TLDR: The paper introduces LITHOS, the largest and most diverse public dataset and experimental framework for automated petrography. It includes over 211,000 high-resolution images and 105,000 expert-annotated mineral grains across 25 categories, captured under dual-polarization conditions. The authors also propose a dual-encoder transformer model, the LITHOS Baseline, which significantly outperforms single-polarization methods in classifying minerals, demonstrating the value of combining different light modalities for geological analysis. The project aims to make automated petrography more accessible and reproducible.
Petrography, a vital branch of geology, involves analyzing the mineral makeup of rocks using microscopic thin section samples. This process is crucial for understanding rock properties across various fields, including geology, archaeology, engineering, mineral exploration, and the oil industry. However, traditional petrography is a labor-intensive task that relies heavily on experts conducting detailed visual examinations through optical polarization microscopes, which limits its scalability and highlights a strong need for automated techniques.
To address this challenge, researchers have introduced the Large-scale Imaging and Thin section Optical-polarization Set, or LITHOS. This innovative framework represents the largest and most diverse publicly available experimental setup for automated petrography to date. LITHOS comprises an impressive 211,604 high-resolution RGB image patches captured under polarized light and features 105,802 expert-annotated mineral grains, categorized across 25 distinct mineral classes. Each annotation is rich with detail, including the mineral class, spatial coordinates, and expert-defined major and minor axes represented as intersecting vector paths, which effectively capture the grain’s geometry and orientation.
The LITHOS project not only provides this extensive dataset but also evaluates multiple deep learning techniques for mineral classification. The researchers propose a dual-encoder transformer architecture that integrates both polarization modalities—plane-polarized light (PPL) and cross-polarized light (XPL)—as a robust baseline for future research. This method consistently outperforms models that rely on a single polarization, clearly demonstrating the significant advantage of combining polarization information for accurate mineral classification. The LITHOS Benchmark, including the dataset, code, and pretrained models, has been made publicly available to encourage reproducibility and further advancements in automated petrographic analysis. You can find the research paper here: Towards Automated Petrography.
The Importance of Petrography and Its Challenges
Petrography describes and analyzes rocks and soils, focusing on their mineralogical composition, texture, and structure. It helps identify minerals, characterize grain size, shape, and porosity, which are critical attributes in fields like economic geology for resource exploration. In archaeology, it helps determine the origin of artifacts, while in engineering, it assesses the quality of construction materials. For the oil industry, petrography is key to evaluating a rock’s capacity to store and transmit hydrocarbons.
Traditionally, petrographers use a point-counting method, superimposing a grid on a thin section and identifying minerals at each point. This process is time-consuming and requires specialized expertise, with each thin section taking around 12 hours to prepare and another 10 hours for annotation. The lack of large, well-annotated datasets has been a major hurdle for automating this process.
Introducing LITHOS: A New Era for Automated Petrography
LITHOS stands out as the most comprehensive publicly available resource for automated petrography. It surpasses existing frameworks by two orders of magnitude in image count and one order of magnitude in annotated mineral grains. The dataset includes 25 mineral categories, with 105,802 annotated grains and 211,604 high-resolution image patches. Each thin section was digitized using an automated petrographic microscope, capturing images under both PPL and XPL conditions. These paired images are crucial as they highlight variations in color, texture, and birefringence patterns, which are essential for mineral identification.
The LITHOS Benchmark and Baseline Model
The LITHOS Benchmark is designed to evaluate mineral classification performance. It includes two main tasks: a binary classification task (Quartz vs. Non-Quartz) and a multi-class task (classifying into one of 25 mineral categories). The dataset naturally has an imbalanced distribution, with quartz and its subtypes making up nearly half of the annotations, reflecting real-world geological samples.
The LITHOS Baseline model is a transformer-based architecture specifically designed to leverage the dual-polarization images. It uses two pre-trained Vision Transformer (ViT) encoders, one for PPL and one for XPL, which are then combined through a dual-decoder architecture. This design allows the model to capture complementary mineralogical features from both polarization modes, leading to more accurate identification.
Also Read:
- DINO-MX: A Flexible Framework for Self-Supervised Learning in Medical Imaging
- Advancing Medical Image Analysis with Adaptable Foundation Models
Key Findings and Future Directions
The LITHOS Baseline consistently outperformed single-modality models across both binary and multi-class classification tasks. For instance, in the binary task, it achieved an F1 score of 0.861, surpassing single-modality Swin Transformer models. In the multi-class task, it achieved an accuracy of 0.623 and an F1-score of 0.432, again outperforming single-modality approaches. This highlights the effectiveness of integrating both polarization modalities, particularly through the cross-attention mechanism in the dual-decoder fusion module.
Despite these advancements, challenges remain, particularly with the long-tail distribution of mineral classes. While common minerals like Monocrystalline and Rock Fragment achieve high F1-scores, rare classes like Opaque and Heavy mineral show lower performance. Future work could explore enhanced supervision strategies and techniques to mitigate class imbalance.
The public release of LITHOS aims to democratize access to high-quality petrographic data, promote reproducibility in mineral classification, and stimulate interdisciplinary collaboration between geoscientists and machine learning researchers. This work offers a positive impact by enabling progress in automated petrography and geological analysis without presenting ethical risks or negative societal impacts.
