New Methodology to Predict Basin or Intrusion from Gravity Data, A Machine Learning Approach

Abstract

     Basins and Intrusions structures are essential features in defining and assessing the evolution of tectonic geo-structures. The gravity effects for Basin and Intrusion refer to (such as a salt dome or granitic pluton) structures that are similar in form, shape, and value. Attempts to characterize these structures from gravity data depend on derivation methods such as second horizontal and absolute second horizontal derivative methods. The task of the discriminator is to determine whether the data presented refers to a Basin or Intrusion. Hence, it is just a binary classifier giving the output as 0 (for Basin) or 1 (for Intrusion). The machine learning approach can solve such types of classification with high accuracy and confidence. Machine learning is a field concerned with algorithms that learn from data sets. Classification is a task that requires machine learning algorithms that learn from data sets how to assign a category label to examples from the problem domain. To learn the machine, how to classify the given data into 0 or 1, big data for training is needed. An easy-to-understand example would be classifying gravity data as "Basin, 0" or "Intrusion, 1". Later on, the learned machine can predict any given test data to the state of (0, 1). Therefore, the procedure is simply to prepare a huge synthetic data set (from 2D gravity modeling) for the Basin and Intrusion case. Then, divide the data sets into 80% data for training and 20% for testing. Label this 80% data set with 0 for Basin and 1 for Intrusion. Next, training these 80% data sets using some algorithms specifically designed for binary classification and do not natively support more than two classes. These include Logistic Regression and Support Vector Machines. A confusion matrix is used to evaluate the accuracy of learning. The following step lets the learned machine predict a label for the 20% data set. Python code programming is usually used for this type of analysis. This study uses an orange program for visual programming and data mining for training and predicting. The result of the prediction is perfect for the tested data. Field data for some cases from the Bougure gravity data of Iraq is tested with the learned machine and gives similar results to the absolute second horizontal derivative used. The saved model of the learned machine can be used to predict Basin or Intrusion case studies for future work.