An Integrated Machine Learning Geological Field Tool In The Age Of Big Data And Drones

Unmanned aerial vehicles (UAVs) and machine learning are relatively new research tools in the geosciences that can be used to collect and analyze large data sets rapidly. We combine the power of rapid data collection using unmanned aerial vehicles with machine learning algorithms to develop a field-based system to identify targeted geological features. For data collection, we have used a commercial-grade UAV which carried visible-wavelength and multispectral (visible- infrared) cameras. We analyzed the data with machine learning and machine vision algorithms that can classify rock units exposed in a field area. We have identified algorithms that in previous literature have proven to be reliable in predicting characteristics. These include k-Nearest Neighbors, Support Vector Machines, K-Means Clustering and Convolutional Neural Networks which we tested to determine their effectiveness in geologic mapping. Several filtering operations were applied to the collected imagery. K-Means clustering was used to generate some preliminary segmentation of lithology on the ground. We then ran 3 supervised machine learning algorithms and the results were field tested to determine how they compared to a professional grade geologic map. The results show that consumer-grade machine learning algorithms and consumer-grade UAV platforms can be integrated into a package that can act as a field assistant and generate preliminary geologic maps. The accuracy results of the algorithms were tested by employing a test/train split of the labeled dataset. The highest accuracy reported was for the SVM algorithm with an accuracy reported of 76% in the classification of 2 visually distinct rock units. We have also shown that by introducing Kuwahara and Median filters, normally used in computer vision applications, we can reduce effects of vegetation and increase the prediction accuracy