Automatic channel detection using deep learning

Training Result

We trained our network on the synthetic data for 102 epochs in 2 hours using two GPUs. Mean value of Intersection over Union (Mean IU) is the accuracy metric defined as

$$(\frac{1}{n_{cl}})\sum\nolimits_{i}\frac{n_{ii}}{\sum\nolimits_{j}n_{ij}+\sum\nolimits_{j}n_{ji}-n_{ii}}$$ (4)

where $n_{cl}$ is number of classes, $n_{ij}$ is the number of pixels of class i predicted to belong to class j (Long et al., 2014). The cross-entropy cost decreases during training (Figure 8) and the mean IU is 87.4% after training. The global accuray defined as the percentage of pixels correctly classified in the image increases during training and reaches 99%. Applying the trained model to 285 unseen validation examples, we obtain the mean IU of 88.1%, which is close to the training mean IU. Comparing with true label of a vertical slice in Figure 9b, channel bodies are picked clearly in the synthetic dataset (Figure 9c).

cost2
Figure 8. Training losses.

The model uncertainty from Bayesian SegNet can be used to understand how confidently we can trust the channel segmentation. At boundaries of the channels, the prediction has high uncertainty (Figure 9d), which reflects the ambiguity of the network surrounding the definition of defining the transition between the channel and non-channel areas (Kendall et al., 2015). Comparing with true label of a horizontal slice in Figure 10b, the model can successfully pick the channel geobodies (Figure 10c). However, it is difficult to distinguish individual channels in the dataset, so there is high uncertainty where there are multiple channels (Figure 10d).

prdw9ver,prdw10ver,prdw7ver,vadw7ver
Figure 9. (a) Training vertial slice. (b) Ground truth of the training vertical slice. (c) Channel probability in the vertical slice. (d) Model uncertainty in the vertical slice.

prdw9hor,prdw10hor,prdw7hor,vadw7hor
Figure 10. (a) Training horizontal slice. (b) Ground truth of the training horizontal slice. (c) Channel probability in the horizontal slice. (d) Model uncertainty in the horizontal slice.

Automatic channel detection using deep learning