Hello. I assume you mean synthetic aperture radar (SAR) images taken from some altitude above the earth. If this is your data source, you can get plenty of images (many thousands). Assuming you want to use a supervised machine learning technique, you will also need ground truth, or annotated images (masks) that say where the fault zone is in each appropriate image. Once you have both image data {X(i)} and the corresponding masks {Y(i)}, you have a supervised learning problem to which many techniques are applicable. The main contenders are regression trees and forests, support vector machines and convolutional neural networks (CNN). CNNs can usually outperform the other approaches if you have enough training data, but they can be difficult to optimise and generally require a GPU rather than a CPU. If you don't use CNNs, you will need some kind of statistical description of what a fault zone looks like, that is, you need to characterise the fault zone in terms of its statistical features (edges, etc). Many feature descriptors are available for this purpose.