Automated mapping of bedrock-fracture traces from UAV-acquired images using U-Net convolutional neural networks

This study presents the results of bedrock-fracture traces mapped using a U-Net convolutional neural network (CNN). Aerial photographs, acquired by unmanned aerial vehicles (UAV) with spatial resolution 0.55 cm, were used for training a U-Net CNN. The main objective was to train a network capable of mapping fracture traces on the Wiborg Rapakivi granite outcrops in the islands off the coast of the Loviisa Region in Southern Finland. The workflow involved optimization of the CNN parameters using root mean squared propagation optimizer and sigmoidal focal loss function for semantic segmentation of input images and pixel-wise identification of fracture traces. Quantitatively the results were assessed using various accuracy assessment metrics. Qualitative evaluations of the results were implemented by comparisons of orientations and length-frequency distributions of automatically- and manually-mapped traces. Results show that the model has the class-balanced accuracy score 0.945, predicting 88.79% of the fracture-trace pixels. Bedrock outcrops with well-exposed surfaces present high true positive rates (>99%). The demonstration (test) site has the class-balanced accuracy score of 0.873 and 75% true positive rate. Additionally, the fracture trace networks were able to replicate the orientation distributions of the manually digitized traces. The length distributions of automatic traces, however, differ with varying intensity from the manual trace length distributions. Nevertheless, this study demonstrates the successful application of deep neural networks and UAV-acquired images for fast and efficient automated mapping of bedrock-fracture traces during the initial phases of structural characterization of a region.