UWSpace is currently experiencing technical difficulties resulting from its recent migration to a new version of its software. These technical issues are not affecting the submission and browse features of the site. UWaterloo community members may continue submitting items to UWSpace. We apologize for the inconvenience, and are actively working to resolve these technical issues.
 

Land Disturbance Extraction in Alberta Oil Sands Satellite Imagery

Thumbnail Image

Files

Hu_Bingxu.pdf (5.22 MB)

Date

2021-12-10

Authors

Hu, Bingxu

Journal Title

Journal ISSN

Volume Title

Publisher

University of Waterloo

Abstract

Tracking and mapping land disturbances from oil and gas development are critical to environmental assessments and land reclamation. Manual extraction of land disturbances is costly, time-consuming, and requires updating given rapid development. Through this study, a methodology and model for oil and gas land disturbance extraction along with an error-correcting algorithm solving rural mapping deficiencies is proposed. Chapter 3 titled “Deep-learning Extraction of Land Disturbances Arising from Oil and Gas Development” explores and finds an optimal strategy for land disturbance extraction methodology. Outlining the 3 findings in Chapter 3 for land disturbance extraction: (1) road and wellpad extraction should be integrated into a single task, (2) land disturbance extraction task should be segmented between forest and farmland backgrounds, and (3) RGB outperforms NDVI in land disturbance extraction.Chapter 4 titled “Maintaining Accurate Maps of Rural Land Disturbances: A Deep-Learning Automatic Change Detection Algorithm” introduces and tests the proposed error-correcting algorithm and explores its hyperparameters. Results in Chapter 4 show that the proposed automated error-correcting algorithm improves performance by 8.3% − 15.4% compared to baseline. Key findings in Chapter 4 explore how hyperparameters affect model performance: (1) alpha and beta in the AEC algorithm need to be carefully selected, (2) careful selection of alpha and beta can reduce the number of transitory artifacts introduced, (3) running the AEC algorithm a few times can greatly improve model performance, (4) adding a threshold to when the AEC algorithm begins stabilizes model performance. Combining findings in Chapters 3 and 4, an accurate fully automatic alternative to manual mapping for oil and gas land disturbance extraction is proposed. The combined model proposed in Chapter 4 is a ready solution to tracking, mapping, and managing land disturbances in the Alberta oil sands for purposes of environmental assessments and land reclamation.

Description

Keywords

deep learning, cnn, oil sands, error correction, land disturbances

LC Keywords

Citation

URI

http://hdl.handle.net/10012/17750

Collections

Theses
Geography and Environmental Management