Dados da Submissão

Título

IMAGE SEGMENTATION FOR AUTOMATIC LITHOLOGY CLASSIFICATION: A REVIEW OF DEEP LEARNING MODELS

Texto do resumo

The lithology classification involves separating geological images into distinct regions that correspond to different rock types. Automating this process reduces the time and effort required for manual analysis. This automation also improves the robustness and reliability of the classification results. Image segmentation is a crucial step in automatic lithology classification using computer vision techniques. This process extracts significant features from images, such as textures, colors, and patterns, which are distinctive for different lithologies. Therefore, high-quality segmentation enhances the accuracy and precision of the lithology classification. Recent advances in Convolutional Neural Networks (CNNs) have brought Deep Learning (DL) to the forefront of image segmentation. This study aims to review the recent literature on the application of DL to image segmentation for lithology classification, covering state-of-the-art models and generalization performance metrics. This review also summarizes the application challenges of DL-based lithology image segmentation and classification methods. The literature analysis method of this review consists of four main steps: (a) search for relevant research; (b) assess the studies’ quality; (c) extract relevant data; and (d) synthesize the extracted data. Two researchers (i.e., a computer scientist and a geologist) performed literature searches using online scientific databases and both assessed the discovered studies’ quality and followed a pre-defined data extraction strategy. In the online search, 85 relevant papers were initially found by this review, and 36 papers were finally selected after the quality review. From the annual publication analysis, we can note that lithology image segmentation using DL has become a research topic of interest in the past three years. CNNs stand out as the most used (>50%) among deep learning-based models for image segmentation in the field of lithological classification. Architectures like GoogLeNet, ResNetSt-50, and ResNeXt-50 have achieved high generalization performances in the drill core image classification. Stacked Sparse Autoencoder (SSAE) can also achieve high overall accuracy, especially in remote sensing data (e.g., synthetic aperture radar imagery). Other architectures, such as Fully Convolutional Networks (FCN) and Vision Transformers (ViT), are also used and achieve satisfactory results. The mapped studies evaluate the performance of DL models focusing on metrics that quantify the model's generalization. The confusion matrix comparing real and classified lithologies is the basis of these metrics. The most commonly adopted metrics were classification accuracy, precision, recall, and F1-score. Despite significant advances in lithology image segmentation and classification, several application challenges are highlighted in the literature (e.g., segmentation accuracy, data heterogeneity, and extensive labeled datasets - more than 15,000 core images in some studies). Furthermore, many studies address simulated environments. However, they did not address complex properties of real applications, such as dust and lack of lighting. These properties reduce the image quality and must be considered to understand the real performance of the models. DL provides greater precision and automation in geological image segmentation for lithology classification. CNNs, SSAE, FCNs, and ViT stand out as promising methods, each with its specific advantages.

Palavras Chave

lithology classification; image segmentation; deep learning; literature review