Understanding what lies beneath the Earth’s surface is important for many reasons, from finding natural resources to understanding geologic history. We use echoes of sound waves (seismic reflection) to do this, but interpreting that data is challenging because the recorded signals are complex and varied. We developed a method to choose which parts of the seismic signal, called “attributes”, are most useful for revealing patterns underground. We grouped these attributes into four categories based on what part of the signal they emphasize: how strong it is (Amplitude), what frequencies it contains (Spectral), how its arranged spatially (Structure/Geometry), and how contrasty or rough it is (Texture). Choosing one from each category helps us get a complete picture of what's happening below the surface.

We tested this method on data from the Taranaki Basin in New Zealand, using machine learning techniques to find patterns in the data without needing prior labels or training in grouping the data into zones with similar features, helping us identify likely geologic structures like channels and sediment layers. This approach can be a guideline in choosing input attributes to interpret complex seismic data using machine learning, and it can be adapted for other location targets and techniques.