Data used in "Deep-Learning-Enhanced Electron Microscopy for Earth Material Characterization"

Understanding the distribution and characteristics of microstructures is crucial for the stability and effectiveness of subsurface activities. However, achieving statistical representativeness typically demands imaging a large number of samples at high magnification. To address this, we developed a workflow to expedite the imaging process. This method was developed, tested, and benchmarked using microstructural data from Back-Scattered Electron (BSE) maps of Berea sandstone, gabbro, serpentinite, and greenschist samples. BSE maps were captured at pixel resolutions of 50 nm (high resolution, HR) and 200 nm (low resolution, LR), with HR maps embedded within the LR maps. Additionally, for the greenschist sample, an extra map was scanned at 800 nm/pixel. Using an image registration algorithm, LR (64x64 pixels) and HR (256x256 pixels) image pairs were created for training Deep Learning (DL) models. This publication includes the following data: full-size and downscaled LR and HR BSE maps, paired LR and HR tiles for overlapping LR/HR data, model-generated data, and metrics of model performance. The image registration algorithm, DL methods, data processing steps, and dataset description are detailed in the research paper "Deep-Learning-Enhanced Electron Microscopy for Earth Material Characterization." For more detailed information about the structure and files in this data publication, refer to the included additional explanation file.