Deep Learning-based Identification and Change Detection of Oil/Gas Well Pads Using Satellite Imagery

Abstract

Global energy extraction and industrial activities have caused widespread land disturbance, making large-scale and accurate environmental monitoring essential for ecosystem protection and land reclamation. Satellite remote sensing provides a powerful tool for monitoring these environmental impacts. In recent years, deep learning algorithms for remote sensing image interpretation have developed rapidly. However, significant gaps remain in bridging algorithmic developments with practical domain applications. In real-world scenarios, existing target extraction algorithms rarely consider the spatial context between targets and their surrounding environments, and optimal training constraints, such as loss functions, remain under-explored. To address these general challenges, this doctoral study develops a series of deep learning methods for target extraction, change detection, and land disturbance and reclamation evaluation, using oil/gas well pad monitoring as a primary application. The research follows a systematic progression: 1. Loss Function Optimization for Road Extraction: Linear feature (Road) extraction plays a foundational role across diverse remote sensing applications. However, how to select a suitable loss function for linear feature extraction is rarely studied. To address this, I conducted a comprehensive comparative study of 12 loss functions for road segmentation, showing that region-based loss formulations (e.g., Log-Cosh Dice and Squared Dice) significantly outperform distribution-based loss functions in maintaining the connectivity of linear features. 2. Well Pad Identification: To overcome the challenge that target extraction algorithms often ignore environmental context, I developed a modified Mask R-CNN network based on the coupled spatial relationship between roads and well pads. By incorporating the road network as a spatial prior, the proposed model effectively solves the spectral similarity problem between abandoned well pads and natural vegetation, improving average precision by over 20% compared with baseline methods. 3. Well Pad Change Detection: To evaluate post-disturbance land recovery and semantic transitions after mining development, I proposed a constrained dual-head HRNet architecture for semantic change detection. By employing a cosine similarity loss to constrain feature structures, the model achieves an 80.05% mIoU on the semantic change detection task. Moreover, to support these methodological advancements and promote benchmark evaluations, we constructed two comprehensive datasets—the Alberta Roads and Wells Dataset and the Alberta Semantic Change Detection dataset—using satellite imagery over oil sands regions in Alberta, Canada. Overall, this research advances the practical capabilities of deep learning in remote sensing by tackling foundational challenges in loss function selection, spatial context modelling, and semantic change analysis, providing a transferable and scalable methodology for broad environmental and geographic applications.