Automated Segmentation of Olivine Phenocrysts in a Volcanic Rock Thin Section Using a Fully Convolutional Neural Network

Abstract

An example of automated characterization and interpretation of the textural and compositional characteristics of solids phases in thin sections using machine learning (ML) is presented. In our study, we focus on the characterization of olivine in volcanic rocks, which is a phase that is often chemically zoned with variable Mg/(Mg + Fe) ratios, so-called magnesian number or mg#. As the olivine crystals represent only less than 10 vol% of the volcanic rock, a pre-processing step is necessary to automatically detect the phases of interest in the images on a pixel level, which is achieved using Deep Learning. A major contribution of the presented approach is to use backscattered electron (BSE) images to: 1) automatically segment all olivine crystals present in the thin section; 2) determine quantitatively their mg#; and 3) identify different populations depending on zoning type (e.g., normal vs reversal zoning) and textural characteristics (e.g., microlites vs phenocrysts). The segmentation of the olivine crystals is implemented with a pretrained fully convolutional neural network model with DeepLabV3 architecture. The model is trained to identify olivine crystals in backscattered electron images using automatically generated training data. The training data are generated automatically from images which can easily be created from X-Ray element maps. Once the olivines are identified in the BSE images, the relationship between BSE intensity value and mg# is determined using a simple regression based on a set of microprobe measurements. This learned functional relationship can then be applied to all olivine pixels of the thin section. If the highest possible map resolution (1 micron per 1 pixel) is selected for the data acquisition, the full processing time of an entire thin section of (Formula presented.) containing more than 1,500 phenocrysts and 20.000 microliths required 140 h of data acquisition (BSE + X-Ray element maps), 8 h of training and 16 h of segmentation and classification. Our further tests demonstrated that the 140 h of data acquisition can be reduced at least by a factor of 4 since only a part of the thin section area (25% or even less) needs to be used for training. The characterization of each additional thin section would only require the BSE data acquisition time (less than 48 h for a whole thin section), without an additional training step. The paper describes the training and processing in detail, shows analytical results and outlines the potential of this Deep Learning approach for petrological applications, resulting in the automatic characterization and interpretation of mineral textures and compositions with an unprecedented high resolution. Copyright © 2022 Leichter, Almeev, Wittich, Beckmann, Rottensteiner, Holtz and Sester.