Visualizing a Marine Oil Spill with Sentinel-2 MSI Imagery

Marine oil pollution is everywhere. Sad but true. While the world has not had a disaster since the BP’s Deepwater oil spill in the Gulf of Mexico, mini oil spills and slicks are still happening. This article has more information on the variety and quantity of oil slicks around the world’s oceans.

I will focus on visualizing and quantifying one such oil spill that was reported it started on August 10, 2017, off the southern coast of Kuwait in the Persian Gulf. The cause was not reported however it is speculated that the source was from an oil pipe leak.

Datasets

Luckily, there were little to no clouds during the days of the oil slick and two Sentinel-2 level 1C datasets were used. One dataset was captured 1 day after the report date (Aug. 11) and the other dataset that was downloaded was captured 5 days after the report date (Aug 16) to track its movement.

In SNAP, the images were downsampled to 20m to calculate band ratios between bands that are at 10 m and 20 m resolution later in the analysis.

After resampling, the images were Rayleigh corrected to produce surface reflectance (Level 1C images are top of atmosphere reflectance). The input parameters — sea level pressure in hPA, and estimated ozone thickness of the atmosphere in DU (Dobson Units) — were found at Columbia Climate School International Research Institute for Climate and Society Monthly Sea Level Pressure tool and at the Government of Canada’s Ozone Archive for the month of August 2017.

For August 16, two subsets of separate images were mosaicked to show the extent of the oil spill.

Images

Figure 1 shows how the oil appears in Sentinel-2 image after the pre-processing. The varying colours of the oil demonstrate the different oil thickness.

Figure 1: True Colour Image of the oil spill in on August 11, 2017

However, the problem with optical imagery is that it is sometimes hard to distinguish oil on the surface of the water because of meteorological and sensor conditions. Take a look at Figure 2A, the mosaicked true colour image on the 16th. The ocean appears dark near the western coast and light coloured at the northern coast due to sediment discharge from the Khawr Abd Allah estuary. No trace of oil is visible like it was on August 11 (Figure 1A) but it’s there. Oil doesn’t disappear in 3–4 days.

Figure 2A: True Colour Imagery of the coast of Kuwait.

This is where band ratios come in. Band ratios help to amplify the differences between spectral reflectance of the different objects on the surface. Sentinel-2 MSI has 13 bands ranging from optical, near infrared (NIR) to shortwave infrared (SWIR). The bands that are used for oil slick detection are:

• B2 (blue)
• B3 (green)
• B4 (red)
• B8 (NIR)
• B11 & B12 (SWIR)

Oil has a high reflectance in the near infrared. This is helpful because at the NIR wavelength, oil and water are easily distinguishable since water has a low reflectance at this wavelength. Oil also has a low reflectance in blue, green, and red portions of the spectrum which helps to distinguish it from algae which has a high reflectance in the blue and green portions.

Many band ratios were created by Rajendran et al. for an oil spill off the coast of Mauritius in 2021. For this oil spill, I used a different combination of band ratios to display the most visual distinction between oil and water. They are:

• R: B3/B2,
  B: (B3+B4)/B2,
  G: (B11+B12)/B8.

When the ratios are applied, the oil appears on August 16, as seen in Figure 2B, 2C, 2D.

Figure 2: False Colour Composite of the August 16 Image with 2 subsets (C and D)

We see that the oil slick spread north and if you zoom in on Figure 2D, you can see boats helping to clean it up.

Using the same band ratio combination, the oil appears more visible on the Aug 11 image, as seen below.

Figure 3: False Colour Composite of the August 11 Image

Comparisons Side by Side

So you don’t have to scroll…

I find these images very cool! But the best band ratio combination is dependent on the location of the spill and the type of oil spilled which unfortunately does not make it a fast repeatable process creating an element of trial and error.

But band ratios do make optical imagery more useful for the detection and classification of oil spills since they help to visualize the oil better, with more contrast.

Classification Results

I ran a unsupervised classification (K-means Clustering Analysis) to
extract the oil pixels in the Aug 16 composite to estimate the surface area polluted by the oil. The classification was only run on a polygon of interest that I created around the oil to reduce complexity and unnecessary clusters like land, clouds etc. Ten clusters seemed to yield the best classification result that visually matched the oil slick in the false colour composite.

The area estimate is 83 km². Although there was no way of quantively evaluating the classification, the number of false negatives seen in Figure 4A and false positives in Figure 4B seem relatively equal making the area estimate a balanced estimate.

Figure 4: K-means classification result for August 16, 2017

Supervised classification was not used because of the lack of true ground data to use as training and testing data. Creating training and testing data points from the Sentinel-2 imagery to train a classifier was, also, not feasible due to the small sample size of visible oil pixels. :/

Object-based classifications and segmentation algorithms may produce a better result reducing the number of false negatives within an oil slick, as seen in Figure 4A since it takes into consideration the neighbourhood around the pixels which can indicate the regions and better capture the
shape of the oil slicks.

Conclusion

Thank you for joining me on this process of visualizing this oil spill with band ratios to emphasize the difference between water, oil, and algae, testing the methods of Rajendran et al. (2021), and attempting to quantify an oil spill with a simple k-means classification.