Plant classification with Convolution Neural Network

1. Usage

   • Instalation
   • Prepare data
   • Configuration
   • Train model
   • Display results

2. Technology

3. Data

4. Tasks

5. Problems

6. Classifiaction

7. Results

Usage

Instalation

Install python (https://www.python.org/downloads/)

If you have GPU, you can install CUDA (https://developer.nvidia.com/cuda-downloads)

git clone https://github.com/jedrzej-put/plant_classification.git
cd plant_classification
python -m venv venv
.\venv\Scripts\Activate.ps1
pip install requirements.txt

Prepare data

First way:

Create data directory(in plant_classification) with subdirectories "healthy" and "sick" which include apppriate images. Check configs/base.yaml (fields healthy_data_path and sick_data_path).

Second way:

Set absolute paths to "healthy" and "sick" directories in configs/base.yaml (fields healthy_data_path and sick_data_path)

Configuration

You can edit:

• Hyperparameters
• Scheduler parameters
• Transformations
• Data path
• Tensorboard and MLflow parameter

To use mlflow you have to set environment variable "MLFLOW_TRACKING_URI"

EXPORT MLFLOW_TRACKING_URI=<YOUR_MLFLOW_TRACKING_URI>

Train model

python scripts/classification-script.py

Display results

Check "./logs" directory. Find directory start with your experiment name and choose your experiment(if you two or more experiments with the same name, the log path to second and subsequent experiments will have a postfix with the experiment number). Run "tensorboard --logdir logs/<experiment_name><experiment_number>"

e.g.

tensorboard --logdir logs/test_experiment

tensorboard --logdir logs/test_experiment2

Open http://localhost:6006/

Technology

• python 3.10.4
• torch==2.1.1
• torchmetrics==1.2.1
• torchvision==0.16.1
• pillow==9.0.1
• tensorboard
• mlflow
• scikit-learn

Data

• The original data includes two plants:
  • rye
    • healthy
    • sick, infected with Rust
  • beet
    • healthy
    • sick infected with Cercospora beticola

Tasks

• loading data, cleaning data
• random under sampling
• adding and compare approach with preprocessing step with own trained Unet(preparation mask with Label Studio):
  • segmentation infected leaves
  • segmentation spots of diseases
• inferencing segmentaion model
• building pipelines with preprocessing step
• building CNN model
• training, validating and testing CNN model
• saving metrics and model
• performing experiments

Problems

• small number of images therefore sometimes only a few images on the test
• unbalanced datasets
• various data source
• various data distribution - main reason to create preprocessing with segmantation spots of disease

Classifiaction

experiment	plant	dataset	use preprocessing	healthy	sick	val acc	val precision	val recall	test acc	test precision	test recall
1	rye	1	no	25	95	0.854	0.823	0.86	0.833	0.75	0.85
2	rye	2	no	118	129	0.875	0.846	0.917	0.917	0.917	0.917
3	rye	3	no	93	34	0.75	0.75	0.75	0.75	0.75	0.75
4	rye	1	yes	25	95	0.92	0.96	0.85	0.88	0.84	0.86
5	rye	2	yes	118	129	0.917	0.95	0.833	0.917	0.92	0.833
6	rye	3	yes	93	34	0.875	0.95	0.75	0.625	0.6	0.75
7	beet	4	no	155	130	0.846	0.8	0.923	0.692	0.692	0.692
8	beet	5	no	155	74	0.75	0.667	0.688	0.688	0.636	0.875
9	beet	4	yes	155	130	0.846	0.85	0.692	0.731	0.8	0.615
10	beet	5	yes	155	74	0.688	0.667	0.75	0.688	0.615	0.62

Results

Accuracy

Precision

Recall

F1 score

Loss