This study applies deep learning (DL) techniques to automate the classification of quartz grain microtextures under scanning electron microscopy (SEM), with the aim of supporting and standardising traditional microtextural analysis. Accurate microtextural analysis is crucial for sedimentary provenance and paleoenvironmental reconstructions, yet traditional methods are still prone to inter-operator variability. To overcome this, convolutional neural networks (CNNs) were trained on a diverse dataset of quartz grains from five depositional environments (dune, beach, alluvial, basal sands, nearshore) and two high-energy events (storm, tsunami). The models’ performance was evaluated in parallel with model SandAI to test robustness across varied sedimentary contexts. CNNs demonstrated high accuracy in identifying grains from mechanically dominated environments, which also represent the largest sample size in the dataset. However, classification declined where microtextures were overprinted or poorly preserved. While unsupervised methods grouped grains by surface texture, they lacked consistency with depositional settings. Nonetheless, CNNs offer a promising, objective approach for sedimentary provenance and extreme event grain analysis, given more balanced and robust datasets. • Pioneering application of AI to extreme-event sedimentology. • Deep learning models classify quartz grains images automatically. • Automated approach reduces subjectivity in microtextural analysis. • Overlapping microtextures limited predictive precision in complex grains. • Model errors reflected genuine microtextural superimposition patterns.
Marques et al. (Sun,) studied this question.