Oil Spill Detection in the North Sea Using Landsat-8/9 Images and Deep Learning

Kurzfassung

Oil pollution of seas and oceans poses a danger to human health and has a big impact on the marine and coastal environment. Oil enters the water from various sources which are of natural (47 %) and anthropogenic origin (53 %). The release of natural seeps is steady but slow over time which allows the ecosystems to adapt. The most common anthropogenic cases are accidents in maritime transportation, on oil platforms or the deliberate discharges of oil from ships where large amounts of oil can be released into the water within a short time. As indicated by data from the German Central Command for Maritime Emergencies, the number of pollution incidents identified in 2023 exhibited a notable increase in the time spent by naval pilots detecting pollution from the air. Consequently, pollution was detected on average every 7.7 flight hours, a significant deviation from the previous average of 12.5 to 20 flight hours observed since 2009. The global risks of the maritime safety and security have increased significantly due to Russia's utilisation of a very outdated tanker fleet. With the advantage of wide coverage, remote sensing can be used for a timely and accurate oil spill detection and help to prevent pollution spread and support clean-up operations to minimize the negative impacts on the environment as well as to identify the polluter. This study presents two different deep learning methods for the detection of oil spills on multispectral optical satellite images acquired from the Landsat-8 and Landsat-9 satellites. The North Sea was chosen as the study area due to the fact that the satellite data is transmitted in direct downlink mode to the DLR ground station, which allows the operational use of the developed methods in a near real-time application as a remote objective. The two deep learning methods used in this study are a (fully connected) deep neural network (DNN) and a convolutional neural network (CNN) in the type of a U-Net architecture. Both networks utilize different techniques for handling training data in order to train the model. A DNN is trained pixel by pixel ignoring the spatial component while the U-Net is able to use the context information and the localization at the same time. The training dataset used consists of training image patches containing oil spills and corresponding binary oil masks. The training images were generated by applying an atmospheric correction and calculating three independent indices based on specific spectral bands. These indices are the Normalised Difference Oil Index (NDOI), the Green-Shortwave Infrared Index (G-SWIR) and the CaBGS index. The oil masks were labelled manually using a segmentation method based on the indices mentioned above and the coastal aerosol band. A precise oil mask is crucial to train a robust model, as any pixel not marked as an oil pixel (pixel value 1) is automatically considered as a non-oil pixel (pixel value 0) including pixels of the land surface, clouds, cloud shadows and no-data values (black border outside the Landsat image). These pixel values show a high variability and they would negatively influence the model training. Utilizing this training dataset, the models were then trained to recognise and classify patterns of oil spills against the complex background of marine and coastal environment. Consequently, the performance and effectiveness of the models was evaluated using precision, recall and F1-score and their efficiency was demonstrated on different datasets. The result of the study has indicated that the proposed methodology is, in principle, an effective approach. The performance of the DNN is quite good with respect to the used test data while the U-Net indicates a less favourable performance outcome. With regard to the performance of the models on some selected entire Landsat image frames, the DNN has demonstrated a tendency to overestimate regions that are supposed to be oil, particularly in areas along the coastline. In contrast, the U-Net has demonstrated superior performance in terms of accuracy but it has tended to underestimate the extend of the spilled oil. The model demonstrated enhanced performance when the oil slicks exhibit a broad shape. The reason for that could be the exclusion of patches with less than 5 % oil pixels from the training process and with this the exclusion of oil spills of narrow and small-scale structures. The current performance is constrained by the restricted scope of the training data, particularly the lack of diversity in the oil types and the inconsistency in weather conditions under which the data were acquired.