If you are using this code in your own project, please cite at least one of our relevant papers:

@inproceedings{maroun2022machine,
  title={Machine Learning Using Support Vector Regression in Radar Remote Sensing for Oil-Spill Thickness Estimation},
  author={Maroun, Charbel Bou and Daou, Georges and Hammoud, Bassel and Hammoud, Bilal},
  booktitle={2021 18th European Radar Conference (EuRAD)},
  pages={221--224},
  year={2022},
  organization={IEEE}
}

@inproceedings{hammoud2022artificial,
  title={Artificial Neural Networks-Based Radar Remote Sensing to Estimate Geographical Information during Oil-Spills},
  author={Hammoud, Bilal and Maroun, Charbel Bou and Ney, Jonas and Wehn, Norbert},
  booktitle={2022 European Signal Processing Conference (EUSIPCO). Piscataway, NJ, USA: IEEE},
  year={2022}
}

Oil Spill Thickness Estimation

A proof of concept for using machine learning with radar data from a UAV hovering over a body of water to detect the presence of oil and estimate its thickness. For proper assessment of an oil spill scenario, thickness measurements need to range from 1 to 10 mm.

Machine learning models

We tested 4 different approaches for oil spill detection and thickness estimation:
SVR
Neural Networks
U-Net Model
Cascaded U-Net Model

In summary:

The SVR and ANN models are point-based models, they perform well when the oil thickness is uniform over the environment.
We used the ANN model to estimate the oil thickness and the oil permittivity.
The U-Net models perform way better since they take into account the spatial correlation within the spill.
SVR and ANN have lower complexity compared to U-Net models making onboard monitoring and processing of oil spill data feasible.
For the U-Net based approach, the environment has to be fully scanned and then processed.

Here are some metrics for the 4 machine learning models:

Model	IoU	Dice	Precision	Recall
SVR	0.29	0.43	0.43	0.51
ANN	0.36	0.49	0.49	0.56
U-Net	0.75	0.86	0.85	0.87
Cascaded U-Net	0.82	0.89	0.90	0.90

Regarding the cascaded U-Net model, here is a diagram showcasing its architecture:

Data

Due to the scarcity of oil spill data, especially for radars operating in wide-band ranges, we developed an oil spill simulator to model the spill. The developed model reflects essential properties of the real world and provides us with realistic oil spill distributions (More info in the github repo).

The second part takes care of populating the simulated environment with radar reflectivities using a Monte-Carlo simulation by introducing the additive white Gaussian noise and roughness loss to the computed reflectivities based on input frequencies. To generate the needed data, the data generator module is used. It utilizes the relative dielectric constant module along with the noise and the export to file modules.

Run

All .py files used to train and evaluate the models are stored in the run directory. The segmentation directory contains all run files related to the segmentation models.

noinget / oil-spill-thickness-estimation Goto Github PK