MicroCrystalNet: An Efficient and Explainable Convolutional Neural Network for Microcrystal Classification Using Scanning Electron Microscope Petrography

Morphological characterization of microcrystalline rock textures typically relies upon the visual interpretation and manual measurement of scanning electron microscopy (SEM) imagery: a practice fraught with subjectivity, inefficiency, sampling bias, and data loss. We introduce a state-of-the-art computer vision pipeline, built on deep learning architectures, for segmenting and classifying individual microcrystals from SEM images. Initially applied to low-Mg calcite carbonate rocks, instance segmentation is achieved using a custom-tuned version of Meta’s Segment Anything Model (SAM). To train and test the classifier, we utilized 48 SEM images of diverse carbonate microtextures composed of Low-Mg calcite from studies performed worldwide. Each individual microcrystal (1852 in total) was labelled according to a bipartite classification scheme, encompassing both crystal shape (rhombic, polyhedral, amorphous, and spherical), and degree of crystal facet definition (euhedral to subhedral, anhedral), with a total of four distinct classes. MicroCrystalNet: our proposed classification model employs a convolutional neural network architecture, incorporating advanced feature map processing (feature normalization, dimensionality reduction, and sparse feature selection), integrated within a novel Normalized Sparse Reduction block. Performance metrics reveal excellent average precision scores (AP = 0.93-0.98) and Area Under Receiver-Operator Curve values (AUC = 0.95-0.99) across all classes, with visual comparison to manual ground truth images demonstrating powerful inter-class discriminatory power, even in the presence of occlusions.