Analyzing thin sections of rocks under the microscope and classifying them is a particularly important task in geological research. The primary tool used in petrography is the light microscope, which is used to observe and characterize rocks. This includes analyzing the geometry and structure of the grains, identifying fossil paleontology, and examining the structural characteristics of the minerals. The microscopic section of carbonate rock can reveal the characteristics of the type and content of the particles, which can provide important reference information for the subsequent exploration work. In traditional geological exploration studies, this task is manually identified by petrology experts and describes hundreds or even more thin sections of rock. This process consumes a lot of time and human resources, resulting in limited objectivity and efficiency of the research. Cheng et al. (2018) found that the increasing number of new rock sections has led to a gradual increase in the number of sections that need to be analyzed and archived by the geological community. These processes not only raise the economic cost of scientific research, but also deplete researchers' energy with repetitive and burdensome tasks, diverting their focus from more creative problem-solving. It is crucial to address the problem in order to improve efficiency and enable researchers to focus on more innovative issues.
In recent years, machine learning and deep learning have continued to be applied in various fields. Machine learning involves algorithms that enable machines to learn features from big data in order to intelligently identify new samples or make predictions about the future. Machine learning models trained with large amounts of data can learn more useful features. Such as face detection (Li et al., 2015; Garcia et al., 2017; Lan et al., 2024), medical image analysis and applications (Havaei et al., 2017), and healthcare (Gasparini et al., 2018), among others. Meanwhile, machine learning has made significant contributions to the geology community. Marmo et al. (2005) applied a multilayer perceptron algorithm to analyze over 1,000 thin slices of carbonate rocks from various marine environments. The model attained an accuracy of 93.3% on 268 test sets and 93.5% on 215 test sets.To address the common challenge of identifying basalt textures, Singh et al. (2010) utilized a multi-layer perceptron algorithm for a geological classification task and achieved a classification accuracy of 92.2% on 140 rock slices. Budennyy et al. (2017) aimed to evaluate the structural properties of rocks. They utilized a random forest classifier to distinguish between sandstone, limestone, and dolomite, achieving up to 95% prediction accuracy for these three rock types. However, these methods still have drawbacks. For instance, the classification is not detailed enough, the model is overly complicated, and it relies on too many complex mathematical methods. Some scholars have found that deep convolutional neural networks have the ability to automatically classify carbonate rock particles from rock thin section photographs (Lima et al., 2019; Koeshidayatullah et al., 2020; Idgunji et al., 2021; Koeshidayatullah et al., 2022). Liu et al. (2020) used convolutional neural network to accurately classify 22 types of carbonate rock particles, achieving over 90% accuracy in all four models. However, the study utilized a sample size of up to 13,000 images. Yu et al. (2021) trained a ResNet model to classify 10 types of carbonate rock bio-fossils, achieving a combined accuracy of 86%. Ho et al. (2023) employed the TaxonNet deep convolutional neural network architecture to classify six types of carbonate rock bio-particles with multiple labels, achieving an accuracy of over 90%. The above methods have improved the efficiency of geological research to varying degrees, and achieved relatively high precision. However, there are still some shortcomings, such as the relatively homogeneous identification of fossil and abiotic grain types by previous authors, and the efficient identification of only a small number of fossil and abiotic grain types. In carbonate rock microscopic thin sections, there is a wide range of fossil and abiotic grains, making it a challenging task to efficiently identify as many types of particles as possible. (2) The number of samples used by previous researchers is excessively large, leading to high training costs. Obtaining samples in the field of geology requires significant human, material, and financial resources. Accessing large numbers of rare fossils and abiotic particle samples is particularly challenging compared to the rich data available in other areas of deep learning. Therefore, it is worth exploring how to achieve higher accuracy rates when dealing with small sample sizes. Achieving more difficult tasks and classifying detailed images using easy and efficient methods has become a major challenge in today's world.
Deep convolutional neural networks are a type of deep learning model (Krizhevsky et al., 2012; Lecun et al., 2015) has higher accuracy compared with traditional neural networks (Garcia et al., 2017). An important feature of CNN is that it can be perceived locally and shared globally, enabling it to automatically extract essential features from original images for accurate classification. Additionally, It greatly simplifies the network parameters and speeds up the network training speed (Krizhevsky et al., 2012; Simonyan et al., 2015; Szegedy & Ioffe et al., 2015; Szegedy et al., 2016; Szegedy et al., 2017). In this study, we aim to utilize deep convolutional neural networks for the classification of fossil and abiotic grain types in carbonate rock flakes under small sample conditions. Our goal is to achieve more efficient and accurate classification. We classify carbonate rock particles into 11 types: V-shaped Crinoid, Irregular Crinoid, Round Crinoid, Ooid, Gastropod, Ostracod, Trilobita, Peloid, Coral, Foraminifer, and Algae. There are a total of 1266 images. Thin-section data were obtained from dense coring of the middle and upper parts of the Upper Ordovician Lianglitag Formation in well TZ72 in the Tazhong area of the Tarim Basin. The core process of this method, DCNN, has advantages such as local sensing and global sharing nature. We utilize this feature to delve into deep modeling to extract features of different types of particles in carbonate rock flakes, enabling us to capture diverse feature information. Subsequently, we introduce nonlinearities by applying activation functions (e.g., ReLU). This step enables the network to learn complex features. Finally, the full join layer is used to transform the mapping of the extracted features in order to classify carbonate rock flake particles. In computer vision tasks, including image classification and object detection, DCNN accuracy has surpassed human accuracy. Deep convolutional neural networks can outperform shallow neural networks when analyzing large datasets. In conclusion, DCNN can still guarantee high efficiency and accuracy in complex classification tasks.
The two main goals of this study are: (1) Expand the paleontological species identified under the microscope of carbonate rocks, including corals, foraminifers, gastropods and other rare paleontological species, so as to provide a data set of carbonate flake particle images for relevant researchers, in order to provide data support for the subsequent development of convolutional neural networks. (2) Achieve a higher accuracy rate when trained under small sample conditions, aiming to establish a framework and foundation for the application of convolutional neural networks in geology. The feasibility of this method has been proved by many experiments. Our training on raw data using the VGG-16 and ResNet-18 deep neural network architectures achieved highly competitive results, with accuracies of 78.9% and 83.5% at 100 epochs, respectively. Ho et al. (2023) also employed deep learning to classify carbonate rock particles, but they trained on the original dataset and achieved an accuracy of only 53%. In addition, we expanded the dataset through data augmentation and trained it using VGG-16 and ResNet-18 architectures. This resulted in achieving 98.8% and 100% accuracy at 100 epochs, respectively. Yu et al. (2021) employed the ResNet architecture for classifying carbonate microorganisms and achieved 98.8% and 100% accuracy with the use of data augmentation. The accuracy was only 86% when data enhancement was utilized. The results of this experiment are highly competitive with the latest carbonate fossil and abiotic grain classification methods.