Grain size distribution (GSD) is useful in understanding the erosion, transport, and deposition of sediment, identifying the trends and patterns in response to surface processes, determining the slope stability, tracing the liquids-particles reactions and studying fluids through the porous sedimentary deposits (Syvitski, 1991; Li et al., 2017).
Although determining GSD of solidified sediments and rocks using image analysis methods is common, it is not so common in obtaining soils' GSD. Given the advantages of the mechanical method (e.g., sieving) to obtain a soil's GSD, it still has some drawbacks: e.g., the breadth of a particle passing through a sieve can also be greater than the seive size, and the determination of GSD by sieving also takes long time (Arasan et al., 2011). Therefore, mechanical method and image analysis may be complementary for a soil's GSD.
The era of geoscientific big data requires the larger-area data of grain sizes, and studying a statistically significant grain population can help the correction of studying 3D GSD using 2D sections. Both require the fast and massive extraction of grain-size data. U-Net is a fully convolutional network (FCN) developed for biomedical image segmentation. Compared to other convolutional neural networks, U-Net can work well with fewer training data to yield more precise segmentations (Ronneberger et al., 2015; Long et al., 2014).
Here, our grain extraction software based on U-Net algorithm can use 40~80 training data of grains to obtain hundreds and thousands of grain data in a few minutes (maybe even ten thousands but not tried due to the lack of data), with the loss of 6~10%, the precision of 85~100%, the recall of 81~97%, and the F1-score of 85~97%.