Oil Pool Detection Dataset

The Northern Ecuadorian Amazon (NEA) has suffered vast oil pollution in the last decades. Remote sensing is key to discovered contaminated areas and to provide local communities with tools to identify specific environmental and social impacts created by petroleum extraction. The use of remote sensing by local communities and governments to identify oil pollution in the NEA is limited by the intrinsic characteristics of the tropical forest and the lack of reliable and inexpensive data and methods. In this study, we focus on a type of pollution where oil waste is “hidden” in pools that are later covered by dirt and vegetation. We conducted a novel deep learning workflow combining object detection and semantic segmentation architectures based on enhanced topographic visualization images derived from UAV-LiDAR data to find buried oil pools left by petroleum companies during oil extraction in the Ecuadorian Amazon. Results show that the combination of DL architectures with LiDAR data achieved, on average, over 0.83 in precision, recall, F1 score, and Average Precision-AP (commonly used assessment metrics in deep learning). Our study reveals that this approach can accurately detect and delineate abandoned oil pools located under tropical forest, and thus provides an effective tool to monitor this type of oil liabilities, which represent a permanent risk for environmental and human health. The study has shown that deep learning architectures have potential for automatic detection of oil pollution features in terrestrial environments. Also, the use of unmanned aerial vehicle (UAV)-LiDAR systems has proved to be promising, highly-accurate tools for oil pollution discovery. Further research should now focus on the integration of other type of sensors, such as RADAR, with the ability to cover larger extensions of surface and the capacity to penetrate deeper into the forest canopy.