DEVELOPMENT OF AN OPEN SOURCE, MACHINE LEARNING BASED TOOLSET FOR THE IDENTIFICATION OF DIKES IN SATELLITE IMAGES THROUGH SEMANTIC SEGMENTATION

Large Igneous Provinces (LIPs) are the largest magmatic events in Earth history with large volumes (> 500000km³) of predominantly basaltic lavas covering huge continental and oceanic regions (> 100,000 km²). LIP emplacement is frequently correlated with severe climate and ecosystem perturbations including mass extinctions. However, we still lack a good understanding of one of the most fundamental questions related to LIPs: How did these large LIP flows erupt from the underlying magma chambers? This is a critical parameter to constrain the nature of the LIP magmatic system and its similarities or differences compared to modern volcanic systems such as Hawaii. For many LIPs (e.g., Columbia River Basalts, Deccan Traps), the lack of large volcanic edifices suggests that the lava flows were fed by an extensive dike system comprising 1000s of dikes, some of which can be 10s of km long. The dikes' density distribution and their orientations provide critical information about the crustal magmatic system and the stress state of the crust at the time of eruptions. However, despite a large number of dikes, accurate maps only exist in only a few cases and can have spatial biases due to accessibility and exposure quality. We describe a new tool to address this challenge using global high-resolution satellite images and digital elevation maps. Our tool applies a machine learning approach to this data to predict the presence of dikes and similar geographic features. We evaluated and trained a number of potential artificial neural net architectures, with the most promising preliminary results from DeepLabV3 and a generative adversarial model. We trained these networks to perform semantic segmentation on training data. The input layer comprises a concatenation of five spectral bands and their corresponding mask, and the final output consists of line segments indicating predicted dikes. Our tool performs curation, training, feed-forward neural network operation on new data and post-processing to generate the desired output. We present results of our trained model for a set of images from the Deccan Traps in India and compare the model predictions with dikes known from fieldwork. We show that machine learning is a powerful method for mapping geologic features such as dikes for LIPs and similar features on planetary bodies (e.g., Venus).