Automated Detection of Oil/Gas Well Sites Detection from Multi-Source High Spatial Resolution Images

With the development of oil/gas production, its adverse impact has drawn much attention. Therefore, automated detecting oil/gas well sites become important. Current research has been focused on detection using sole source RGB images, which was not the common case in remote sensing. In this study, we explored the use of a combination of the Residual Channel Attention Network (RCAN) and the state-of-the-art object detection method, You Only Look Once (YOLO) v4, to detect oil/gas well sites from multi-sensor images. To testify the feasibility of the combination, we selected 18 RapidEye and 15 WorldView images which cover the oil sands area in Alberta, Canada. We applied a pre-trained RCAN to unify the spatial resolution of different images to 2m/pixel. To maximize the feature space, we preserved 5 bands which were available in both images. YOLO v4 was applied on cropped image patches to detect oil/gas well sites. The experiment results showed that using the framework proposed in this study, oil/gas well sites can be localized accurately although the bounding boxes of the sites may not perfectly align with the objects.