Preliminary geological mapping with convolution neural network using statistical data augmentation on a 3D model.

Airborne magnetic data are commonly processed and interpreted by geologists to produce preliminary geological maps of unexplored areas. Machine learning can partly fulfill this task rapidly and objectively as convolutional neural networks applied to image segmentation showed promising results in comparable applications. As these algorithms require a large and high-quality dataset to be trained, we developed an innovative geostatistical data augmentation workflow that uses a 3D conceptual lithological and magnetic susceptibility model as input. The workflow uses soft-constrained Multi-Point Statistics to create equiprobable synthetic 3D geological models and Sequential Gaussian Simulation to populate these models with a geologically meaningful magnetic susceptibility spatial distribution. Then, geophysical forward modeling is computed to get the airborne magnetic responses of the synthetic models, which are associated with their counterpart synthetic geological maps. We applied this workflow on a 3D model of the Malartic Mine area to obtain a large synthetic airborne magnetic and counterpart geological map dataset. Then, a Gated Shape Convolutional Network is trained on this dataset to perform a preliminary geological mapping. A semi-supervised approach using clustering on the feature maps from the trained network is also implemented. Testing the trained network on synthetic data and remote areas of similar geological context shows that the methodology is suitable for producing preliminary geological maps using airborne magnetic data. Notably, the clustering module shows a precise and adaptive segmentation of the magnetic anomalies into pertinent preliminary geological maps. The quality of the results empirically validates our data augmentation method. Our method allows producing geological maps by training a convolutional neural network using an area where a detailed geological and petrophysical 3D model exists and then permits applying the pre-trained model in new areas of the same geological context, where only airborne magnetic data is available. • Multi-Point Statistics, Sequential Gaussian Simulation and Forward modeling can be used to perform data-augmentation using a 3D Magnetic Susceptibility and Geological Model as input. • Convolutional Neural Networks perform to generate preliminary geological mapping using only airborne magnetic data. • Apply clustering on the feature maps of the CNN provides pertinent preliminary geological mapping. • Gated-Convolutional Layers provide pertinent representations of an airborne magnetic data. [ABSTRACT FROM AUTHOR]