Traditional description methods for describing rock traces heavily rely on the operator’s knowledge and expertise. For the characterization of hard rock pillar traces in complex underground mines, this subjective evaluation strategy not only limits the generalization of the description method, but also causes local trace omissions, increasing the potential for quantization error. To that end, 19,780 trace images with a resolution of \(1000 \times 1000\) pixels were collected to train and test the DeepLabV3 + model. This paper introduces a novel description method termed the intersection grid description (IGD) to describe and quantify traces predicted by the DeepLabV3 + model. In addition, a novel performance index, namely precision efficiency covalent (PEC), is established to evaluate the trace description accuracy. The results indicate that the proposed method can effectively captures the similarity and integrity of hard rock pillar traces. Furthermore, trace features (length, angle, density, and intensity) quantified by the proposed method closely match actual survey results. Overall, the proposed IGD method with high performance in characterization of traces on hard rock pillar surfaces suggests its potential for further application in trace investigation of other rocks in the complex environments.
