Connectivity-informed Drainage Network Generation using Deep Convolution Generative Adversarial Networks

Stochastic network modeling is often limited by high computational costs to\ngenerate a large number of networks enough for meaningful statistical\nevaluation. In this study, Deep Convolutional Generative Adversarial Networks\n(DCGANs) were applied to quickly reproduce drainage networks from the already\ngenerated network samples without repetitive long modeling of the stochastic\nnetwork model, Gibb's model. In particular, we developed a novel\nconnectivity-informed method that converts the drainage network images to the\ndirectional information of flow on each node of the drainage network, and then\ntransform it into multiple binary layers where the connectivity constraints\nbetween nodes in the drainage network are stored. DCGANs trained with three\ndifferent types of training samples were compared; 1) original drainage network\nimages, 2) their corresponding directional information only, and 3) the\nconnectivity-informed directional information. Comparison of generated images\ndemonstrated that the novel connectivity-informed method outperformed the other\ntwo methods by training DCGANs more effectively and better reproducing accurate\ndrainage networks due to its compact representation of the network complexity\nand connectivity. This work highlights that DCGANs can be applicable for high\ncontrast images common in earth and material sciences where the network,\nfractures, and other high contrast features are important.\n