Accurate segmentation of digital rock images is essential for characterizing pore-matrix systems and predicting petrophysical properties. However, the diversity of rock textures across different lithologies poses a significant challenge for conventional segmentation networks, especially under limited training data. To address this, we introduce DRI-SAM (Digital Rock Image - Segment Anything Model), a hybrid segmentation framework that leverages the powerful visual prior of the Segment Anything Model (SAM) and adapts it to the digital rock domain. Specifically, we apply LoRA-based fine-tuning to SAM's image encoder to better capture rock-specific microstructures, while U-Net is employed to generate prompt points, guiding SAM toward accurate pore-matrix delineation. This approach retains the encoder's representational power while allowing domain-specific adaptation via LoRA, enabling effective cross-domain generalization under limited training data. The model is trained exclusively on 200 annotated images of Bentheimer sandstone, covering two distinct voxel resolutions, and is evaluated on digital rock images of varying lithologies, resolutions and imaging modalities. The results confirm that DRI-SAM achieves accurate segmentation on both sandstone and more challenging carbonate samples, including synthetic and SEM images, without additional retraining or parameter adjustments. Compared to DeepLabV3 + and the only LoRA-tuned SAM, DRI-SAM demonstrates superior performance under limited supervision, highlighting its strong generalization and practical value in digital rock image analysis. Moreover, the findings suggest that foundation models like SAM, when properly adapted, also hold great promise for broader geoscientific imaging tasks.

This study systematically evaluates the performance of 15 conventional global single-threshold segmentation algorithms and three representative convolutional neural network (CNN) models across multi-modal geotechnical material images, including computed tomography (CT) and scanning electron microscopy (SEM) data. Based on their characteristics, thresholding methods are classified into three categories: histogram-based, entropy-based, and other approaches. Four types of geotechnical material CT images and two types of SEM images were selected as the evaluation datasets, and an objective assessment criterion that does not require manual annotation was proposed. The results indicate that the performance of different thresholding algorithms varies considerably across imaging modalities: the Otsu method performs best on coal and sandstone CT images, the Liu-S method (implemented in the JHNY-DPM software) excels on sandy soil CT images, and the Yen method demonstrates strong robustness on SEM images. However, all thresholding methods fail to effectively segment granite images with uneven grayscale distributions. In contrast, deep learning models exhibit superior performance across all modalities, with U-Net achieving the highest accuracy and stability during both training and validation without noticeable overfitting, significantly outperforming Fcn and Deeplabv3. Further experiments on a combined CT-SEM dataset reveal that despite domain adaptation challenges, U-Net can consistently segment complex geotechnical structures across different imaging modalities. Overall, the analysis demonstrates that deep learning models substantially enhance the accuracy and robustness of multi-modal geotechnical image segmentation, providing guidance for algorithm selection and supporting the unified processing of multi-source imaging data toward automation and intelligent analysis in digital geotechnical research.

While size distribution has traditionally been the dominant metric in rock fragmentation, studies have shown that both size and shape characteristics are influential in determining energy consumption, equipment wear, flowability, and efficiency. The main objective of this research is to provide a scalable, quantitative, and light-independent method for characterizing the 3D shape of rock fragments. This work leverages an existing deep learning approach that combines LiDAR-based point cloud acquisition with deep learning instance segmentation. Trained on a synthetic 3D labeled dataset, the deep learning model, SoftRock, accurately segments individual rock pieces and provides key shape metrics such as sphericity, angularity, aspect ratio, and longest dimension for each rock in the pile. The model's performance was then validated on three distinct rock piles curated to represent different spectrum of rock types and sizes, including blasted limestone from a quarry, rounded pebbles, and crushed copper ore from a conveyor belt. The model demonstrated a high level of accuracy across these diverse samples, with the error for key shape metrics ranging from 2% to 16%. While some inaccuracies were observed, primarily due to the sensitivity of sphericity and angularity to noise in the point cloud data, our findings validate the model's ability to capture key shape characteristics. This study provides a foundational framework for integrating comprehensive 3D particle morphology into mining workflows, offering more accurate data to inform decisions that enhance operational efficiency and equipment longevity.

This study presents an AI-driven framework for predicting groundwater storage (GWS) in the arid to semi-arid regions of Agdz and Zagora in southern Morocco, where sustainable water resource management is increasingly critical. Four machine-learning models Random Forest (RF), CatBoost, AdaBoost, and Multi-Layer Perceptron (MLP) were trained using a comprehensive dataset integrating Gravity Recovery and Climate Experiment (GRACE) mission-derived Terrestrial Water Storage (TWS), remote sensing indicators such as The Normalized Difference Vegetation Index (NDVI) and Land Surface Temperature (LST), and key climatic variables. To improve predictive accuracy, model hyperparameters were optimized using the Swan Optimization Algorithm (SOA), a bio-inspired metaheuristic technique. Among the tested models, RF achieved the highest performance, with root mean square error (RMSE) values of 4.70 mm and 4.29 mm and NSE scores of 0.998 and 0.999 for Agdz and Zagora, respectively. TWS consistently emerged as the most influential predictor across all models. These results highlight the potential of integrating artificial intelligence, satellite remote sensing, and bio-inspired optimization for periodically updated monitoring and prediction of groundwater storage in data-scarce regions. The proposed framework provides a valuable decision-support tool for smart irrigation planning and climate-resilient water management in agriculture-dependent areas.