Is Deep Learning more effective in detecting natural oil slicks? A comparison of semi-automated and AI techniques

Marine pollution has been traditionally addressed in Earth Observation studies through the use of Synthetic Aperture Radar (SAR) imagery. In operational processes, the contrast between the dark oil surfaces, characterized by a low backscatter return, and the rough, bright sea surface with higher backscatter has been exploited for decades.

Many of the processing techniques involve the use of semi-automatic workflows. Dark spot segmentation and feature classification are, undisputedly, common computational tasks. However, effective discrimination between oil slicks and other ocean phenomena (e.g. biogenic slicks, wind streaks, greasy ice) remains a challenge. To complete this task, a trained human operator is often employed in the final validation step. Thus, the process is time and resource consuming over large expanses, while the results are highly subjective. Automating this process and reducing computation and analysis time is the ultimate goal.

New algorithms based on the use of artificial intelligence for oil slick detection have recently emerged. While there are studies proving their effectiveness in successfully segmenting and classifying oil slicks, questions regarding their operational feasibility remain. Do they improve the quality of the detection? Are they more capable of discriminating between the various dark formations? What are the computational and data resources required for training, validation, and deployment of such an algorithm?

This project focusses on the development of a new customized algorithm for natural oil slicks detection. As part of this development, we analyzed the state-of-the-art methods and performed a comparison of the latest deep learning methods and classic semi-automatic techniques. Here, we present an in-depth analysis of selected segmentation and convolutional neural networks algorithms and various frameworks. The primary objective is to evaluate their effectiveness and expose their deficiencies for the detection and classification of natural oil slicks against anthropogenic pollution and other dark formation caused by ‘look-alikes’.

This presentation centers on the results that have been obtained by utilizing high-resolution open SAR data acquired by the Copernicus Sentinel-1 satellites. The evaluation is based on study sites located in the Black Sea, where two known oil seepage areas are actively generating consistent productive slicks as well as underdeveloped oil traces.