Multi-Source Hyperspectral Data Integration using Image Feature Extraction for Mineral Mapping

The fusion of remote sensing datasets for combined classifications has recently received great attention. Separate data acquisition and subsequent fusion allow for sensor-specific adjustments of experimental parameters and flexible sensor combinations. In mineral exploration, multi-sensor approaches are particularly promising to extend the range and accuracy of mineral phase detection. However, the time-efficient and accurate integration of data with differing resolution, field of view or acquisition modi remains challenging. This motivates us to promote a multi-source data integration based on efficient feature extraction strategies. An important pre-requisite for a successful integration is data co-registration. A workflow combining automated keypoint detection and matching provides an accurate and fast method to align multi-sensor datasets of different spatial sampling distances for a joint processing. For feature extraction, we employ innovative methods that consider both spatial and spectral aspects such as Orthogonal Total Variation Component Analysis (OTVCA). Besides the reduction of dimensionality and consequently processing time, the method allows the integration of textural and spectral information, such as short wave and long wave hyperspectral imagery or reflectance and luminescence data that are not directly interpretable with a single method. Application-relevant mineral assemblage classes such as alteration zones, ore zones or mineralized veins, become then discriminable as the spatio-spectral patterns are evident by extracting relevant features from all datasets. We use a Support Vector Machine with Radial Basis Function Kernel (SVM-RBF) to demonstrate the classification of mineralogical domains from such multi-source feature fused sets. We choose SVM as it is particularly robust when handling high-dimensional data with low number of training samples and against the heterogeneity of classes that is typical for mineralogical datasets. The optimal classification parameters are determined by five-fold cross- validation to ensure the best possible classification result with the given data. Training data required for the classification can be defined according to user knowledge, high resolution mineralogical analyses or spectral point measurements. The proposed multi-source data integration workflow shows to exceed the classification accuracy of single-source data and could be beneficial for many potential application fields in mineral exploration, mineral processing, recycling or food industry.