Terrestrial oil spill mapping using satellite earth observation and machine learning: A case study in South Sudan

Terrestrial oil spills are a major threat to environmental and human well-being. Rapid, accurate, and remote spatial assessment of oil contamination is critical to implementing countermeasures that prevent potentially lasting ecological damage and irreversible harm to local communities. Satellite remote sensing has been used to support such assessments in inaccessible regions, although mapping small terrestrial oil spills is challenging - partly due to the pixel size of remote sensing systems, but also due to the distinguishability of small oil spill areas from other land cover types. We assessed the usability of freely available Sentinel satellite images to map terrestrial oil spills with machine learning algorithms. Using two test sites in South Sudan, we demonstrated that information from the Sentinel-1 and-2 instruments can be used to map oil spills with more than 90 % classification accuracy. Classification accuracy was significantly increased (>95 %) with the addition of multi temporal information and spatial predictor variables that quantify proximity to oil production infrastructure such as pipelines and oil pads. The mapping of terrestrial oil spills with freely available Sentinel satellite images may thus represent an accurate and efficient means for the regular monitoring of oil-impacted areas.

Automated summary

This study demonstrated the usability of freely available Sentinel satellite images and machine learning algorithms for mapping terrestrial oil spills with more than 90% classification accuracy. The addition of multi temporal information and spatial predictor variables, quantifying proximity to oil production infrastructure, significantly increased the classification accuracy to over 95%. Freely available Sentinel satellite images may present an accurate and efficient means for the regular monitoring of oil-impacted areas.