8f.1:

• hpo_data/
• inference_data/
• data_patches/

Place them in the repository root.

3. Usage Options

You can use the repository in three main ways, depending on your goal.

3A. Quick Start Inference

Run a minimal demo using default settings.

python Inference_quick_start.py \
    --encoder convnext_tiny \
    --task multiclass \
    --test_data_path Sample_test_data \
    --run_leaf_evaluation \
    --leaf_evaluation_output quick_start_results

Output: quick_start_results/ containing predictions, leaf-level visualisations, and evaluation metrics.

3B. Figure Reproduction

Reproduce all figures from the manuscript.

cd Figure\ Reproduction/
python Reproduce_figures.py

The figures are in the respective folders in Figure Reproduction

3C. Complete Pipeline

A full workflow: 2-phase HPO → encoder selection → retraining → inference.

Supported segmentation tasks:

• binary_ram
• binary_rust
• multiclass

Run Pipeline

cd Complete\ Pipeline/

python Complete_pipeline.py \
    --task multiclass \
    --skip-hpo

If --skip-hpo is omitted, the pipeline performs broad + refined Optuna HPO (~6 days on a single GPU).

---

Pipeline Arguments (Reference Table)

Argument	Description	Default
--task {binary_ram, binary_rust, multiclass}	Segmentation task (required)	—
--skip-hpo	Use existing HPO results instead of running Optuna	False
--trials N	Number of HPO trials (if HPO active)	60
--epochs N	Epochs for final retraining	300
--experiment-name	MLflow experiment name	None
--dry-run	Print commands without executing	False

The pipeline automatically executes:

• Broad-phase HPO
• Refined HPO
• Encoder benchmarking
• Final retraining with best hyperparameters
• Leaf-level inference and evaluation

---

4. Citing This Work

This work is currently under review:

Deep Learning–Based Identification of Visually Similar Foliar Diseases in Field-Grown Barley  
Sofia Martello, Nikita Genze & Dominik G. Grimm