Creating a contour for the oil spill allows a qualitative and quantitative measurement of the pollution spread. The latter permits to better monitor the oil spill.

Proposed Methodology: Geodesic Active contour model

The idea is to implement a new loss function, which calculates the difference between the estimated geodesic ACM and the ground truth, and then feed the loss function into the UNet network to optimize the results.

Dataset:

The dataset used for this project was acquired from MKLab. The research group of this data have been contacted for data request and it has been searched privately. The dataset comprises of jpg images extracted from satellite Synthetic Aperture Radar (SAR) data collected from European Space Agency (ESA) databases, the Copernicus Open Access Hub. The downloaded SAR images were acquired using the Sentinel-1 European Satellite. The required geographic coordinates and time of the confirmed oil spills were provided by the European Maritime Safety Agency (EMSA) based on the CleanSeaNet service and its records covering a period from 28/09/2015 up to 31/10/2017. The size of the dataset is 2000 images for training and 220 images for testing. The dataset has two types of labelling, mask and 1D, which allows for flexibility in training different types of networks. Additionally, the dataset does not label oil spill only, but also sea surface, “look-alike” ship, and land.

UNET:

UNET architecture is implemented and trained on the dataset described above. This network consists of two paths; encoder and decoder. The encoder consists of convolution and pooling layers, and the decoder uses transposed convolutions. The network is implemented using Keras Python library.

Adam is used as an optimization function, with a learning rate of 0.01. Intersection over Union (IoU), dice coefficient, and loss are used as evaluation metrics to assess the performance of the network. The figures below demonstrate the performance of the training and testing accuracy according to the aforementioned metrics. The performance is decent, but there is room for more improvement, which will be the next step in this project.

Sample result from UNET algorithm for oil spill detection
Performance metrics in terms of IoU, dice coefficient, and model loss

Duration:

6 months (April 2020 – October 2020)

Team Members:

Deep CNN for Oil Spill semantic segmentation in SAR Images