Affiliation:
1. Centro de Pesquisas, Desenvolvimento e Inovação da Petrobras CENPES
2. Laboratório de Engenharia de Software PUC-Rio
3. GeoSpatial Petroleum
Abstract
Abstract
The development and application of predictive models to distinguish seepage slicks from oil spills are challenging, since Synthetic Aperture Radar (SAR) systems detect these events as dark spots on the sea surface. Traditional Machine Learning (ML) has been used to discriminate the Oil Slick Source (OSS) as natural or anthropic assuming that the samples employed to train and test the models in the source domain (DS) follow the same statistical distribution of unknown samples to be predicted in the target domain (DT). When such assumptions are not held, Transfer Learning (TL) allows extracting knowledge from validated models to predict new samples. This research aims to apply well-trained and validated models developed in the Gulf of Mexico (GoM) to predict the OSS of 105 unknown seepage slicks detected in the Brazilian Equatorial Margin (BEM), employing TL.
To accomplish this, 26 geometric features extracted from 6,279 validated oil slick polygons were used to develop predictive models in the GoM, utilizing different ML algorithms: Artificial Neural Network, Random Forest, Linear Discriminant Analysis, Support Vector Machine, and Logistic Regression. The knowledge learned from these models was transferred to predict unknown samples employing Data Interpolation as a TL method. Since the seepage slicks were detected by different satellites in the DS (RADARSAT: RDS) and in the DT domains (RDS and Sentinel-1: SNT1), a deeper analysis was conducted to evaluate the effect of different SAR sensors and image beam modes (BM).
Predictions considering all SAR sensors did not overtake the global accuracy (GA) of 34.29%, due to the high divergence among seepage slicks detected by different sensors in the DS (RDS) and in the DT (RDS and SNT1) domains. As seen in the prediction results, GoM models were trained to recognize the OSS of samples detected by RSD (37.78%), not by SNT1 (13.33%). Analyses per RDS BM made difference, once 78.20% of the oil slicks used to build the models were detected by ScanSAR Narrow (SCN), and only 10.64% by Wide modes. Consequently, the GoM models were better trained to predict seepage slicks detected by SCN achieving GA of 58.82%, while using Wide modes only 10.26% of samples were correctly predicted. Detailing, the higher GA of 61.70% was obtained using the SCNA, since 51.51% of the SCN samples used for training the GoM models came from this BM.
Results suggested that there are similar geometric patterns between seepage slicks detected in the GoM and BEM, being possible to predict samples in distinct geographic regions when using compatible SAR sensors. This perspective allows saving time and budget to collect, validate and annotate new samples for training new models from scratch. This value-added approach contributes to minimizing geologic risks for oil generation and migration in offshore exploration frontiers.
Reference19 articles.
1. Oil spill detection by imaging radars: Challenges and pitfalls;Alpers;Remote Sens. Environ,2017
2. Sensors, Features, and Machine Learning for Oil Spill Detection and Monitoring: A Review;Al-Ruzouq;Remote Sens,2020
3. Arnold, A.; Nallapati, R.; Cohen, W.W.
A Comparative Study of Methods for Transductive Transfer Learning. In Proceedings of the Seventh IEEE International Conference on Data Mining Workshops, Omaha, NE, USA, 28–31 October 2007; pp. 77–82. DOI 10.1109/ICDMW.2007.109 Available online: https://ieeexplore.ieee.org/document/4476649 (accessed on 01 March 2022).
4. Oil spill detection by satellite remote sensing;Brekke;Remote Sens. Environ,2005
5. Ben-David, S.; Blitzer, J.; Crammer, K.; Pereira, F.
Analysis of representations for domain adaptation. Conference: Advances in Neural Information Processing Systems 19, In Proceedings of the Twentieth Annual Conference on Neural Information Processing Systems, Vancouver, BC, Canada, 4–7 December 2006.