Oil Spill Detection Using Deep Neural Networks
Abstract
Oil spills have catastrophic effects on the environment, wildlife, economy, and human health. Therefore, timely detection of oil spills can reduce these disastrous impacts. Existing oil spill detection practices include in-situ (e.g., acoustic method, vapor sampling, pressure-point-analysis, and negative pressure wave) and remote sensing methods (e.g., traditional image processing and image processing using artificial intelligence). These methods rely mostly on skilled personnel for data collection, processing, and analysis, thus leading to slow, costly, and subjective results. Furthermore, oil platforms and pipelines are often situated in remote, harsh areas, making inspections hazardous. To remedy this problem, in this Thesis, three state-of-the-art artificial intelligence (AI) models, namely VGG16, YOLOv3 (you-only-look-once), and mask R-CNN (mask region-based convolutional neural network) are used in a transfer learning scheme to facilitate the process of detecting oil spills and surrounding objects such as vessels and oil rigs. Keyword search, a semi-supervised machine learning approach, is used to collect red-green-blue (R-G-B) imagery for training and testing these models. The methodology includes image classification, object detection, and instance segmentation. The VGG16 model is used to predict the existence of an oil spill in an image, yielding an accuracy of 93%. The YOLOv3 model is implemented to detect and mark the location of vessels and oil rigs. The mean average precision for detecting these two object classes is 61.5% (46% for vessel and 77% for oil rig). The mask R-CNN model is utilized to identify oil spill boundaries at the pixel level in the input image. Results (considering all test images) indicate an average precision of 62%, and an average recall of 71%. Findings of this Thesis are sought to benefit oil and gas industry stakeholders and coastal communities by creating operational AI-assisted technologies for timely detection and response to oil spills and other environmental pollutions, ultimately contributing to human health, environment preservation, and profitability of energy exploration projects.
Citation
Ghorbani, Zahra (2019). Oil Spill Detection Using Deep Neural Networks. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /189041.