Feature fusion-enhanced t-SNE image atlas for geophysical features discovery
摘要
The discovery and identification of geophysical features from diverse gridded datasets play a pivotal role in understanding geological phenomena. Traditional tools tailored to identify specific signatures, such as lineaments in magnetic data, do not account for (1) the naturally occurring complexity of the causative geology which results from competing and overprinting processes and (2) often overlook important related patterns present in other forms of data (e.g., gradients in the associated gravity data). To address those limitations, we propose a two-step data-driven approach that integrates diverse gridded datasets and autonomously reveals inherent patterns without predefined assumptions in terms of feature geometry. Utilizing Haralick texture descriptors, we encode data patches from different geophysical datasets as points in a high-dimensional unified representation space, facilitating seamless data fusion. This representation is then nonlinearly projected into a two-dimensional space using t-distributed stochastic neighbor embedding (t-SNE), forming an interactive “t-SNE Atlas”. This atlas positions tiles from each dataset according to their position in the t-SNE while linking them back to their original geographical coordinates. Such a setup allows earth scientists to intuitively navigate through the data, exploring complex relationships between geophysical responses and geological structures, thus facilitating the discovery of new insights, the formulation of hypotheses, and the exploration of non-trivial connections. This is illustrated using magnetic and gravity data covering