Differentiating outcropping strata from oblique photogrammetric data using an octree-based convolutional neural network with spatial parameters

Abstract

Differentiating strata is a prerequisite for subsequent stratum outcropping research. Because of the large scale of outcropping strata and the complex surrounding terrain environments, differentiating the layers in traditional artificial field geological surveys is time-consuming and laborious. The emergence of oblique photogrammetry technology provides a new way to overcome the inherent challenges in traditional methods. This paper proposes a new method for differentiating outcropping strata from oblique photogrammetric data using an octree-based convolutional neural network (O–CNN) with spatial parameters. This method enables O–CNN to obtain high-level semantic information of point clouds and additional prior knowledge by fine-tuning network parameters and mining attributes and spatial parameters as network input, thereby improving the performance of the network for outcropping stratum differentiation. This study used the outcrop Xuankongsi in Fugu, Shaanxi, China as an example for stratum differentiation experiments. Compared with the O–CNN models using only attributes or spatial parameters as network input, the O–CNN model integrating both attributes and spatial parameters obtained better results, reaching an overall accuracy of 87.5%. After the stratum differentiation results were further optimized, the accuracy reached 97.2%, and the differentiation results were more consistent with the manual results of stratum differentiation. The experimental results show that the method proposed is effective and provides an intelligent way to differentiate outcropping strata.