AI-driven analysis of SEM images of thin layers of volcanic rocks: a test with Segment Anything for Microscopy

In recent decades, image analysis has become established as a standard quantitative method. However, with a few exceptions, this analysis has predominantly relied on labour-intensive procedures that require extensive intervention by human experts. Consequently, physicochemical methods, which can be more readily automated for the analysis of large sample sets, have been more commonly employed to provide scientific evidence. This tendency is particularly pronounced in the geosciences, where the analysis of light microscopy and scanning electron microscopy (SEM) images of thin rock sections is highly informative but can only be applied to a limited number of samples, thereby restricting conclusions at larger scales. In recent years, advances in artificial intelligence have elevated image analysis to a new level by automating tasks traditionally performed by human experts, thus enabling the processing of much larger datasets. In this study, we examine the Segment Anything for Microscopy (micro-sam) package, which is based on the widely used deep learning tool Segment Anything Model (SAM), in order to assess its practical application for the analysis of SEM images of thin sections of volcanic rock samples. To this end, we first conducted a grid search to determine optimal SAM parameter values using a set of three SEM images, exploring the impact of different parameter combinations on automatic mask generation. We generated a dataset comprising more than 300 objects per image through interactive delineation and identification (“labeling”) and used this dataset to evaluate the results and identify the best parameter sets for each image, as well as common parameter configurations that performed well across all three images. A shared parameter set was subsequently used to compare the results obtained with the three available SAM models. The results indicate that two distinct parameter configurations are particularly noteworthy. The first configuration maximizes object detection and is intended for subsequent automatic instance segmentation using deep-learning methods