Oil Spill Identification from Satellite Images Using Deep Neural Networks

Abstract

Synthetic aperture radar (SAR) sensors are commonly used for this objective due to their capability for operating efficiently regardless of the weather and illumination conditions. Black spots probably related to oil spills can be clearly captured by SAR sensors, but it is hard to discriminate from look-alikes. Several methods have been proposed to automatically detect and classify these dark spots. Most of them employ custom-made datasets that don´t generalize well. Others assign a single label to the entier SAR image, resulting in simplistic informations. To overcome these limitations, semantic segmentation with CNNs is proposed as an efficient approach. Also, a public dataset of SAR images is introduced, aiming to consist on a benchmark for future oil spill detection methods. This dataset is employed to review the performance of well-known DCNN segmentation models. DeepLabv3+ presented the best performance, in terms of test set accuracy and related inference time.

Introduction

Sea oil pollution is considered a major threat to oceanic and coastal ecosystems, as well as for various naval-related human activities. Accidents at offshore oil drilling platforms or oil pipeline networks can provoke severe oil spills.

Synthetic aperture radar (SAR) mounted on aircrafts or satellites is the most common sensory equipment in marine remote sensing systems. the SAR sensor is a microwave-based technology, that emits radio wave pulses and receives their reflection in order to capture a representation of the target scene, widely known as SAR images. The sensor is considered an ideal option due to its robustness against weather and illumination changes. Oil spreading over the sea surface dampens the capillary waves, and so the backscatter radio waves are suppressed. As such, oil spills are depicted as black spots, contrary to the brighter regions which are usually related with unspoiled polluted sea areas. Also, the wide coverage of the sensor is of relevant importance, since oil spills could cover kilometers, as well as further contextual information, such as close-coastal regions or vessels, which can be enclosed in the acquired image.

The detection can be challenging, since oceanic natural phenomena such as low wind speed regions, weed beds and algae blooms, wave shadows behind land, grease ice, etc. can also be depicted as dark spots, resulting in possible false positives look-alikes.

Automated detection of oil spills commonly involve three main steps: 1) automatic detection of dark spots in the processed SAR image, 2) feature extraction from the initially identified regions, and finally the 3) classification as oil slick or regions including look-alikes. In the first step, binary segmentation is usually applied to the input image representation in order to retrieve the depicted black spots. The second phase involves the extraction of statistical features from the aforementioned segmented areas that might include potential oil spills. In the last phase, the whole image and/or the segmented region is classifyied as oil spill or look-alike (binary classification).

To help the classification, some authors propose a probabilistic approach to distinguish between oil spills and look alikes. Also, wind history information was also used to enhance the oil spill recognition system. Although the classification process is automated, an extensive pre-processing phase is required in order to initially extract geographic features. Multi-level segmentation schemes were also used to reduce the false positives rate.

Several works proposed the use of neural networks for the feature extraction and classification of oil spills. Although performative, these approaches heavily rely on pre-computed features. To this end, genetic algorithms were used to identify the optimal subset of the extracted features and the optimal number of nodes in the network's hidden layer.

Due to its capabilities in monitoring wide territories, SAR sensors can provide rich contextual information, about shores, vessels and oil platforms. This information can be useful for a more meaningful detection and alert system, while also help the distinction between oil spills and look-alikes, e.g., it is expected that a dark spot with linear formation close to a ship might correspond to an oil spill discharged from the vessel, rather than a look-alike. Moreover, information about nearby coastal territories or ships os important for an early warning system and a decision making module towards mitigating the overall danger. This represents a paradigm shift, where a different classification approach is required in order to identify properly multi-class instances enclosed in SAR imagery. Also, given the physical properties and events related to oil slicks, handcrafted features are not enough to represent the possible geometric shapes and volumetric sizes of oil spills affected by wind speed, sea currents, etc. Taking this into consideration, among with the existence of multi-class instances, semantic segmentation models could be deployed as robust alternatives to extract the rich informative content from SAR images. Even though several works explored this technique, they weren't able to fully leverage its capabilities efficiently, due to the lack of data, or simplistic approach, where only oil spill and look-alikes were classified. Only one study expanded this technique by also classifying land and ship instances.

The use of custom-made datasets also made the comparison of techniques irrelevant. To this end, the authors present a new publicly available oil spill dataset, aiming to establish a benchmark dataset for the evaluation of future oil spill detection algorithms.