首页|Optimizing image-based deep learning for energy geoscience via an effortless end-to-end approach
Optimizing image-based deep learning for energy geoscience via an effortless end-to-end approach
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
The rapid growth of artificial intelligence (AI) technology and its applications in recent years has transformed the process of data analytics in many scientific fields, including geoscience. Geoscience has traditionally been a descriptive science and fundamentally relies upon visual recognition and identification of different geological features, from satellite images to subsurface seismic, to study Earth's history. Geological image data provides immense potential to apply advanced AI methods, such as deep learning to improve and optimize different geological and geophysical characterization workflows. Despite the increasing efforts and interest toward using AI in geosciences, its actual potential remains untapped, and further exploration is required. The prospect of AI application in geosciences is primarily hindered by the following: (i) limited availability of high-quality labeled datasets and (ii) inherited imbalance dataset distribution. These limitations are compounded by overexploitation of the transfer learning method to mitigate such issues, discarding the interpretability of the AI black-box problems. In this study, a robust and effortless strategy is proposed to overcome the limitations and simultaneously reduce our dependency on to the transfer learning method. Among the various methods available to mitigate these issues, only traditional data augmentation is heavily used in geosciences. This study, therefore, explored and developed a workflow by combining three readily available methods to maximize the performance of machine learning algorithms when dealing with a limited and imbalanced geoscience dataset. Here, the proposed method follows three robust and straightforward end-to-end steps: (i) combining traditional and advanced data augmentation (e.g., CutOut and CutMix) techniques to enhance localization and generalization performance; (ii) employing an algorithm-level class weight method to minimize detrimental impact and performance bias due to class imbalance; and (iii) applying a regularization label smoothing technique to improve the generalization and avoid overconfident prediction. Across the study datasets, the overall accuracy is typically improved up to 12% when comparing CNN without and with the proposed strategy. In addition, when combined with transfer learning, these methods could minimize model overfitting, and optimize generalization and model performance. This study further highlights that the proposed method should apply to different applications of AI in geosciences and could provide an alternative approach to the transfer learning method in analyzing limited and imbalanced geoscience datasets and improve the interpretability of how AI models work when combined with subject matter expertise.
Department of Geosciences, College of Petroleum Engineering and Geosciences, King Fahd University of Petroleum and Minerals, Dhahran, 31261, Saudi Arabia
NSTL
