Spatial modeling and uncertainty analysis for subsurface feature mapping : integration of geostatistical concepts and image-based machine learning model validation



Journal Title

Journal ISSN

Volume Title



Spatial modeling of subsurface features and uncertainty analysis plays a pivotal role in the integration of data analytics and machine learning techniques in the petroleum industry. As the energy landscape is always changing, and new technologies are emerging, the demand for accurate assessments of uncertainty to inform high-value decision-making is of utmost importance. Nonetheless, the same longstanding methods are used due to their simplicity and the lack of immediate necessity for change. However, with improvements and the implementation of proper workflows, the current methods for calculating uncertainties and validating machine learning models can be more effectively addressed. We developed multiscale methods for data analytics and machine learning. These approaches integrate geostatistical concepts to enhance the precision and reliability of subsurface modeling techniques. We address the challenge of integrating multiple datasets with varying accuracies and volume support sizes. We emphasize the importance of accounting for different sources of uncertainty in spatial modeling workflows. Leveraging geostatistical concepts, such as semivariograms, and dispersion variance, a novel approach is introduced to calculate a more precise measure of error when imputing smaller scale datasets with larger scale datasets. This refined measure of error allows for the direct integration of these datasets in spatial modeling workflows. Once all the uncertainty in our models is accounted for, we must check if our models are accurate. Therefore, we focus on the validation of machine learning models, particularly those tailored for image data. Image-based models often necessitate pre-processing steps, such as resizing and augmentation, to improve data quality for training. To ensure the performance and suitability of these models for real-world datasets, proper validation techniques are imperative. We propose integrating the concept of minimum acceptance criteria with the multi-scale Structural Similarity Index (MS-SSIM) for improved model checking. This enables a more accurate evaluation of model performance in reproducing original images and predicting new ones, surpassing conventional approaches such as mean squared error (MSE) and single-scale SSIM. Our multiscale approaches for data analytics and machine learning establish a comprehensive framework for addressing uncertainty and validating image-based models. The incorporation of geostatistical principles in calculating uncertainty and proper selection criteria for image-based model validation are showcased on subsurface data; however, they are versatile and applicable across various domains. Ultimately, they contribute to the safe and effective deployment of machine learning models for spatial modeling, advancing the field towards more reliable and informed decision-making.


LCSH Subject Headings