A machine-learning-based approach to predict potential oil sites: Conceptual framework and experimental evaluation

Farouk Ferjani; Tarek Sboui; Moncef Ben Smida

doi:10.1515/geo-2022-0714

Article Open Access

A machine-learning-based approach to predict potential oil sites: Conceptual framework and experimental evaluation

Farouk Ferjani , Tarek Sboui and Moncef Ben Smida

Published/Copyright: December 10, 2025

Published by

Become an author with De Gruyter Brill

Submit Manuscript Author Information

From the journal Open Geosciences Volume 17 Issue 1

Abstract

Petroleum exploration requires careful planning due to high risks, expenses, and time consumption. Predicting oil sites is the preliminary step in petroleum exploration. Oil site prediction can be optimized by machine learning (ML) algorithms and techniques. This article highlights the issue of optimizing oil site prediction, which has not been focused on in existing research studies. The article proposes an approach that aims to improve the prediction of potential oil sites. The proposed approach is based on a conceptual framework and an assessment of geological criteria. The criteria are assessed and validated based on a sensitivity analysis. The framework is based on convolutional neural networks (CNNs), a subsidiary of ML, and image processing techniques. It consists of four main steps, which are data preparation, semantic segmentation, model training, and validation. The framework is implemented in a case study in the Fort MacKay region of Alberta, Canada. The implementation of the proposed approach shows promising accuracy for predicting potential oil sites, with a 95% intersection between the predicted potential zone and the actual petroleum site. The main contribution of the proposed approach is a sensitivity assessment of geologic criteria and a CNN-based model to enhance the prediction of potential petroleum sites.

Keywords: oil site prediction; machine learning; convolutional neural network; image processing; sensitivity analysis

Abbreviations

ML: Machine learning
CNN: Convolutional neural network
AI: Artificial intelligence
RNN: Recurrent neural network
GPR: Gaussian process regression
SVM: Support vector machine
IR: Infrared
UAV: Unmanned aerial vehicle
LSU: Linear spectral unmixing
SAM: Spectral angle matching
AGS: Alberta Geological Survey
OLI: Operational Land Imager
SAM: Segment-anything-model
ROI: Region of interest
CVAT: Computer Vision Annotation Tool
IoU: Intersection over union
N: Total number of wells
WP: Well potential
wL: Surface lineaments weights
wM: Aeromagnetic value weights
wSM: Lineaments and aeromagnetic values combines weight
IL: Well intersections with surface lineaments
IM: Well intersections with aeromagnetic survey
ISM: Well intersections with lineaments and aeromagnetic survey combined

1 Introduction

Upstream in the oil industry is composed of two main parts: exploration and production. The exploration phase consists of predicting and locating potential oil-bearing sites or “prospects” suitable for extraction [1]. Oil site prediction can be optimized by machine learning (ML) algorithms and techniques. The work presented in this article aims to improve the prediction of potential oil sites.

In recent years, the oil industry has seen major technological developments. With the first explorations mainly relying on surface indicators such as natural oil seeps way back in the nineteenth century, the industry has evolved to include multi-dimensional methods and techniques to locate and extract oil [1,2]. These techniques are essential in every modern endeavor. Geologic surveys, aeromagnetic and gravity surveys, seismic surveys, geochemical surveys, and remote sensing are the core techniques used in the exploration phase [2,3,4,5]. With the development of new techniques and tools came the increase in costs and management.

Oil exploration expenditures can lead to high costs. For instance, exploration of offshore operations in the Grand Banks of Newfoundland Canada reached $4.3 billion from the years 2009 to 2018 [6]. The cost of an unsuccessful exploration can range from 5 to 100 million dollars, contingent on the tools employed and whether it is an onshore or offshore exploration, rendering it a highly risky endeavor [2].

Advances in ML techniques have the potential to optimize oil exploration and prediction. ML, which is a subsidiary of artificial intelligence (AI), has offered new ways of data acquisition, processing, and interpretation. Its advantages include its ability to analyze vast amounts of data and facilitate interpretations. Another advantage is its flexibility in terms of applicability, making it deployable for many tasks including different phenomena classifications and predictions [7,8,9,10].

Many studies attest to the ability of ML in oil exploration, optimizing the upstream phase and increasing its efficiency.

The application of ML algorithms within geological disciplines has witnessed significant advancements. Various studies have explored the utility of ML in classifying data derived from well logging [11,12,13,14], as well as forecasting construction parameters and uncertainties in tunnel geology [15].

In terms of upstream applications, studies include interpolation and gradient boosting techniques bundled with deep learning to evaluate subsurface geology [16], accelerating interpretation of seismic cubes using deep learning [17], and automating geological feature detection in 3D seismic data using semi-supervised learning [18].

Other studies have reviewed prominent methods for AI applications in critical areas such as seismic data processing, facies and lithofacies classification, and the prediction of key petrophysical properties (e.g., porosity, permeability, and water saturation) [19].

With regard to the optimization of reservoir management, a convolutional-recurrent neural network (CNN-RNN) model was used to predict well-by-well oil and water rates based on time-varying well bottom-hole pressure schedules [20]. Another study employs a deep learning CNN model to automate the tracking of seismic horizons. These horizons help build geologic structures and stratigraphy models. The use of a CNN model alleviates the challenging problem of tracking seismic horizons in 3D seismic data [21]. In production, Gaussian process regression (GPR), CNN, and support vector machine were used to predict shale gas production and optimize parameters for multistage fractured horizontal wells. Porosity, permeability, fracture half-length, and horizontal well length were the key factors in the prediction output [22].

In terms of oil prospecting, ML-based approaches have shown credible ways of optimizing certain procedures. A study by De Kerf et al. [23] employed infrared (IR) camera images captured by an unmanned aerial vehicle drone in the port of Antwerp in Belgium, and an ML model that automatically detects oil spills. Oil exhibits higher IR emissivity than water facilitating its distinction from different water bodies. The study used a multitude of CNN segmentation architectures and feature extractors to determine which combination best suits the required task. Tian [24] proposed a hyperspectral remote sensing approach grounded in hydrocarbon microseepage theory. Validating their methodology in the Qaidam Basin and Liaodong Bay, the study demonstrated the direct identification of oil reservoirs and indirect detection using linear spectral unmixing and spectral angle matching. Another study by Huang et al. [25] focuses on identifying geological faults in 3D seismic volumes using CNNs. By tackling the challenge of analyzing extensive seismic data using deep learning, they created a “Big Data” platform that showed promising accuracy and potential in improving hydrocarbon prospect identification.

While the existing works have shown promising results in automatically predicting oil sites, they have based their research on unique criteria, mainly oil spills [26,27,28] or geologic faults. Solely basing on oil spills can limit the prediction accuracy, as the absorption band of hydrocarbons can intertwine with other minerals such as calcite. The choice of additional geological criteria such as geologic faults is key in oil prediction. Integrating other geological parameters could enhance accuracy. Our research proposes additional criteria in order to enhance the prediction aspect of petroleum prediction. It aims to optimize cost expenses and reduce its overall time frame by automating the prediction of potential prospects.

The main contribution of this study is the use of relevant geologic criteria to predict potential petroleum sites which are surface lineaments and aeromagnetic surveys. This limited set of criteria does not intend to be exhaustive but rather to provide a basis to forecast oil sites. We propose an ML-based framework that uses geologic criteria to forecast zones where the highest potential petroleum sites reside. The framework centers around a CNN model. CNNs are adept in their ability to automatically generate a comprehensive set of features from original geospatial images, minimizing reliance on human intervention for the computation and fine-tuning of attributes. Classification is done through semantic segmentation, allowing it to analyze and learn efficiently. This method enhances its predictive accuracy and overall output quality.

Another contribution of our study research is the evaluation of the model’s performance through a sensitivity assessment of the oil-predicting criteria. The criteria and their combination are evaluated, and their contribution percentage is measured in a sensitivity analysis to illustrate which criteria are more adept at predicting oil sites.

The novelty of this study lies in the combination of different aspects, notably the prediction of potential oil sites, the definition and the assessment of geological criteria, the sensitivity analysis, and the use of an ML CNN model in order to enhance the accuracy of the model results.

In the next section, we present our proposed method to enhance the prediction of oil sites. Our approach consists of a CNN-based framework and the assessment of a set of criteria to detect potential oil sites. In Section 3, we present the study zone. Following that is Section 4, where a prototype is developed to implement our approach. In Section 5, we showcase results and validation assessments. Finally, Section 6 discusses and concludes the overall study while providing suggestions for future endeavors.

2 Methods

The framework involves several processes that should be implemented to forecast oil sites: data preparation, identifying and assessing criteria, and model training and validation. The framework is depicted in Figure 1.

Figure 1

ML-based framework for predicting oil sites.

We acquire and preprocess datasets to prepare for the model training phase, as shown in Figure 3. Once the images have undergone preprocessing and are ready for training and validation, we define a CNN model for carrying out petroleum predictions. The CNN model we implemented is based on Python Libraries Keras and TensorFlow.

Given the intricacies of segmenting ever-changing satellite images of the Earth’s surface, we choose the CNN U-net convolutional neural network model. U-net was initially tailored for biomedical image segmentation. It has proven its adaptability to diverse biological elements. Notably, U-net’s architecture lends itself effectively to topographic landscapes, harnessing the inherent variability of such environments [29,30].

For model development, we used the Google Colab platform as it is equipped with essential tools and libraries like TensorFlow and PyTorch, which are important assets for ML endeavors. The implementation of Google Colab significantly enhances computational performance and renders it a highly suitable platform for executing this project [31].

Subsequently, we transition to training the CNN U-net model using the prepared images and their corresponding masks. After ensuring the model is adequately trained with the appropriate number of epochs, we proceed to validate it. The validation process involves assessing the prediction accuracy between the model’s predictions set and its respective validation set. Additionally, we validate the model’s predictions based on the presence of potential petroleum sites, considering the geologic criteria we propose. Simultaneously, we provide a sensitivity analysis for these criteria.

2.1 Proposed criteria

In order to train our proposed model, we assessed two criteria that are surface lineaments and aeromagnetic surveys. Geologically significant linear features observed on the Earth’s surface are commonly referred to as surface lineaments. These lineaments can assume different shapes and sizes and are associated with geological structures such as faults, folds, rivers, and joints. They are typically identified through visual observation of aerial photography or by utilizing remote sensing techniques. Surface lineaments hold importance from geological, structural, and geomorphological perspectives. In the context of petroleum exploration, surface lineaments can serve as valuable indicators of potential hydrocarbon reservoirs. Consistent correlation between surface lineaments and subsurface oil and gas traps has been demonstrated by studies [32,33]. By examining the orientations of different surface lineaments, it is possible to identify and allocate different fault systems to distinct reactivation timelines, providing insights into the history of a structural zone. This information is particularly useful for understanding migration channels, which constitute hydrocarbon pathways, and improving our knowledge of hydrocarbon migration [33]. The integration of surface lineament analysis with other geophysical and geological data sets enhances the accuracy and reliability of petroleum exploration efforts. It provides valuable information for identifying potential drilling targets and optimizing exploration strategies. The combination of surface lineament analysis with techniques such as seismic surveys, magnetic surveys, and subsurface mapping enables a more comprehensive evaluation of potential hydrocarbon reservoirs [34]. Figure 2 shows the surface lineaments across the province of Alberta in Canada.

Figure 2

Surface lineaments in Alberta, Canada.

Complementing surface lineaments, aeromagnetic surveys emerge as a highly precise technique for identifying potential petroleum-rich sites. This method involves equipping an aerial vehicle with a magnetometer and scanning a designated region to measure its magnetic response. This surveying method offers comprehensive coverage, making it particularly effective for studying vast areas. By detecting and utilizing the Earth’s magnetic properties, aeromagnetic surveys provide valuable insights into the subsurface. The magnetic properties of different rocks and the variations in the magnetic field can be indicative of fault systems, fracture zones, and other structural features that play a role in the presence and migration of hydrocarbons. Aeromagnetic surveys enable the mapping of magnetic properties, which form the basis for the interpretation and understanding of the subsurface geology. Moreover, aeromagnetic surveys are effective in detecting hydrocarbon traps such as salt domes and igneous intrusions, which generate magnetic anomalies, providing valuable information about the presence of hydrocarbon traps and serving as indicators of potential petroleum reservoirs [34,35]. Figure 3 presents a map of aeromagnetic survey values in the province of Alberta, Canada.

Figure 3

Aeromagnetic survey of Alberta, Canada.

3 Study area

Our study zone is based on the Fort MacKay region in Alberta, Canada. This area has been the focus of many subsurface investigative techniques, namely seismic and aeromagnetic surveys. We gathered images of this zone from the Landsat Satellite alongside geoscientific datasets. We employ geoscience datasets primarily obtained from the Alberta Geological Survey (AGS) website. The Fort MacKay zone is strategically situated within the Athabasca Oil Sands, hosting significant reserves of non-conventional oil, comparable to global conventional reserves. Specifically, these reserves are embedded in the McMurray formation, functioning as a reservoir rock with shale, sandstone, and oil-impregnated sands developed during the Cretaceous period. Figure 4 shows the geology of the Fort MacKay region in Alberta, Canada. The geographical distribution of the McMurray Formation spans a landscape distinguished by valleys and ridges [36].

Figure 4

Geology of Alberta, Canada, in the region of Fort MacKay.

4 Implementation

4.1 Data acquisition

Satellite images of the study zone from the Landsat 8 OLI were procured for this study. Landsat 8, a renowned satellite, offers global coverage with orbits completed approximately every 99 min at an altitude of 705 km [37,38]. Leveraging Landsat 8 satellite imagery, this project harnesses the technological advancements and historical depth of the Landsat program, enabling the acquisition of multispectral data at varying resolutions, global coverage, and frequent revisits for comprehensive monitoring and analysis. Figure 5 showcases a multispectral image of the study zone.

Figure 5

Multispectral image of the study zone: Fort MacKay region of Alberta, Canada.

Subsequently, extracting the surface lineaments involved digitizing and georeferencing structural features from various geological reports into shapefile format. We perform a sequence of preprocessing steps on the surface lineaments and aeromagnetic survey datasets. This preparatory phase ensures that the data have been edited to fit our study scope and range, as shown in Figures 6 and 7.

Figure 6

Surface lineaments in the study zone: Fort MacKay.

Figure 7

Aeromagnetic survey values in Fort MacKay expressed in nanotesla (nT).

Following this preparatory phase, we performed an intersection between the surface lineaments dataset and the areas exhibiting high values in the aeromagnetic dataset. The outcome is an integrated image that effectively combines petroleum potential indicators derived from both datasets. This composite image is then processed further by employing the Segment Anything Model (SAM) in Figure 8, after which a binary mask is generated as shown in Figure 9.

Figure 8

Generated composite image highlighted in the region of interest) by SAM.

Figure 9

Generated binary image mask generated using Computer Vision Annotation Tool (CVAT).

4.2 Model training and evaluation

With the image and its corresponding mask prepared, the following phase involves training a ML model. To implement this process, we use the Python library “patchify” to process the division of the image and mask into multiple patches. By performing this data augmentation technique, we enhance the model’s training by providing a diverse dataset, a crucial component to enhance ML model performance. Then, we divide the data into training and validation sets to help train and validate the model output. To do this, we employ the “train_test_split” function from the SciKit library. The data is divided into 20% for validation and 80% for training the model. In our study, the U-net model is selected, utilizing Adam as the optimizer and Imagenet weights as the encoder weights. A total of 250 epochs or iterations are configured to ensure the model achieves the desired level of accuracy. To assess the model’s efficacy, we employ the intersection over union (IoU) evaluation metric. In the context of semantic segmentation, accuracy is most effectively quantified by measuring the degree of overlap between the model’s predicted image and the ground truth mask image at the pixel level. This assessment is commonly expressed through the IoU calculation. IoU provides a more comprehensive evaluation of the model’s predictive accuracy as it evaluates the alignment of each pixel in the predicted image with its corresponding pixel in the mask image [39].

5 Results

Once the model is fully trained and prepared for predictions, we proceed to the next step by selecting specific image patches from the validation set. Subsequently, we employ the model to make predictions on these chosen image patches. The model’s output is then compared to the corresponding mask patches from the validation set using the IoU metric, allowing for a thorough assessment of its predictive accuracy. The results can be seen in Figure 10.

Figure 10

Representation of study zone image patches (left column), corresponding training masks (middle column) and model predicted images (right column).

Once the model has provided its predictions, we proceed to validate the results. Validation focuses on two aspects: the model’s accuracy and the appropriateness of selected criteria. For our model validation, we assess the accuracy of the predictions by comparing them with the set of images predefined for validation. As for the criteria validation, we measure the number of intersections between the criteria and the well locations in the study zone.

5.1 Model accuracy validation

The results in Figure 10 showcase a visual representation of the model’s capabilities. Each image and mask patch shown in Figure 10 is a 2 by 2 cm image. The level of overlapping (IoU) between the mask patches and model outputs are measured to indicate how accurate the model predictions are. The IoU of the first row indicates an overlap percentage of 74.36, and 84.35% for the second row. However, the third row exhibited a lower overlap percentage of 53.80%. Overall, the model showcases decent result values on resized image patches. Various factors contribute to the inaccuracies in the predictions.

5.2 Criteria accuracy validation

With the model being operational validation of petroleum existence in the predicted zone remains unconfirmed. That is why we introduced an additional dataset for validation presented in Figure 11. The dataset displays annotated core descriptions and geophysical logs from the oil sands areas in Alberta where Alberta Geological Survey staff and contractors have logged core from the early 1990s to the late 2000s. This supplementary dataset serves as a valuable resource for cross-verifying the model’s predictions against real-world geological data, providing a more comprehensive validation process. Surface lineaments and Aeromagnetic survey values located in the study zone are intersected with well locations in the study zone. The result is used in the sensitivity analysis of the criteria.

Figure 11

Contour map with added well locations dataset in the study zone of Fort MacKay.

Individually, each criterion exhibits a contribution percentage toward potential petroleum. This percentage can be calculated using the location of the wells dataset for validation. Table 1 shows the contribution percentage of each criterion. Surface lineaments intersected with 26 out of 101 wells in the study zone, marking a 25.7% intersection percentage. High aeromagnetic survey values intersected with 41 wells in the study zone, exhibiting a 40.6% intersection percentage.

Table 1

Contribution percentage of each parameter

Parameter	Well intersections	Intersection percentage
Lineaments	26 out of 101	25.7%
High aeromagnetic survey values	41 out of 101	40.6%

To illustrate each parameter’s potential, as well as their combined potential, we propose a weighted average formula. The formula presents a different perspective to oil site prediction, in which it quantifies the intersections of the geoscientific datasets with the total number of wells N, and offers a numerical value that describes the potential for an oil-producing well site noted WP (well potential). In our study, the weights for surface lineaments (wL) and aeromagnetic values (wM) are 1/4. Their combined weight (wSM) is 1/2. Well intersections with surface lineaments are notated IL, IM for aeromagnetic values and ISM for their combined intersections count. The well potential for these two geoscientific parameters came out to be 33.4%

(1) WP = ( wL ⁎ IL + wM ⁎ IM + wSM ⁎ ISM ) /N .

Figure 12 depicts the well locations dataset which is modified via geospatial tools to display a concentrated zone of the well density in the area. Subsequently, SAM is then utilized to create a binary mask of the well density based on this data which is presented in Figure 13. We then calculate the IoU between the well density image (Figure 13) and the potential petroleum binary mask (Figure 9). The result is a 95% intersection as shown in Figure 14.

Figure 12

Well locations dataset in the region of Fort MacKay.

Figure 13

Well density binary image mask in the study zone.

Figure 14

Overlap between well density and generated composite image.

6 Discussion

This analysis integrates the use of a geological map in identifying a highly favorable region for petroleum exploration. By leveraging the valuable insights provided by the geological map, alongside surface lineaments, aeromagnetic survey, and well location datasets, a comprehensive assessment of the study area’s petroleum potential was achieved. The geological map plays a pivotal role in understanding the distribution and characteristics of geological formations, particularly the significant McMurray Formation, which serves as the primary reservoir rock for bitumen in Alberta. By analyzing the geological map, valuable information regarding the presence and extent of the McMurray Formation, as well as other potential reservoir rocks, was obtained. In combination with the geological map, the analysis considered surface lineaments and aeromagnetic survey parameters to further refine the identification of the petroleum-rich region. The surface lineaments, representing linear geological features, and the aeromagnetic survey parameters, providing insights into subsurface geological structures, were crucial in narrowing down the areas with high petroleum potential within the geological map. The integration of these three parameters resulted in the identification of a compelling region for petroleum exploration, demonstrating a remarkable 95% intersection with the well locations density map within the study area. This strong correlation between the identified petroleum rich zone and the distribution of existing wells confirms the effectiveness of the methodology employed and provides valuable guidance for future exploration efforts.

With regards to previous works, De Kerf et al.’s result of 89% accuracy was established on a local scale (i.e., port of Antwerp in Belgium) with IR as the sole oil identification parameter [23]. Our 95% accuracy result spans a bigger area (i.e., Fort McMurray region of Alberta, Canada) using three different oil-identifying parameters (i.e., geological map, surface lineaments and aeromagnetic surveys).

As for Tian’s approach, it uses hyperspectral remote sensing, which can vary from region to another in terms of availability [24]. On the contrary, our approach uses multispectral remote sensing which is largely available and has global coverage. We should note however that hyperspectral imagery offers higher resolution than multispectral imagery.

We expect that our obtained accuracy results could be further enhanced through modifying our experimental system. We could incorporate the use of hyperspectral imagery and integrate additional oil-identifying parameters.

7 Conclusion

The combination of remote sensing techniques and ML algorithms holds potential to be effective in optimizing the exploration process. In this work, we proposed and implemented an ML-based approach to predict potential oil sites. In the implementation, we gathered crucial visual data for analysis based on Landsat 8 satellite images. Additionally, geoscience data, including geological maps, lineaments, and aeromagnetic survey data, were integrated into the study to identify potential petroleum sites.

The specific study zone was the province of Alberta, known for its extensive oil sands reserves, with particular emphasis on the area surrounding Fort MacKay. This region was chosen due to its significant petroleum potential. The CNN ML model employed in the study is based on the U-net architecture for prediction. This CNN model plays a key role in predicting potential oil sites based on geospatial images, minimizing reliance on human intervention for the computation and fine-tuning of attributes.

In order to facilitate predicting potential oil sites we used surface lineaments and aeromagnetic surveys datasets as indicators of areas with potential petroleum content.

The validation accuracy of the CNN model along with the selected petroleum predicting criteria showed promising results with a 95% overlap level (IoU) between the predicted zone and the existing zone. The aeromagnetic surveys parameter demonstrated a higher intersection value of 40.6% than the surface lineaments parameter which showcased a 25.7%, marking the former parameter as more accurate. The CNN U-net model developed in our approach demonstrated promising results, indicating the potential presence of petroleum sites in the study area. Still, certain model outputs exhibited slightly above-average inaccuracies, revealing areas where the model’s performance can be further improved.

While our approach aims to be a step forward in predicting potential oil sites, there are some improvements which can be made:

One potential avenue for enhancement lies in the geological aspect, where additional petroleum-indicating parameters, such as gravity and alteration maps, can be incorporated to increase prediction accuracy. In addition, petrophysical factors such as permeability and porosity can be incorporated to help detect oil reservoirs.

Experimenting with different model parameters can be beneficial in determining the optimal configuration for achieving the highest accuracy.

Overall, our approach serves as a step forward toward the effectiveness achieved by integrating remote sensing techniques, CNN ML models, and geospatial analysis tools in optimizing the prediction process for petroleum resources. The findings may help the potential for employing similar methodologies in different regions to efficiently and accurately identify and assess valuable petroleum sites. In addition, the proposed approach can be extended to applications beyond petroleum exploration, including mineral exploration and groundwater exploration.

Acknowledgments

The authors received no funding for the research, authorship, or publication of this article.

Funding information: Author state that no funding information.
Author contributions: FF and TS defined the conceptual framework of the approach. FF developed the model code and performed the model training and validation. FF, TS, and MBS elaborated the manuscript with contributions from all co-authors. FF and TS prepared the manuscript revision.
Conflict of interest: Authors state no conflict of interest.

References

[1] Petroleum Exploration - [Internet]. Oil-gasportal. https://www.oil-gasportal.com/upstream/petroleum-exploration/.Search in Google Scholar

[2] Burclaff N Research guides: Oil and gas industry: A research guide: introduction [Internet]. Library of Congress. 2021. https://guides.loc.gov/oil-and-gas-industry.Search in Google Scholar

[3] The Editors of Encyclopedia Britannica. Seismic survey | Description, methods, & facts. In Encyclopædia Britannica [Internet]; 2017.Search in Google Scholar

[4] Gravity and magnetic geophysical methods in oil exploration | Hart energy [Internet]. www.hartenergy.com. 2023. https://www.hartenergy.com/ep/exclusives/gravity-and-magnetic-geophysical-methods-oil-exploration-176138.Search in Google Scholar

[5] USGS. What is remote sensing and what is it used for? [Internet]. www.usgs.gov. 2022. https://www.usgs.gov/faqs/what-remote-sensing-and-what-it-used.Search in Google Scholar

[6] Kaiser MJ. A review of exploration, development, and production cost offshore newfoundland. Nat Resour Res. 2021 Jan;30(2):1253–90. 10.1007/s11053-020-09784-3.Search in Google Scholar

[7] IBM. AI vs Machine learning vs. Deep learning vs. Neural networks: What’s the difference? [Internet]. IBM Blog; 2023. https://www.ibm.com/blog/ai-vs-machine-learning-vs-deep-learning-vs-neural-networks/.Search in Google Scholar

[8] Misbahuddin M. Estimating petrophysical properties of shale rock using conventional neural networks CNN. In SPE Annu. Tech. Conf. Exhib., Denver, Colorado, USA, 5–7 October 2020; 2020.10.2118/204272-STUSearch in Google Scholar

[9] Ciresan DC, Meier U, Schmidhuber J. Multi-column deep neural networks for image classification. In Proceedings CVPR, Providence, RI, USA, 16–21 June 2012. p. 3642–9.10.1109/CVPR.2012.6248110Search in Google Scholar

[10] Hastie T, Tibshirani R, Friedman J. The elements of statistical learning: Data mining, inference, and prediction. New York, NY, USA: Springer; 2009.10.1007/978-0-387-84858-7Search in Google Scholar

[11] Merembayev T, Yunussov R, Yedilkhan A. Machine learning algorithms for classification geology data from well logging. 2018;11:206–12. 10.1109/ICECCO.2018.8634775.Search in Google Scholar

[12] Asante-Okyere S, Ziggah YY, Marfo SA. Improved total organic carbon convolutional neural network model based on mineralogy and geophysical well log data. Unconv Resour. 2021;1:1–8. 10.1016/j.uncres.2021.04.001.Search in Google Scholar

[13] Al-Mudhafar WJ. Integrating well log interpretations for lithofacies classification and permeability modeling through advanced machine learning algorithms. J Pet Explor Prod Technol. 2017;7:1023–33. 10.1007/s13202-017-0360-0.Search in Google Scholar

[14] Imamverdiyev Y, Sukhostat L. Lithological facies classification using deep convolutional neural network. J Pet Sci Eng. 2019;174:216–28. 10.1016/j.petrol.2018.11.023.Search in Google Scholar

[15] Mahmoodzadeh A, Mohammadi M, Daraei A, Farid Hama Ali H, Ismail Abdullah A, Kameran Al-Salihi N. Forecasting tunnel geology, construction time and costs using machine learning methods. Neural Comput Appl. 2020 May;33(1):321–48. 10.1007/s00521-020-05006-2.Search in Google Scholar

[16] Sircar A, Yadav K, Rayavarapu K, Bist N, Oza H. Application of machine learning and artificial intelligence in oil and gas industry. Pet Res. 2021 Jun;6(4):379–91. 10.1016/j.ptlrs.2021.05.009.Search in Google Scholar

[17] Koroteev D, Tekic Z. Artificial intelligence in oil and gas upstream: Trends, challenges, and scenarios for the future. Energy AI. 2021 Mar;3:100041. 10.1016/j.egyai.2020.100041.Search in Google Scholar

[18] Li H, Song X, Liu S. Artificial intelligence applications in petroleum exploration and production. Appl Sci. 2023 Jan;13(10):6214. 10.3390/app13106214.Search in Google Scholar

[19] Lawal A, Yang Y, He H, Baisa NL. Machine learning in oil and gas exploration: A review. IEEE Access. 2024;12:19035–58. 10.1109/ACCESS.2023.3349216.Search in Google Scholar

[20] Kim J, Yang H, Choe J. Robust optimization of the locations and types of multiple wells using CNN based proxy models. J Pet Sci Eng. 2020;193:107424. 10.1016/j.petrol.2020.107424.Search in Google Scholar

[21] Yang L, Sun SZ. Seismic horizon tracking using a deep convolutional neural network. J Pet Sci Eng. 2019;187:106709. 10.1016/j.petrol.2019.106709.Search in Google Scholar

[22] Wang H, Qiao L, Lu S, Chen F, Fang Z, He X. A novel shale gas production prediction model based on machine learning and its application in optimization of multistage fractured horizontal wells. Front Earth Sci. 2021;9:726537. 10.3389/feart.2021.726537.Search in Google Scholar

[23] De Kerf T, Gladines J, Sels S, Vanlanduit S. Oil spill detection using machine learning and infrared images. Remote Sens. 2020 Dec;12(24):4090. 10.3390/rs12244090.Search in Google Scholar

[24] Tian Q. Study on oil-gas reservoir detecting methods using hyperspectral remote sensing. ISPRS - Int Arch Photogramm Remote Sens Spat Inf Sci. 2012 Jul;XXXIX-B7:157–62. 10.5194/isprsarchives-xxxix-b7-157-2012.Search in Google Scholar

[25] Huang L, Dong X, Clee TE. A scalable deep learning platform for identifying geologic features from seismic attributes. Lead Edge. 2017 Mar;36(3):249–56. 10.1190/tle36030249.1.Search in Google Scholar

[26] Tysiąc P, Strelets T, Tuszyńska W. The application of satellite image analysis in oil spill detection. Appl Sci. 2022 Apr;12(8):4016. 10.3390/app12084016.Search in Google Scholar

[27] Lennon M, Babichenko S, Thomas N, Mariette V, Mercier G, Lisin A. Detection and mapping of oil slicks in the sea by comined use of hyperspectral imagery and laser induced fluorescence. EARSeL eProceedings. 2006;5(1):120–8.Search in Google Scholar

[28] Chenault DB, Vaden J, Mitchell DA, Demicco ED. Infrared polarimetric sensing of oil on water. Proceedings of SPIE; 2016 Oct 19. 10.1117/12.2241866.Search in Google Scholar

[29] U-Net: Convolutional networks for biomedical image segmentation [Internet]. Uni-freiburg.de. 2015. https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/.Search in Google Scholar

[30] Mandal M. CNN for deep learning | Convolutional neural networks (CNN) [Internet]. Analytics Vidhya; 2021. https://www.analyticsvidhya.com/blog/2021/05/convolutional-neural-networks-cnn/.Search in Google Scholar

[31] PhD HR. Benefits of using Google Colab (Python) [Internet]. Medium; 2020. https://medium.com/@sagihaider/benefits-of-using-google-colab-python-8f246c91bc52.Search in Google Scholar

[32] Mohammed A, Kathiresan P, Kumanan C, Ramasamy SM. Significance of surface lineaments for gas and oil exploration in part of Sabatayn Basin-Yemen. J Geogr Geol. 2010;2(1):119. 10.5539/jgg.v2n1p119.Search in Google Scholar

[33] Guo G, George SA. An analysis of surface and subsurface lineaments and fractures for oil and gas exploration in the mid-continent region [Internet]; 1999. www.osti.gov. https://www.osti.gov/biblio/5611.10.2172/5611Search in Google Scholar

[34] Gadirov V, Kalkan E, Ozdemir A, Palabiyik Y, Gadirov K. Use of gravity and magnetic methods in oil and gas exploration: Case studies from Azerbaijan. Int J Earth Sci Knowl Appl. 2022 Jul;4(2):143–56.Search in Google Scholar

[35] Pei Y, Liu B, Hua Q, Liu C, Ji Y. An aeromagnetic survey system based on an unmanned autonomous helicopter: Development, experiment, and analysis. Int J Remote Sens. 2017 Jan;38(8-10):3068–83. 10.1080/01431161.2016.1274448.Search in Google Scholar

[36] Geologic features of the Athabasca oil sands - Regional Aquatics Monitoring Program (RAMP) [Internet]. www.ramp-alberta.org. http://www.ramp-alberta.org/river/geography/geological+prehistory/mesozoic.aspx.Search in Google Scholar

[37] USGS. Landsat Missions | U.S. Geological Survey [Internet]. www.usgs.gov. https://www.usgs.gov/landsat-missions.Search in Google Scholar

[38] USGS. What are the band designations for the Landsat satellites? | U.S. Geological Survey [Internet]. www.usgs.gov. https://www.usgs.gov/faqs/what-are-band-designations-landsat-satellites.Search in Google Scholar

[39] Rezatofighi H, Tsoi N, Gwak J, Sadeghian A, Reid I, Savarese S. Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression. Thecvfcom [Internet]. 2019;658–66.10.1109/CVPR.2019.00075Search in Google Scholar

Received: 2024-03-12

Revised: 2024-08-20

Accepted: 2024-08-22

Published Online: 2025-12-10

This work is licensed under the Creative Commons Attribution 4.0 International License.

Articles in the same Issue

https://doi.org/10.1515/geo-2022-0714

Keywords for this article

oil site prediction; machine learning; convolutional neural network; image processing; sensitivity analysis

Creative Commons

BY 4.0