Look-Alike Detection using Deep Convolutional Neural Networks for Ocean Oil Spills

Abstract

The marine and coastal ecosystems are placed in grave danger whenever there is a spill of oil. For the sake of protecting future generations, the authorities need a dependable mechanism that allows them to respond quickly and effectively to oil spills. As a result of their consistency also effectiveness in a wide series of weather and light situations, synthetic aperture radar (SAR) sensors are frequently utilized for this determination. The dark spots that are typically associated with oil spills remain easy for SAR sensors to spot, but it may be difficult to differentiate them from other objects or phenomena. A great number of methods have been outlined in order to automatically differentiate and categories these dark spots. The results are frequently unable to be compared with one another because of the heterogeneous nature of the data collected. It is often difficult to fine-tune settings or extract meaningful information because SAR images remain often classified through a solo label that is applicable to the entire picture. Because of this, the process of dealing with the images is made more difficult. The Random Forest Classifier and deep convolutional neural networks (also known as DCNNs) are examples of approaches that have been suggested as potential workarounds for these constraints. In addition, we make available to the public a search-and-rescue photo library that might serve as a standard for the growth of upcoming oil spill recognition tools. On the dataset that was provided, a number of well-known DCNN segmentation models as well as the Random Forest technique are evaluated and compared with one another. When compared to other approaches, it was discovered that Random Forest had the best test correctness and the quickest implication time. Additionally, the dataset is used to explore and explain the challenges of the provided issue, which in this case is identifying between genuine and fabricated oil spills. These investigations and demonstrations are carried out with the help of the dataset. According to their training and evaluation on the dataset that was provided to them, DCNN segmentation models, when paired with a Random Forest classifier, did quite well when attempting to detect oil spills. The novel method is anticipated to be of significant use in subsequent research pertaining to the identification of oil spills and the dispensation of SAR images.