DEEP LEARNING NEURAL NETWORK IN GEOLOGICAL INTERPRETATION OF GROUND PENETRATING RADAR IMAGES

Abstract

The post-processing of GPR subsurface survey data is both time-consuming and involves the highly subjective task of geological interpretation. In this study, a Deeplab v3+ convolutional neural network is used to automize the process in an attempt to improve efficiency and to reduce the impact of subjectiveness. For this, a semantic segmentation approach is used for sedimentology classification of radar facies in GPR images. Manual interpretation of 143 images with size 150x360 pixels, from a beach area with mostly sand and clay sediments, resulted in the identification of seven different radar facies. In three experiments, the neural network was evaluated with subsets of images with these facies. From a subset of 20 images with two radar facies, the neural network was able to correctly identify facies with a precision of >90%. From another subset of 44 images with two radar facies and a separate classification of remaining pixels, the facies were correctly identified with a precision in the range of 43%-49%. In a subset of 43 images with six radar facies, the network could identify five of the facies with a precision in the range of 21% - 54%, and one of the facies could not be identified. The lower precision for the dataset with six facies was probably due to a lower representation in the dataset. The study shows that a convolutional neural network is a feasible method for automizing GPR image interpretation for sedimentology applications. It also demonstrates the importance of using a substantial and representative dataset for learning the neural network.