A Probabilistic Framework for Fusing Classifications Derived From Multi-Temporal Hyperspectral Imagery

A new framework to fuse probabilistic classifications from replicate hyperspectral imagery of the same scene is presented. To improve scene classification accuracy, probabilistic outputs from a classifier, such as a Gaussian Process (GP), are fused. The framework allows fusion of several $(n\geq 2)$ images simultaneously or sequentially. The framework has been tested using hyperspectral imagery acquired from field-based platforms from a mine face at four different times during the day under different illumination conditions. Classification results of individual images showed large variations, however, using the fusion framework, the fused map showed a better agreement with the geology mapped in the field.1