Physics-Informed Annotation for Learning-Based Hyperspectral Mineral Mapping

The retrieval of drill cores is a costly component of mineral exploration. Improving the spatial overview of mineral abundances within a deposit can substantially reduce the need for drilling. We present an unsupervised, automated annotation strategy for pixel-wise mineral labeling in hyperspectral imagery of simple deposit styles. In this context, a simple deposit style refers to deposits with very low or no mineral transitions and predominantly homogeneous, dominant mineral occurrences.

The automated annotation is based on handcrafted, mineral- and deposit-specific normalized difference indices (NDI). The objective is to extract a large number of representative mineral spectra for each occurring mineral. These spectra are subsequently used as training data for a targeted hyperspectral neural network with positional encoding, which is expected to generalize better to more complex deposit styles.

As a first step, the normalized mineral indices were successfully learned by the network, achieving an F-score of 0.98. This result represents a promising step toward physics-informed, neural-network-based mineral classification in hyperspectral imagery.