Big Data Techniques for Applied Geoscience: Compute and Communicate

Big Data techniques have the potential to be paradigm-changing for applied geoscience if they are used widely. A significant number of such techniques, under the umbrella of Earth informatics, involve Machine Learning applied to high dimensional data to create new forms of value. This contribution presents two case studies of successful Earth informatics computation and the communication of the value of results, which provide insight into the uptake of ‘Big Data’ in geosciences. Machine Learning techniques split naturally into either supervised or unsupervised approaches. Supervised algorithms, such as Random Forests™ (RF), support vector machines or neural networks, share the concept of training a classifier using an initial (training) dataset. They are generally applied to predictive tasks, such as our first case study, predicting lithology from remote sensing and airborne geophysical data. Unsupervised algorithms, such as Self-Organising Maps (SOM), allow patterns inherent in the data to emerge without the use of a training dataset. They are generally applied to tasks which seek to explore patterns in data, such as our second case study, which identifies new potentially prospective river catchments. We find that calculating and presenting explicitly the newly extracted value, of the result obtained through computation, is an essential component of the post-compute evaluation. As strong advocates for the use of a range of Big Data techniques in applied geosciences, we conclude that the benefits to be gained from the way that we ‘compute’ can be lost if we do not also take considerable care with the ways that we ‘communicate’.