Automated recognition of oil and gas production infrastructure using satellite imagery

Recent studies have highlighted the need to identify oil and gas (O&G) facilities to connect anonymous aerial or satellite emissions surveys to specific facilities in the production basin. This study proposes a novel deep learning architecture that processes high-resolution (30 cm GSD) 3-band, pansharpened, visible-spectrum (RGB) satellite imagery from Maxar to identify facility outlines and key on-site equipment. The architecture utilizes a dual-branch model paradigm that combines few convolutional layers and Adaptive Fourier Neural Operators (AFNO) within a Transformer-based structure. One branch focuses on detecting two major equipment types at the site, while the other branch identifies overall facility boundaries. Comparison to test data indicates our architecture achieves 93% accuracy for facility identification with as few as 260 positive training samples. Comparing resulting facility data to regulatory reporting indicates that satellite deep-learning–based detection identifies more facilities, with greater detail and specificity than current reporting programs. To implement widespread anonymous aerial or satellite sampling, this type of facility recognition is likely required to attribute emission detections to O&G facilities.