An Advanced Image Processing Technique for Backscatter-Electron Data by Scanning Electron Microscopy for Microscale Rock Exploration

0. INTRODUCTION

Backscatter electron analysis from scanning electron microscopy (BSE-SEM) produces high-resolution image data of both rock samples and thin-sections, showing detailed structural and geochemical (mineralogical) information (De Boever et al., 2015; Reed, 2005). This allows an in-depth exploration of the rock microstructures and the coupled chemical characteristics in the BSE-SEM image to be made using image processing techniques. Although image processing is a powerful tool for revealing the more subtle data "hidden" in a picture, it is not a commonly employed method in microstructural analysis of geological samples.

Image processing can be regarded as a form of data mining, delving into the image data to unearth hidden information (Cheng, 2021; Gonzalez and Woods, 2018). Unlike the traditional primary focus on the characteristics of individual mineral grains and rock microstructures in an image, image processing-assisted studies widen the scope to the entire image, thereby revealing the overall characteristics of the target via specific visualizations and/or mathematical functions (i.e., Gonzalez and Woods, 2018; El-Gabry et al., 2014). This allows microscale studies of samples to be enlarged to an exploration of their coupled "structural-chemical-spatial" characteristics.

Since the applications of powerful image processing methods are rare in geological studies, we provide, for the first time, below a general work-flow for using advanced image processing techniques to reveal "structural-chemical-spatial" coupling in BSE-SEM data (Figs. 1–2), with a ready-to-use MATLAB code.

Figure  1.  Working paradigm for advanced image processing on BSE-SEM images.

DownLoad: Full-Size Img PowerPoint

Figure  2.  Detailed work-flow for advanced image processing on BSE-SEM images (modified from Hou et al., 2023a).

DownLoad: Full-Size Img PowerPoint

1. BACKGROUND FOR IMAGE PROCESSING METHODS

Image processing is fundamentally based on the two-dimensional properties of digital images (Gonzalez and Woods, 2018): an image can be expressed by a function f(x, y), where the x and y represent spatial coordinates in the established x-y plane, and f is the specific pixel colour or the grey-scale value corresponding to each (x, y) point. When x, y and f are finite, the picture is a digital image. This framework allows dimensional operations and mathematical adjustments to be made on the images, laying the groundwork for image processing.

In the context of studying BSE-SEM images for geological purposes, the goal of image processing is to deliver a high-quality representation of the structures or minerals of interest. This involves directly emphasizing these features in images and/or revealing their quantitative characteristics (Zhang et al., 2021; Cnudde and Boone, 2013; Prêt et al., 2010). Notably, computer scientists have developed numerous ready-to-use mathematical functions and algorithms in recent decades to improve the theoretical foundations of image processing (i.e., Gonzalez and Woods, 2018; Boyat and Joshi, 2015; Sonka et al., 2013). Image processing can be easily performed on various platforms, such as, but not limited to, MATLAB (Image Processing Toolbox: https://uk.mathworks.com/products/image.html), Python (Scikit-image library: https://scikit-image.org/) and Fiji ImageJ (https://imagej.net/software/fiji/; Goldstein et al., 2018; Schindelin et al., 2012). This advancement makes image processing an efficient technique, allowing its diverse application without the necessity of a comprehensive mathematical background from the users.

Image processing is categorized into three levels, based on the study objectives of the images; "junior" (low-level), "middle" (mid-level) and "advanced" (high-level) image processing (Fig. 1; Gonzalez and Woods, 2018; Sonka et al., 2013). These levels can also be perceived as narrow image processing, image analysis, and image understanding (Gonzalez and Woods, 2018). Successful advanced image processing relies on the effective output of middle image processing, which, in turn, is fundamentally determined by the quality of the junior processing.

In geological studies, geologists need to design a coherent workflow and correctly call for the mathematical functions to generate the precise processing output desired (Zehner et al., 2015). However, image processing is still rarely discussed in the geoscience literature. Consequently, we describe here for the first time an efficient and self-developed workflow for advanced image processing suitable for geological and in particular petrological microscale studies. To aid the description, the techniques created by Hou et al. (2023a) in a detailed microstructural study of stylolites will be used as an example.

2. WORKFLOW FOR ADVANCED IMAGE PROCESSING ON GEOLOGICAL MICROSTRUCTURAL STUDY

Advanced image processing involves the interpretation of visual information, extending beyond the middle level by scrutinizing the properties and interrelationships among diverse image targets, employing computer vision to gain insights into the image content (image recognition) and elucidate the original objective scene. In the context of BSE-SEM image studies, the goal is to mathematically explore the coupled "structural-chemical-spatial" relationships of the research target across the segmented image. To achieve this aim, a suitable large area, focusing on the structures and minerals of interest research targetshould be chosen. Subsequently, segmentation, data matrices creation, and choosing the proper mathematical procedure are crucial (i.e., Fig. 2).

Step 1: Image acquisition and junior processing. No image is perfect in quality (Karras et al., 2020; Wang and Bovik, 2006; Hong et al., 1998). Acquiring a high-quality image serves as a pivotal precursor for facilitating smooth image processing. Depending on the research targets, "high-quality" here denotes an image with a suitably high resolution and size. The better the quality of the obtained images, the more efficiently the later processing can be conducted, since image processing is time-consuming.

The coupled "structural-chemical-spatial" of the study area (i.e., Fig. 2a) can be documented from an SEM mosaic (i.e., Fig. 2b), stitched from numerous overlapping high-resolution BSE images of the area of interest. In our case of studying stylolite microstructures (Hou et al., 2023a), 2 715 high-resolution BSE images were collected, covering an area of 17 440 × 3 100 µm² (Fig. 2b). Each BSE image was stitched in Fiji (Schindelin et al., 2012) using the "stitching plugin". To facilitate the later segmentation (see Step 2), the stitched images were performed denoising using the Median Smoothing Filter. Notably, image enhancement can be simply achieved by various Denoising Filter functions (i.e., Gaussian Filter, Mean Filter, Median Filter, Maximum and Minimum Filter; Swamy and Kulkarni, 2020; Gonzalez and Woods, 2018; Jain and Seung, 2008), and its selection relies on the research target and original quality of the image.

Step 2: Segmentation (middle processing Ⅰ). Segmentation revolves around feature recognition and extraction; striving to classify image regions and delineate descriptions of targets within the image. This process transforms images into numerical values or symbols, transitioning from input images to output data (Gonzalez and Woods, 2018; Xu et al., 2010). Depending on the aims of the study, the segmentation of SEM images can be efficiently assisted by the Weka Trainable Segmentation plugin (Arganda-Carreras et al., 2017). This plugin uses machine learning to classify images. Accordingly, computers can repeatedly learn the input characteristics across various regions of an image or an image stack until the generated segmented phases reach a high probability of a successful segmentation (typically at > 85% confidence; Minaee et al., 2022). The segmentation of Hou et al. (2023a) results to four different phases, including clay minerals, K-feldspars, pyrites and pores (Fig. 2c). Notably, the segmented phases can be assigned different false colors to improve the readability of figures.

Step 3: Image rescaling (middle processing Ⅱ). Advanced processing is most efficiently performed on the data matrices derived from the segmented phase image. To achieve this, middle image processing should be used for designing and creating the data matrices. The stitched SEM image mosaic (Figs. 2b–2c; Hou et al., 2023a) proved to be too large to be analyzed in its entirety. After testing, we proposed that the simplest, and likely quickest, approach is to subdivide the image mosaic into several rows and/or columns of a suitable smaller size. Significantly, the columns and rows chosen were used for designing and creating the data matrices for the later in-depth image analysis (in Step 4). Here, the stitched SEM image mosaic was cropped into 12 200 subregions (61 rows and 200 columns) all with the same dimension of 512 × 281 pixels (Fig. 2d). These images provided an excellent base for a detailed quantitative analysis of the research targets, which were grain size, porosity, and distribution of various mineral phases.

Step 4: In-depth Image analyses (advanced processing). In the context of BSE-SEM image studies, the goal here is to mathematically explore the coupled "structural-chemical-spatial" relationships of the research target across the segmented image. Establishing the proper mathematical procedure for analyzing the segmented pixels of each sub-image is crucial. These objectives include, but are not limited to, concentration density, pixel numbers, area ratios, and weight point distributions. In the case of Hou et al. (2023a), the percentage of pore space, area of pyrite, clay minerals and K-feldspars per unit area (size of each sub-image) were calculated and the resulting values used to populate an array (Fig. 2e), and subsequently establish data matrix (for each segmented phase; Fig. 2f), which was then plotted and contoured in MATLAB (Gonzalez and Woods, 2018), revealing significant hidden relationships (Fig. 2g). Establishment of the array, based on the target information extracted respectively from each 12 200 sub-images (Figs. 2d–2e). The processing performed on the image stack of the entire 12 200 sub-images (Fig. 2e), can significantly save computational time in a user-friendly image processing environment on home PCs. Note that in advanced image processing, there is no universal optimal mathematical approach to explore pixel characteristics. Instead, the most suitable method is constrained by the specific characteristics of the research target; essentially, a trial-and-error procedure.

In addition to the use of image processing in the study stylolite microstructures, we have also applied the technique, with similar workflows, in other microscale studies. For example, in the study of Hou et al. (2023b), prior to simulating the growth of mineral dendrites, advanced image processing was applied to examine the correlation between porosity and manganese-oxide concentration in mineral dendrites (Fig. 3). In the study, an extremely high-resolution BSE mosaic (Fig. 3a) was stitched from three individual images (resolution: 2 048 × 1 768 pixel, 0.06 µm/pixel). The mosaic was then cropped to 6 930 × 1 680 pixels, which were trained and segmented using Weka Trainable Segmentation (the machine-learning assisted approach), with a > 90% result confidence. Four components were classified, including pores (green in Fig. 3b), matrix minerals (white in Fig. 3b) and euhedral Mn oxides (red in Fig. 3b), and fine-grained Mn oxides (yellow in Fig. 3b). Relative changes in components were determined by variations of the area fractions of the pixels for each component on a series (195 in number) of overlapping (50%) columnar subareas (70 × 1 680 pixels) that traversed the segmented image. The segmentation and quantification along the SEM traverse not only show "structure-chemistry-spatial" information across dendrites (Figs. 3b, 3c), with EPMA exploration (Figs. 3d, 3e), it also confirms the replacement of porosity by manganese oxide (Fig. 3c). Moreover, employing a similar methodology, Zhang et al. (2021) studied the spatial distribution of minerals based on QEMSCAN mapping, which, together with a Fe isotope study, revealed the fluid-rock interaction processes associated with the formation of a porphyry Cu-Mo deposit. Based on a BSE-SEM and EPMA geochemical scan, image processing by Tschegg et al. (2020) visualized fault zone microstructures in a natural porous rock.

Figure  3.  Advanced image processing revealing the "structural-chemical-spatial" coupling. (a) High-resolution BSE mosaic showing dendrite microstructures from cores towards rock matrix. (b) Machine-learning assisted segmentation revealing the coupled variations in mineral concentration and microstructures across dendrites. Four phases were classified, including euhedral manganes oxides (red), fine-grained manganese oxides (yellow), matrix minerals (white) and pores (green). (c) Advanced-level image analysis showing variations between manganese oxide, matrix and rock porosity across Fig. 3b. (d) Variation of Mn concentration with porosity from matrix to dendrite core. Porosity data were calculated based on the result of image processing. Mn concentration was detected by EPMA. (e) Variation of Mn and Si from matrix to core. Both Mn and Si amount were detected by EPMA (Figs. 3d–3e were modified from Hou et al. 2023b).

DownLoad: Full-Size Img PowerPoint

3. PERSPECTIVE

Although we discuss the workflow design and application of advanced image processing in rock microscale exploration, notably, image processing is a flexible technique, and there is no 100% suitable method. The significance is to design a suitable working pathway for the specific image data set and mosaic. This short discussion is expected to inspire more image-based study in the rock microscale analysis. To encourage and facilitate the use of image processing in geological microscale study, the self-developed MATLAB codes can be access via: https://github.com/KunfengQiu/image-processing.

"One picture is worth more than ten thousand words" (Gonzalez and Woods, 2018). Image processing is not only an efficient approach to visualize rock details but also a powerful technique to deeply explore the microscale characteristics in the coupled "structure-chemistry-spatial" aspects of BSE-SEM image data. The basic research paradigm here, involving the transformation of visual sensory information into numerical data which is then subjected to in-depth data mining, holds broader implications and potential for studies of structure-material coupling across many scales.