Description

Ground Penetrating Radar (GPR) has been an 
essential nondestructive geophysical tool in Saskatchewan's potash 
mines for over six decades. This innovative technology, used in 
conjunction with an active boring machine, facilitates real-time 
data collection, particularly for imaging the immediate clay seam 
(414-clay seam) above the mine roof. Its reliability is demonstrated 
by the accuracy with which the roof beam thickness (mine roof to 
414-clay seam) is interpreted in real-time, crucial for making 
informed safety decisions during mining operations. The 
imperative for a robust auto-picking algorithm tailored to handle 
complexities in potash mine GPR data is emphasized.
The previously developed algorithm called the Clustered Ratio 
Derivative (CRD) algorithm, employing unsupervised machine 
learning for real-time GPR interpretation showcased promising 
results. However, the CRD algorithm faces limitations due to 
potential sensitivity to variations in input data, especially in the 
presence of noise or anomalies. Geological variations, such as the 
presence of "stray clays" within the roof beam, pose challenges to 
the algorithm's performance and accuracy. 
In response to these challenges, this paper proposes a novel 
deep learning-based algorithm leveraging two distinct 
Convolutional Neural Network (CNN) architectures. These CNNs 
are designed to navigate the intricacies of the GPR data pattern 
specific to potash mines. The presented results indicate promising 
levels of accuracy, with the new method achieving 95.4% accuracy 
in detecting the 414-clay seam and an average of 88% accuracy in 
finding “stray clays” seen in the dataset when compared to a 
geophysicist interpretation. The overall analysis suggests that this 
new approach has the potential to detect mining room roof 
features with a high degree of accuracy.