Oil Spill Segmentation using Deep Encoder-Decoder models

Abhishek Ramanathapura Satyanarayana, Maruf A. Dhali

Paper available on ArXiv

• The results of this research is available in the ArXiv paper.

Info about the project

• This repo contains the project work done as part of the HTSM Masterwork at University of Groningen.
• Research work carried as part of the HTSM Masterwork to train deep learning CNN model for segmentation task to detect oil spills from the satellite Synthetic Aperture Radar (SAR) data.
• This research is towards application of AI for good, particularly towards conservation of nature with AI.
• This project was done under the supervision of Mr. Maruf A. Dhali

Dataset information

• The details about the dataset used in this Masterwork project can be found here - dataset details.
• This dataset contains labels for 5 classes --- sea surface, oil spill, oil spill look-alike, ship, and land.
• ThIs dataset is a relatively smaller dataset when compared to other popular benchmark datasets for the segmentation task.

Required dependencies for training

• To install pytorch use the following command

pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113

• The other required dependencies for training are available in requirements.txt.

Instructions to run the code for training

• For any python script, use the following to list all the command-line options

python3 script_name.py --help

• Run the script src/training/compute_stats_for_train.py to compute the statistics for training set. This will generate src/training/image_stats.json with the mean and std. dev. of images in the training set.
• To train the models, run the script src/training/train.py.
• To compute the k-fold validation metrics, run the script src/training/compute_kfold_validation_metrics.py.
• Run the notebook src/training/EDA.ipynb or the script src/training/exploratory_data_analysis.py to generate the plot of class distribution in the dataset.
• For plotting the variation of loss, accuracy, IoU during the training phase, use the notebook src/training/plot_graphs.ipynb.
• To infer on the test set, run the script src/training/inference.py.

Docker deployment instructions

• The streamlit app has been developed for deployment.
• The detailed python package requirements for the streamlit app can be found in src/requirements_deployment.txt.
• To build the container, run the following command inside src directory

docker build -t app_oil_spill .

• To the run the container, run the following command inside src directory

docker run -p 8000:8000 -t app_oil_spill

Huggingface deployment

• A streamlit application, with the best performing model, has been deployed to Huggingface

Qualitative results - sample test set predictions

• Since the dataset is not publicly available, the original test set images from the dataset are not uploaded but only their predictions are included in the repo.

Quantitative results

• The best model's class-wise and mean IoU performance is presented below.

class	class IoU (%)
sea surface	96.422
oil spill	61.549
oil spill look-alike	40.773
ship	33.378
land	92.218
mean IoU	64.868

Sphinx docstring generation

• The following are the steps to generate docstrings using sphinx
• Create a directory named docs and go to that directory

mkdir docs && cd docs

• Run the everyone of following commands inside docs directory
• Run the following command with appropriate options

sphinx-quickstart

• Open the file index.rst and add modules to it
• In the file conf.py, make the following changes
  • set html_theme = 'sphinx_rtd_theme'
  • set extensions = ["sphinx.ext.todo", "sphinx.ext.viewcode", "sphinx.ext.autodoc"]
  • add the following to the beginning of the conf.py file

  import os
  import sys
  
  sys.path.insert(0, os.path.abspath(".."))
  

• Run the following command

sphinx-apidoc -o . ..

• Create html files with documentation

make html

References

• HTSM Masterwork
• https://m4d.iti.gr/oil-spill-detection-dataset/