Real exploration activity data were offered by the Laboratory of Subsurface Energy, Water, and Environmental Systems (SEWES) at Virginia Tech. Two-hundreds raw CT scan images of rocks are used as training data, and both global thresholding methods and manual segmentation are applied to generate ground truth in order to test models.

The format of inputs and labels are not the same, so we make a normalization for both original and segmented images. We used min-max normalization to map the images to a range of 0-1.

We tested multiple loss-functions including Dice-coefficient loss and binary cross-entropy, and we decided to use binary cross-entropy for both models as it provides the best result.

Both U-net and simple CNN models were tested and used to process CT images.

[Zhu et al, 2017]

Overview of U-net model

Kernels :5

Epoch : 100

Learning Rate : 1e-05

[SuperDataScience ]

Overview of simple CNN model

Conv2d:6

Maxpool:3

Randsample:3

Epoch: 100

Learning Rate: 1e-05