Digital rock physics has emerged as a powerful complement to laboratory core analysis, supplying pore-scale insight that underpins reservoir evaluation, geothermal production and subsurface energy storage (Blunt et al., 2013). By converting high-resolution images into binary maps of solid and void, researchers can simulate single- and multiphase flow, geomechanics and geochemistry with unprecedented fidelity. These capabilities are of growing importance to low-carbon technologies such as carbon capture and storage (CCS) and the geological disposal of radioactive waste, where long-term safety hinges on predicting how CO₂ or brine migrates through complex pore networks (Blunt et al., 2013). Geological CO₂ storage is highly sensitive to micro- and meso-scale heterogeneity because capillary- and gravity-dominated flow regimes amplify the influence of pore connectivity on plume migration (Ringrose et al., 2022).

A persistent bottleneck is segmentation: most workflows still rely on thresholding greyscale scanning-electron-microscope (SEM) or X-ray micro-computed-tomography images, followed by manual post-processing of the masks when pores are filament-thin or sparsely distributed (Sarkar et al., 2018). In these cases, edges lie near the noise level and small porosity channels or bridges of matrix that separate pores must be decided by hand, which increases annotation time and introduces operator bias. Openly available pore-scale datasets with pixel-accurate masks are scarce, limiting reproducible benchmarking (Da Wang et al., 2021). Recent CCS case studies further show that subtle layering and fracture networks can redirect or immobilise dense-phase CO₂ over decadal timescales (Ringrose et al., 2022). To address this bottleneck, we introduce high-quality pixel-accurate labels and benchmark strong classical segmentation baselines for carbonate SEM images.

Recent studies apply convolutional neural networks to porous media and report strong results for delineating pore space in µCT/SEM images (Alqahtani et al., 2022; Bihani et al., 2022; Bangaru et al., 2022). These models can integrate broader context and better preserve thin structures, but they require pixel-accurate labels, topology-aware training, and non-trivial compute, and many published datasets are not class-balanced or publicly reusable. Artificial intelligence model training also takes exponentially more computational resources, which necessitates quicker and more efficient methods.

Classical computer vision segmentation techniques often uses e SEM/µCT (micro Computed Tomography), typically combining thresholding (Otsu, 1979; Sauvola and Pietikäinen, 2000), distance-transform/watersheds (Beucher and Meyer, 1992; Soille, 2003), edge detectors (Canny, 1986) and morphological operations (Soille, 2003). These pipelines are fast and explainable, but performance depends on a number of issues, such as illumination and noise, overlapping pore/matrix histograms, and the requirement for hand-tuned parameters. In addition, thresholding can be unstable under local shading (Sezgin and Sankur, 2004), watersheds can over-segment at weak boundaries unless seeds are carefully controlled (Beucher and Meyer, 1992), edge-based methods often miss pore interiors when boundaries are diffuse (Canny, 1986), and morphology can erode 1–2 px filaments unless parameters are tailored per image (Soille, 2003). This motivates releasing pixel-accurate labels that preserve thin isolated pores and a transparent benchmark across carbonate textures to document both strengths and failure modes of classical approaches. Investigating the optimal uses of these methods can help to inform better architectures in machine learning methods, especially important for very large images with high resolution. In three dimensions this requirement is even more important.

Connected porosity governs flow and thus permeability/relative-permeability upscaling, whereas isolated porosity mainly contributes to capillary trapping and (via wetted surface) adsorption, where the separation has direct consequences for CCS modelling. In practice the distinction is challenging because connectivity is scale- and physics-dependent (sub-pixel throats may allow diffusion but not viscous flow at the prevailing pressure gradient), 2D slices can miss 3D bridges, and thin, low-contrast links near grain contacts make edges ambiguous while intensity/texture overlap undermines simple thresholds. Our pixel-wise labels therefore preserve 1–2 px filaments and topology, and the benchmark explicitly reports the conditions under which classical pipelines succeed or fail.

This study aims to (i) publish an open, fully-labelled carbonate SEM data set, (ii) benchmark eight classical segmentation pipelines, and (iii) introduce a Hybrid voting ensemble. To address the identified gaps we present a high-quality, fully labelled 2-dimensional (2D) SEM image set of outcrop carbonates. The original mosaic (29 056 × 22 952 px, 8-bit depth) was subdivided into 155 tiles of 2048 × 2048 px. Each carbonate grain was manually annotated and traced using hand-drawn polygon boundaries, with pixel classification defined as follows:

Black pixels inside a polygon represent isolated, disconnected, intragranular pores.

Recently there has been a large amount of investment and research in evaluating AI and generative AI in various domains. However, it is important to understand the baseline results of standard and extremely efficient models, such as older computer vision algorithms, to see where improvements can be made. Thus, we evaluate eight representative classical pipelines covering the main families used in computer vision algorithms: thresholding (global/adaptive), distance-transform/watershed, edge-based, morphological and texture filters. Together, they expose complementary strengths and failure modes directly relevant to CCS and radioactive-waste performance metrics, balancing the recovery of thin, isolated intragranular pores (relevant to capillary trapping and radionuclide immobilisation) with precisely mapping the connected network that governs potential migration pathways. Each method is particularly suited for specific use cases.

We have evaluated these algorithms noting their typical behaviour and performance in order to ascertain, which algorithm is best suited to which application. In future, further machine learning approaches can be built upon this more interpretable and explainable understanding.

The eight segmentation pipelines evaluated are:

Watershed (Beucher and Meyer, 1992) – a seed-and-grow algorithm on a distance map that excels at splitting touching pores when using markers to control over-segmentation.

Hybrid voting is an ensemble scheme that labels a pixel as “isolated” only when at least two of the three highest-performing filters (Distance Transform, Watershed and Local Contrast) concur, thereby combining their complementary strengths while suppressing individual false positives as discussed in (Isensee et al., 2021) and (Soille, 2003).

The contributions of this study are fourfold:

Open data: release of the carbonate SEM tiles, polygons and masks, plus full acquisition metadata, under a CC-BY licence.

The analysis was carried out on ultra-high resolution SEM images from a number of samples. The sample images are from an alluvial sand sheet or sabkha packstone, with grain-size codes ranging from upper-fine (UF) to medium (M) sand, from an industrial partner's field (exact location withheld per confidentiality agreement), but from a depth in the range 4346–4788 m. Its helium porosity at ambient stress lies between 0.016 and 0.02, and Klinkenberg-corrected nitrogen-gas permeability at ambient pressure (k∞), lies between 0.032 and 0.087 mD. The Klinkenberg correction was carried out by linear extrapolation of apparent permeability (kapp) versus 1/Pm (with Pm the mean gas pressure). k∞ is reported as metadata only (not used in analyses; see Table 1).

Four back-scatter SEM mosaics were exported as 29 056 × 22 952 px, 8-bit greyscale, LZW-compressed TIFFs. Each mosaic has a calibrated pixel size of 0.195 µ m px⁻¹, yielding a field of view of ∼ 5.7 mm × 4.5 mm. All four mosaics were captured in the same session, ensuring consistent greyscale calibration. After partitioning the mosaics into 2048 × 2048 px tiles, we randomly retained 100 tiles, with 25 from each mosaic (≈ 4.2 × 108 px) retained for analysis.

Each raw tile was first binarised with a global greyscale threshold to separate pore space from solid matrix (Fig. 1a). Carbonate grains were then outlined manually in Microsoft Paint using the free-hand polygon tool (Fig. 1b). Black pixels inside a polygon represent isolated (intragranular) pores, black pixels outside all polygons represent the connected fracture–pore network, and grey matrix pixels are ignored. The three-region map is collapsed to a three-class mask (0 for isolated, 1 for connected and 2 for mineral/background) and saved as an 8-bit PNG registered one-to-one with the raw tile. Annotation time per ring was ∼ 1–2 min, where some images with more rings took longer to annotate then those with fewer. A simple morphological clean-up filled stray matrix pixels inside wide fractures and removed speckle noise smaller than five pixels. This strategy allows us to highlight easily all the larger regions containing isolated pores, using an algorithm to process the pixels inside the annotated yellow rings. This benefitted us by not having to annotate each pixel for either class. All annotations were first rechecked after labelling all images, and again once using the software to create the isolated and connected classes to further validate the classes, where it was manually corrected if needed.

This study focuses on unsupervised classical computer-vision pipelines to establish transparent, reproducible baselines using the released pixel-accurate labels. We used the yellow bounding rings to determine whether pores were isolated or connected. We evaluated eight classical methods, ordered from most to least conservative as summarised in Table 2 and illustrated in Fig. 2. Table 3 provides a summary of pixel class distributions across the dataset the results of which will be described in detail in the next section.

Pixel-wise performance was assessed with the standard metrics defined below; the numerical values are reported in Table 4 and representative qualitative results are shown in Fig. 3.

Accuracy: Out of all pixels, the percentage the method labelled correctly.

Watershed delivers the best overall balance, missing only 26 isolated-pore pixels (high recall) while maintaining a moderate false-positive rate. The trade-off is that one in every two “isolated” pixels it proposes is actually connected (precision ∼ 0.54). The Refined Morphology method is uncompromising by demanding strict roundness, size of connected pixels (5–100 px²) and ≥ 80 % grey surrounding pixels, it achieves near-perfect precision (0.97) but sacrifices recall (0.15). This makes it ideal for quality-control checkpoints where only the most certain isolated pores matter.

Regarding computational details, the methods were tested on a M3 Macbook Pro with 16 GB of memory. As the methods are based on parameters and not through machine learning, there is no time taken to train any model, but it is needed for calibration. Inference time with these methods were typically between 356 to 1458 ms per 2048 × 2048 px image. Parallel computation by splitting each image into 9 tiles of a 3 by 3 grid improved performance to 122 to 714 ms per image. However, as these are static methods, they are not generalisable to other datasets without further calibration.

We report results for eight classical segmentation pipelines and a simple Hybrid voting ensemble on the labelled carbonate SEM dataset. Our comparative analysis demonstrates a clear trade-off between recall and precision, a common challenge in image segmentation. The Watershed algorithm emerges as the most balanced method, achieving the highest mean IoU (0.73) and an exceptional recall for isolated pores (0.97). This indicates that it successfully identifies the vast majority of true isolated pores. Its primary drawback is a modest precision (0.54), meaning it misclassifies a significant number of connected-pore pixels as isolated. For applications in CCS and radioactive waste storage, this behaviour could lead to an overestimation of the rock's capacity for capillary trapping or radionuclide immobilisation. Nonetheless, for a comprehensive first-pass characterisation of a sample, Watershed provides the most complete picture of the potential isolated pore system.

In stark contrast, the Refined Morphology method is defined by its conservatism. It achieves near-perfect precision (0.97) at the cost of very low recall (0.15). This pipeline is engineered to only identify pixels that are unequivocally part of an isolated, intragranular pore. While it misses the majority of such pores, its classifications are highly reliable. This makes it an ideal tool for specific quality-control tasks or for generating a high-confidence “minimum case” estimate of isolated porosity, which is critical for safety case assessments where false positives are unacceptable. The performance of the other methods, such as Distance Transform and Edge-based segmentation, falls between these two extremes, but none offer a more advantageous balance of metrics than Watershed. The poor performance of methods like Morphological Gradient and Global Thresholding highlights the difficulty of relying on simple intensity or gradient features in these complex carbonate images, where pore boundaries can be subtle and indistinct. Global Threshold and Gabor Texture perform poorly for isolated pores (Precision = Recall = 0.00) because they produce no isolated predictions on our carbonate SEM tiles. For Global Threshold, local normalisation suppresses small dark intragranular pores; for Gabor, the filter bank responds to edges rather than interiors, and our conservative post-processing removes the resulting small islands. These two methods are therefore not recommended when isolated-pore quantification is required. By contrast, the Watershed approach is preferable when completeness is the priority (high Recall; good at splitting touching grains), while the Distance Transform is useful when thin filaments must be recovered (with light pre-filtering to limit noise). The Edge-based method is best where boundaries are sharp and continuous, and the Refined Morphology pipeline should be chosen when false positives must be minimised (near-perfect precision at the expense of Recall). These benchmarks provide a transparent reference and document failure modes that can inform future learning-based baselines.

This initial study effectively validates our polygon-based annotation workflow and establishes a robust performance baseline using traditional, explainable methods. It underscores that even without deep learning, significant microstructural insights can be extracted. More importantly, it quantifies the specific challenges, such as distinguishing finely connected fracture networks from truly isolated pores, that a future deep-learning model must overcome.

The injection of CO₂ into underground reservoirs for carbon capture and underground storage (CCUS) is very different from the well-known process of extracting gas (Ringrose et al., 2022). There are many reasons for this, of which three of the most important are (i) that supercritical CO₂ injection is highly sensitive to heterogeneity and anisotropy (Reynolds and Krevor, 2015; Trevisan et al., 2017), (ii) the CO₂ flow regime varies from the well, to the near-wellbore space, then more distally (Rui et al., 2025), and (iii) variations in rock microstructure at small scales control the overall fluid flow (Al-Zainaldin et al., 2017). All three of these issues are controlled by the amount of void space in the rock (porosity) and the size and connectedness of the void space; in other words, the microstructure of the rock. Therefore, general pore-segmentation without taking into account the different kinds of pores is important for all aforementioned applications.

The pixel-wise masks produced here separate connected from isolated pores. From these masks one can derive simple inputs for modelling: (i) the connected-pore fraction (share of pixels labelled connected) to constrain permeability/relative-permeability upscaling by excluding non-flowing porosity; (ii) the isolated-pore fraction to bound capillary-trapped volumes and residual saturations; and (iii) where surface-area information is available (e.g., from 3-D µCT or stereological estimates), the connected wetted surface area to parameterise adsorption-capacity models (e.g., Langmuir-type).

Heterogeneity in the lithology and particularly in microstructure and its influence on transport processes has a strong effect on the capillary- and gravity-dominated flow processes which are typical of CO₂ storage projects (Reynolds and Krevor, 2015; Trevisan et al., 2017). A good example of this is the effect of the complex layering of the Sleipner field, which dominated the shape and development of the CO₂ plume, which represents the effects of strong vertical heterogeneity (Furre et al., 2017, 2020). Flow-in gas and oil production systems is much less affected by rock heterogeneity because fluid transport occurs in the viscous-dominated flow regime (Bentley and Ringrose, 2021).

Consequently, CCUS processes can only be managed if the CCUS geological modelling and simulation takes account of the rock microstructure, and that in turn requires us to be able to segment any image or measurement of a rock into its different phases as that small scale.

This paper has attempted to use computer vision algorithms to automate and accelerate the ability to segment the rock into various mineral phases, into connected and unconnected pore space. Future work would involve using artificial intelligence and machine learning approaches, which have previously been used in CCUS. One approach used machine learning techniques to help upscale core data so that it could be used in a conventional large-scale reservoir model and simulation campaign (Yu et al., 2022). Unfortunately, such an approach has the outcome of averaging the small-scale heterogeneities which control the supercritical CO₂ flow.

A combination of CT-scan images and various machine learning approaches, including convolutional neural networks (CNNs), was proposed by (Panaitescu et al., 2024). They recognised the importance of understanding how multi-scale, image-derived data can enhance understanding of porous features can be unlocked with the use of machine learning, but did not specifically focus their work on CCUS reservoirs, targeting instead marginal oil and gas reservoirs. Building on their earlier work (Panaitescu et al., 2024), they do not attempt to segment connected and unconnected porosity, as in our case, but porosity from fractures. They noted that simple histogram thresholding is not sufficient in accurately perform semantic even if threshold optimisation techniques (Otsu, 1975) or Gaussian Mixture Models (Huang and Chau, 2008) are used, because simple differences in grey scales does not provide enough information to separate phase.

Consequently, they used CNNs, especially using U-net-inspired architectures (Ronneberger et al., 2015), providing results that are extremely useful for application to CCUS reservoirs in coal (Karimpouli et al., 2020) and other sequestration reservoirs (Kim et al., 2020; Reinhardt et al., 2022; Pham et al., 2023).

Elsewhere in CCUS science and engineering, deep learning is also making an impact. These applications include simulation (Chu et al., 2022), process engineering (Sabeena, 2023) and optimising carbon storage (Wang et al., 2025).

The findings and methodologies presented in this paper provide a strong foundation for microstructural analysis, but they are subject to several limitations that also define the direction of our future research.

This study addressed a critical gap in digital rock physics: the quantitative segmentation of isolated versus connected porosity in carbonate reservoir analogues. Our primary contributions are:

An Efficient Annotation Workflow: We detailed a polygon-based method that preserves pore-edge fidelity more effectively than pixel-wise tracing, reducing annotation time while maintaining high quality.