RockDNet: Deep Learning Approach for Lithology Classification

Abstract

Analyzing rock and underground layers is known as drill core lithology. The extracted core sample helps not only in exploring the core properties but also reveals the lithology of the entire surrounding area. Automating rock identification from drill cuttings is a key element for efficient reservoir characterization, replacing the current subjective and time-consuming manual process. The recent advancements in computer hardware and deep learning technology have enabled the automatic classification of various applications, and lithology is not an exception. This work aims to design an automated method for rock image classification using deep learning technologies. A novel CNN (Convolution Neural Network) is proposed for lithology classification in addition to thorough comparison with benchmark CNN models. The proposed CNN model has the advantageous of having very low complexity while maintaining high accuracy. Experimental results on rock mages taken from the “digitalrocksportal” database demonstrate the ability of the proposed method to classify three classes, carbonate, sandstone and shale rocks, with high accuracy, and comparisons with related work demonstrated the efficiency of the proposed model, with more than 98% saving in parameters.

1. Introduction

The detailed analysis of rock and the underground layers confronted during drilling of boreholes or wells is known as drill core lithology. The extracted core sample helps not only in exploring the core properties but also reflects the lithology of the whole surrounding area [1]. Previously, lithology classification was performed manually by visually inspecting the drill core samples to identify and characterize rock types according to color and texture. However, this method requires expert examiners with extensive experience and can be labor intensive and time consuming, in addition to being highly subjective due to the diversity of the rock structure [2].

Automatic classification of drill cores’ lithology is essential in order to provide a better understanding of subsurface rock formations [3,4]. The automation of drill core lithology classification is of high interest, as the demand to gather geological information is growing due to ongoing mining for various purposes including water and oil exploration. In this regard, automatic lithology classification mostly concentrates on well log data, which includes a wide range of measurements taken from boreholes, such as gamma-ray, resistivity and sonic log [5,6]. However, these traditional log data methods give limited rock information due to the omission of the heterogeneities below log resolution. In addition, well log data typically have limited vertical resolution and sampling intervals. Therefore fine-scale lithological variations may not be accurately detected, such as small fracture and rock details’ texture, which results in misinterpretation of lithological information. Such high details can be identified with X-ray and CT scans, as these imaging techniques allow for detailed examination of the internal structure and composition of core samples. CT images also provide 3D information about the internal structure of the cores and hence become a routine check in lithology detection [4].

Recently, the rapid development of computer hardware and artificial intelligence (AI) technology has paved the way for automatic classification of different applications, and lithology is not an exception [1,7]. Incorporation of artificial intelligence methodologies into the oil and gas sector has emerged as a recent development, presenting robust approaches for the efficient management of complex, multi-dimensional data systems. This integration has the potential to mitigate human bias and enhance the intelligence of operations throughout the entire oil and gas value chain. Despite the widespread use of deep learning and CNN models, the utilization of these methods for digital rock analysis, particularly in the context of lithology classification of micro-CT images, remains an ongoing area of development [8,9]. Hence, this study aims to demonstrate the practical implementation of deep learning techniques for the purpose of identifying lithological characteristics for rock samples through the analysis of such images.

The novelty inherent in our approach lies in its capacity to employ a novel CNN to recognize the complex connections between three-dimensional features extracted through convolution and the lithology classes derived by experts. Accordingly, we extend methods designed for two-dimensional imagery to construct a workflow that directly utilizes high-resolution three-dimensional CT images to be employed for lithology classification. This paper progresses in the following manner: the following section provides an overview of the work conducted in this regard. Section 3 presents the proposed method along with other performance evaluation metrics. In Section 4, results are listed and discussed in addition to a comparison with related work. Finally, Section 5 concludes this paper.

2. Related Work

Several works have utilized machine learning technology for lithology classification. A preliminary study to classify lithology based on core images was presented by Thomas et al. in 2011 [4], employing an object-based image analysis approach. Their study utilized grayscale core images from a single well to categorize three lithological types: carbonate–cement, shale, and sandstone. The nearest-neighbor classifier was employed for classification with an accuracy of 94%. However, their method is not applicable for large databases and not fully automatic, as it requires manual selection of images in the testing phase. Caja et al. [5] presented a method for lithology identification of thin images using image analysis and supervised machine learning. The training dataset included four classes: quartzites, siltstone, claystone, and carbonate. Nonetheless, the methodology employed a shallow model that lacked the capacity to be generalized. Furthermore, it depended on high-resolution thin section images and did not undergo evaluation using core tray images. Yang et al. [6] used the AdaBoost classifier to analyze volcanic lithology with an accuracy of 90% using logging parameters, such as rock fabric and pore structure. Likewise, Al-Mudhafar [7] applied the generalized boosted regression model to lithology data using well core permeability for the purpose of lithofacies classification. The findings indicated that the GBM algorithm demonstrated a superior performance compared to the conventional regression method. Moreover, in a number of studies, support vector machines (SVMs) are exploited in rock lithology [8,9]. As a case in point, Chawshin et al. [10] employed a SVM for lithofacies classification based on statistical features extracted from whole core CT images.

On the other hand, several works employed deep learning approaches for lithology classification. For example, De Lima utilized deep learning and transfer learning in a number of works [11,12,13] to classify images of rocks based on rock thin section and borehole image logs and CT images. Gonzalez et al. [14] introduced an automatic rock classification method using a process flow that merges core CT images, optical core photographs, conventional well logs, and routine core analysis (RCA) data. In the context of this process flow, the relevant rock-fabric features are employed to identify the rock classes by the utilization of a clustering algorithm after being extracted from all core CT images and core photographs. Firstly, several rock classes were assumed by input image. Subsequently, the optimization of the number of assumed classes involves a recursive increase in the number of classes and a minimization of a permeability-based cost function under a specific threshold. Finally, an artificial neural network is employed to predict the classes from well-log data using the obtained rock classes. Antariksa et al. [15] have utilized two classifiers, the gradient boosting and the random forest, for the purpose of classifying petrographic of log data from the Tarakan basin. The obtained results outperformed other machine learning methods. The authors of [16] discuss computing challenges in nature-oriented civil engineering, mentioning lithology classification and geophysical disaster prevention. It highlights the computational demands of Big Data and complex algorithms, suggesting AI, data filtration, and advanced technologies to reduce computational load.

Recently, the attention of research has been shifted towards deep neural networks (Deep learning) [17]. Deep learning modalities, specifically CNNs [18], have been revisited in the last ten years due to the growth of computing power, dominating the field of computer vision. Well known CNN architectures, such as AlexNet [19], VGG [20], GoogLeNet [21], and DenseNet [22], etc., have enabled researchers to enter into the realm of deep networks for problem solving within their specialized areas. Deep level features can be automatically extracted by the kernel of CNN compared to manual feature extraction [23].

In the field of microcomputed tomography (μCT) images of rocks, CNN yielded promising results. Alqahtani et al. [24] employed a CNN for the prediction of rock physical properties in μCT images involving specific surface area, average pore size, and porosity. This procedure yielded error rates below 7%. Similarly, in Karimpouli and Tahmasebi [25], a CNN based segmentation is conducted on sandstone μCT images; this procedure achieved better segmentation compared to the multi-thresholding approach and yielded a 96% accuracy in testing set of images. Wang et al. [26] introduced a method for enhancing the resolution of μCT images of rocks and removing noise by utilizing a super resolution CNN. Their suggested CNN achieved a 70% reduction in relative error when compared to conventional approaches according to their results. In the work of Ran et al. [27], a CNN model under the name of Rock Type deep CNNs (RTCNNs) is proposed; this model is developed for lithology classification using patches of images. The utilized six types of rocks, i.e., granite, conglomerate, mylonite, limestone, shale, and sandstone, were classified, and an accuracy of 97% was achieved. Baraboshkin et al. [28] classified 20,000 drill core images collected from Russia into six lithologies: massive, siltstone, laminated sandstone, granite, shale, and limestone. A number of well-known CNN-based architectures were utilized for lithology classification based on the optical core images. Notably, GoogleNet achieved accuracy of 72% on new core images, while attaining 70% on ResNet [29]. In [30], image patches were unitized and classified using a pre-trained CNN model ResNetXt-50. Later, the obtained results were compared against the results of [29] utilizing ResNet-18, and Inception-v3 [31] models. Comparison results showed that their model outperformed the previous results, achieving an accuracy of 93.12% with new core images. Additionally, Anjos et al. [32] utilized a deep learning technique in favor of lithological patterns identification in carbonate rocks based on microtomographic images. In Zhang et al. [33], three lithologies were classified, i.e., conglomerate, sandstone, and shale, using various CNNs on a 1500 gray-scale images dataset, with a size of 64×64. The obtained accuracy is reported to be 95%. Valentın et al. [27] employed the micro-resistivity borehole and ultrasonic logs for inputs, and a deep residual CNN is also employed for feature extraction. Later, the determination of the lithology of each sample is conducted. For blind sample testing, an average accuracy is 81.45%.

In the aforementioned publications, the features are either from pore scale micro-CT images, or they are extracted from two-dimensional cross-section images or from two-dimensional image slices taken from three-dimensional stacks of micro-CT or whole core CT images. Zhang et al. [34] introduced a method named ConvXGB that combines both the XGBoost and CNN to provide accurate classification results. The CNN is utilized as a trainable feature extractor, while XGBoost is employed for labels prediction. Lin et al. [35] introduced a two-layer XGBoost model using ZY1-02D hyperspectral images for lithology classification. The reported results showed good accuracy and adaptability to limited datasets compared to traditional methods through field validation. The XGBoost model is also reported to be effective in lithology prediction in both studies [34,35]. Other works [36,37] employed the extraction of thin-section image features of rocks for the purpose of lithology classification and good results were also reported.

3. Methodology

In contrast to conventional neural networks, CNNs can preserve spatial information in images and are resilient to noise. Therefore, CNNs can model non-linear characteristics, which can be found in lithological samples data, and hence can be generalized well to unseen examples [11]. The next subsection illustrates the pipeline of the proposed method.

3.1. Image Augmentation

Data augmentation has been applied in order to add more images and overcome the problem of imbalance classes [12]. In this context, data augmentation can be accomplished by translation or rotation (affine transforms) or image distortion. In this work, image augmentation is applied during the training to enable the model of learning from a large number of images. Specifically, we applied rotation with 45 and −45 degrees in addition to image flipping across both axes. In this way, general features of images do not change, while pixels are translated and the model generalized to more images.

3.2. Transfer Learning

In transfer learning, a pre-trained model is reused to boost the performance on a new problem. We used transfer learning in this work to save training time and improve generalization [13], as the original number of samples is small. In this regard, the most popular pre-trained networks are employed for the sake of classification: Alexnet, VGG16, VGG19, Restnet19, Darknet19 and Darknet53. The number of neurons in the last classification layer in these models is tailored to be equal to the number of output lithology classes.

3.3. Proposed CNN Model

In this work, a novel CNN model is proposed for lithology classification. The architecture of the proposed CNN model is depicted in Figure 1.

Figure 1. Proposed CNN model architecture.

This model has the advantageous of having low complexity and light weight while being robust. The proposed model has nine layers with only 69.6 K of learnable parameters which is a huge model agility compared to the next best model (shown later in Section 4), which has a number of parameters equals to 23.6 M. This saving represents around 98.32% saving in terms of learnable parameters and hence the proposed model has a light model size with fast training time.

This specific architecture was chosen to achieve a trade-off between complexity and accuracy. The performance of the model was then evaluated using 5-fold cross validation and the score is computed as the average among the obtained score across validation scores. The training hyperparameters for all CNN models are unified as illustrated in Table 1.

Table 1. Training hyperparameters.

Training Options
Optimizer	SGDM
Initial learn rate	0.0001
Learn rate drop factor	0.01
Max epochs	40
Activation function	ReLU
Shuffle	Every epoch
Validation frequency	60
Execution environment	Single GPU
Mini batch size	64
Computer specifications used in training and classification: CPU CORE I7-12700H/RAM 16GB/GPU RTX3060-6GB

3.4. Batch Size

The number of training samples per one iteration is known as the batch size. This size plays a crucial role in training generalization and timing performance [32]. Large batch size results in a smooth convergence and less noise. It also results in faster training time per epoch. However, large batch size requires more memory. In this work, different batch sizes are tested and a size of 64 is found to be the best in terms of both accuracy and training time.

4. Experimental Results

The automation of drill core lithology classification is the main goal of this work. The experiments were performed on a lithology dataset taken from digitalrocksportal repository [38]. In this work, the data preprocessing, models training and evaluation experiments are performed under a MATLAB R2022b environment on a PC with a corei7 I7-12700H CPU with 16 GB of RAM, an SSD hard drive, and RTX3060-6GB in terms of GPU.

4.1. Dataset

Geological drill cores serve as a direct representation of geological formations. However, there is a limited number of open-source datasets related to borehole cores. This limitation can be attributed to the substantial cost and time associated with drilling through strata. Additionally, the task of drilling is typically undertaken by specific authorities, making it challenging for external parties to access this data. In this work, a drill core CT scans images dataset is utilized, which contains a total of three well-known lithology categories: carbonate, sandstone and shale. The images are taken from the Digital Rocks portal [38]. Table 2 illustrates the details of the employed CNN while Figure 2 depicts samples from each class in this dataset.

Table 2. Network specifications.

Model Name	Number of Layers	Number of Connections	Depth	Size on Disk (MB)	Parameters	Classification Time (Second)	Input Size
Proposed CNN	15	18	5	0.204	69,602	1.8	128 × 128 × 3

Figure 2. Samples from each class.

The images included three classes: carbonate, sandstone and shale. The original size of the dataset images is 1000 × 1000, which is modestly large; thus, preprocessing is required to reduce the size of these images, which will eventually decrease the size of the whole dataset to around 224 × 224 depending on the employed CNN model.

4.2. Evaluation Metrics

To comprehensively assess the performance of the models, three performance measures are utilized in this research: precision, recall, and F1 Score. These are fundamental metrics in statistics and machine learning and are delineated by Equations (1)–(3), respectively [39,40].

$$Precision=\frac{TP}{(TP+FP)}$$

(1)

$$Recall=\frac{TP}{(TP+FN)}$$

(2)

$$F1Score=2\ast \frac{Precision\ast Recall}{(Precision+Recall)}$$

(3)

where TP is the True Positive, where the positive class is correctly predicted by the model. TN is the True Negative, where the negative class is correctly predicted by the model. Moreover, FP is the False Positive, where the positive class is incorrectly predicted by the model. In contrast, FN is the False Negative, where the negative class is incorrectly predicted by the model.

4.3. CNN Models Training and Evaluation

The remaining 20% is utilized for testing to evaluate the CNN models after training is conducted. Table 3 illustrates the network structure. On the other hand, Table 4 presents the details of the employed database. Table 5 illustrated precision, recall, and F1 score for each class of all employed CNN models in addition to the metrics of the proposed method.

Table 3. Network structure.

No.	Layer Name	Output Shape
1	Input Image	128 × 128 × 3
2	2-D Convolution	128 × 128 × 10
3	Batch Normalization	128 × 128 × 10
4	2-D Max Pooling	64 × 64 × 10
5	2-D Convolution	64 × 64 × 20
6	Batch Normalization	64 × 64 × 20
7	2-D Max Pooling	32 × 32× 20
8	2-D Convolution	32 × 32 × 64
9	Batch Normalization	32 × 32 × 64
10	2-D Max Pooling	16 × 16 × 64
11	2-D Convolution	16 × 16 × 30
12	Batch Normalization	16 × 16 × 30
13	Fully Connected	1 × 1 × 3
14	SoftMax	1 × 1 × 3
15	Classification Output	1 × 1 × 3
Maximum number of learnable parameters		69,602

Table 4. Dataset details.

Class\Precious Stones Type	Number of Samples	Training (80%)	Testing (20%)
Carbonate	239	191	48
Sandstone	196	156	40
Shale	211	169	42
Total	646	516	130

Table 5. Precision, recall, and F1-score for each class across all employed CNN models.

Classification Model Name	Class	Precision % for Testing Data	Recall % for Testing Data	F1-Score % for Testing Data
Darknet-53	Carbonate	97.8	100	98.9
	Sandstone	100	100	100
	Shale	100	96.6	98.2
Darknet-19	Carbonate	95.7	100	97.8
	Sandstone	100	100	100
	Shale	100	93.1	96.4
Resnet-50	Carbonate	90	100	94.7
	Sandstone	100	100	100
	Shale	100	82.8	90.6
Proposed 15 layers CNN	Carbonate	91.8	100	95.7
	Sandstone	100	100	100
	Shale	100	86.2	92.6
VGG 19	Carbonate	95.7	100	97.8
	Sandstone	100	100	100
	Shale	100	93.1	96.4
VGG 16	Carbonate	93.8	100	96.8
	Sandstone	100	100	100
	Shale	100	89.7	94.5
Alex-Net	Carbonate	91.8	100	95.7
	Sandstone	100	100	100
	Shale	100	86.2	92.6

Upon the examination of Table 5, DarkNet-53 exhibited the highest classification accuracy of 99.22% amongst all CNN models, followed by DarkNet-19 and VGG19 with an accuracy of 98.4%. It can also be seen that the proposed CNN models maintained a classification accuracy above 96.9%. Despite the higher accuracy achieved by DarkNet53, the developed CNN model achieved a comparable accuracy of 96.9% with a huge saving in both parameters and model size of 69.6 K and 204 KB, respectively. This saving in terms of size on disk is around 98.32% compared to that of darknet-53 model with a loss in accuracy of merely 2.32%. This small loss is negligible compared to the benefit saving in both time and memory, where small model sizes offer advantages in terms of applicability and efficiency for deployment in embedded systems. On the other hand, VGG16 has a huge size of 476 MB with an increased accuracy of merely 1.6%

The compact architecture reduces the overall memory footprint, making it well-suited for devices with limited storage capacity and computational resources. The reduced model size facilitates quicker deployment and lower latency in real-time applications in addition to energy efficiency. Moreover, the reduced model size not only facilitates integration into portable devices but also contributes to faster classification, allowing for rapid lithological analysis at the point of data collection, where real-time identification of rock types can significantly impact decision-making processes. In such a system, a lithology classifier can be inserted in the drill core for the purpose of real-time lithology classification inside the hole, which saves time and cost compared to lithology core extraction.

Figure 3 depicts the confusion matrices across all CNN models. The best result was recorded with DarkNet-53 with an accuracy of 99.2% followed by a similar performance of both DarkNet-19 and VGG19. A slightly lower accuracy of 96.4% was recorded with Alexnet and RestNet-50. The proposed method achieved a similar performance to the aforementioned networks with the benefit of being lightweight.

Figure 3. Confusion matrices of all CNN models.

Table 6 gives the average precision, recall and F1-socre across the employed CNN models in this work. On the other hand, Table 7 presents a complexity comparison across the employed models in terms of number of layers, depth, timing and size of each model. Comparison with other works in the literature is presented in Table 8. It can be clearly seen that the proposed method holds an advantage in comparison to similar approaches. It has a very small desk size with many fewer parameters, while maintaining a competitive accuracy. On the one hand, the proposed method outperforms the work in [2] and [32] in terms of accuracy. On the other hand, although the accuracy is slightly less compared with the work of [1], this is negligible considering the huge saving in terms of size on disk and number of parameters.

Table 6. Average Precision, Average recall, and Average F1-score across all CNN models.

Employed Model	Average Precision %	Average Recall %	Average F1-Score %	Average Accuracy %
Darknet-53	99.3	98.9	99.1	99.22
Darknet-19	98.6	97.7	98.1	98.4
Resnet-50	96.7	94.3	95.5	96.1
Proposed CNN	97.3	95.4	96.3	96.9
VGG 19	98.6	97.7	98.1	98.4
VGG 16	97.9	96.6	97.2	97.7
Alex-Net	97.3	95.4	96.3	96.9

Table 7. Time and complexity comparison of different CNN model against the proposed model.

Classification Model Name	Number of Layers	Depth	Number of Learnable Parameters	Number of Connections	Size on Disk	Classification Time (Second) for Testing Set	Accuracy %
Darknet-53	185	53	41 M	207	148 MB	4.66	99.22
Darknet-19	65	19	19.5 M	64	74.1 MB	2.94	98.4
Resnet-50	177	50	23.6 M	192	83.7 MB	3.68	96.1
Vgg19	47	19	143 M	46	495 MB	6.35	98.4
Vgg16	41	16	138 M	40	476 MB	5.8	97.7
AlexNet	25	8	60.02 M	24	201 MB	3.78	96.9
Proposed CNN	15	5	69.6 K	18	204 KB	1.8	96.9

Table 8. Comparison with related work.

Model Name	Number of Classes	Accuracy %	No. of Parameters	Size on Disk
Proposed CNN	3	96.9%	69.6 K	0.204 MB
ResNeSt-50 [1]	10	99.60%	≈25 M	97.5 MB
ResNeXt-50 [2]	3	93.12%	N/A	N/A
Anjos et al. [32]	3	81.33%	177,395 k	N/A
FCN-8s [41]	8	96%	≈140 M	≈400–500 MB
FC-Res [42]	6	97.60%	≈10 M	≈30 MB

Upon examining Table 8, our schemes outperform other works in terms of number of parameters and size on disk. Additionally, our work achieved a comparable classification accuracy to the work in [1,41,42], with exceptionally low memory footprint and parameters making it applicable in real-time systems. The closest model in [3] achieved a low accuracy of 81.33% with almost 2.5 the number of parameters used in our work.

5. Conclusions

The detailed analysis of rock and underground layers is known as drill core lithology, where the extracted core provides essential information about the lithology of the surrounding area. Manual rock identification from drill cuttings is a subjective and time-consuming process. The recent advancements in computer hardware and deep learning technology have enabled the automatic classification of various applications, and lithology is not an exception. In this work, an automated method for rock image classification using deep learning is proposed. A novel CNN network was proposed for lithology classification in addition to employing transfer learning with various benchmark CNN models for the sake of lithology classification. Experimental results on rock mages taken from the “digitalrocksportal” database demonstrate the ability of the proposed method to classify three classes, i.e., carbonate, sandstone and shale rocks, with high accuracy and low memory footprint. The proposed model achieved an accuracy of 96.9% with only 69.6 K learning parameter and 204 KB model size. Comparisons with related work demonstrated the efficiency of the proposed model.