Figure 3

Figure 3.

Improved modality data from multiple rotations of crossed polarizers for transmitted light imagery of thin sections. (A) Red-green-blue (RGB), cross-polarized (XPL) image of a granite thin section from the Golden Horn Batholith showing the full field of view (FOV) possible with our setup compared to those obtainable with a microscope camera. (B) False color image obtained using green (530 nm) light at three separate XPL orientations, 18° apart. (C) In principal component (PC) space, the pixel values for the four mineral classes (quartz, plagioclase, orthoclase, and mafics) in a single rotation RGB XPL image mostly overlap in one area of the plot. For an RGB XPL image containing five 18° rotations stacked into a 15-channel image, the pixel values spread out into a cone, where the position on the cone occupied by a given pixel relates to the class of the mineral and the relative orientation of its crystallographic axis. This added separation of the classes in the PC space of the five rotation XPL image improves the accuracy of pixel classifications from machine learning models, like the example given in (D). (E–G) In a zoomed-in portion of the image (E), we see that a support vector machine (SVM) using just a single rotation XPL RGB image (F) is 27% less accurate at classifying pixels compared to an SVM that is given the five-rotation image (G). Even with accurate classifications, analyzing only a relatively small FOV can add uncertainty. We see in (H) that the resulting modality data from the classification in (C) have highly variable values when assessed within the FOV of a traditional petrographic microscope. Each point in the plot represents the modality assessed in a randomly selected area of the segmentation equal to the size of a microscope FOV using either a 2.5× or 10× objective. The variation in these errors between classes stems from the characteristic size and relative abundance of the minerals. (I) To show the effect of crystal size and abundance, we calculate the number of images that correctly estimate the modality of a given mineral in a view size normalized to the mineral abundance (determined using a 4.5 × 5.5 × 4 cm 3D grinding, imaging, and reconstruction instrument [GIRI] reconstruction of the sample). In an experiment randomly drawing thin sections from the full volume of this granite sample, we see that an approximately equal fraction of images estimates the mafic mineral modality within a 90% correctness threshold when comparing GIRI to a 2.5× microscope objective. However, at the 95% threshold, as well as with the larger plagioclase crystals, the GIRI FOV performs nearly twice as well.

previous | next