Semi-universal geo-crack detection by machine learning

Introduction: Cracks are a key feature that determines the structural integrity of rocks, and their angular distribution can be used to determine local or regional stress patterns. The temporal growth cracks monitored in order predict impending failures materials structures such as weakened dam. Thus, spatial-temporal distributions should automatically for assessing integrity, associated patterns potential failure. Method: We show U-Net convolutional neural network, semantic segmentation transfer learning accurately detect drone photos sedimentary massifs. In this case, crack assess safest areas tunnel excavation. Compared coarse performance ridge detection, accuracy identifying images high 98% when evaluated against human identification, which is sufficient general properties rock faces engineering project. Result: Based on approximately 100 h manual labeling 127 20 network training, was able successfully 23,845 high-resolution photographs less than 22 using two Nvidia V100 GPUs. Meanwhile, more 80% observable volcanic outcrop Idaho without additional training. With modest amount extra we found significantly improved. surprising outcome research detector laboriously trained rocks also effectively applied faces. This important real-time assessment geological hazards lithology information dam inspection planetary exploration by autonomous vehicles. For another application, detected fractures faults with scale tens kilometers from Martian photographs. Conclusions: summary, our methodology CNN training suggests it semi-universal across range diverse settings.