A convolutional neural network for semi-automated lineament detection and vectorisation of remote sensing data using probabilistic clustering: A method and a challenge

Abstract

In this paper we present the framework and open-source software code to train and apply Deep Learning Convolutional Neural Networks (CNNs) to the prediction of geological lineaments using topographic, magnetic, and gravity raster data. Many important applications relate to the recognition of linear geological structures from remote sensing data, such as thrust faults, bedrock fault and shear zones, lithological contacts, fractures and fold structures. The digitization of fault lineaments is conventionally performed by geologists or geophysicists with working knowledge of the relevant data, e.g., topographic Digital Elevation Model, magnetic data, and gravity data. Visual inspection and extraction is simple but subjective; the process is also time-expensive with efficiency and accuracy depending on the individual's knowledge, experience, and skill. For decades there has been interest in ways to automate this process. Our CNN approach is trained using publicly available lineament GIS data from the Quest BC project in British Columbia, Canada, and the Loch Lilly-Kars area of New South Wales, Australia. The datasets used to train the prediction models resulted in interesting predictions proximal to the training areas: some major lineaments are indicated, some are missed, and potential new (valid) lineaments are indicated. The results indicate potential for use as a semi-automated lineament detection solution. In contrast, but as anticipated, application of the model to the blind-test area of the Swayze greenstone belt, Ontario, produced poor lineament prediction results (as compared to publicly available interpretations). We interpret this result as related to insufficiently large training data inputs. However, it is inferred that results could be improved through feature engineering (e.g., use of topographic slope, rather than simply elevation) without the need to simply create larger training datasets. We hope that, by making the code open-source, the geoscience community will use this platform to gradually improve an open source fault prediction model.