Infinite Gaussian mixture models for robust decision fusion of hyperspectral imagery and full waveform LiDAR data

In this work, we propose a multi-sensor decision fusion approach which effectively combines hyperspectral data with full waveform LiDAR (Light Detection And Ranging) data for robust geospatial image classification. The proposed approach provides a platform to exploit the most relevant information provided by these two remote sensing modalities. The hyperspectral data and full waveform LiDAR data are classified via a multi-classifier system, following which a decision fusion system based on logarithmic opinion pools (LOGP) is employed. The “base” classifier employed here is the infinite Gaussian mixture model (IGMM) based classifier which can capture the potentially multi-modal statistical structure of hyperspectral data and full waveform LiDAR data very well. As a preprocessing step, we use Local Fisher´s Discriminant Analysis (LFDA) for dimensionality reduction since we expect it to preserve the non-Gaussian, potentially multi-modal structure of the hyperspectral data in the lower dimensional subspace. We compared our proposed multi-classifier based decision fusion method with a feature fusion scheme (the traditional way to do such analysis) which stacks the hyperspectral data and LiDAR waveform data and invokes a single classifier on this extended data cube. The results of our experiments show that the proposed IGMM-LOGP algorithm achieves a promising classification performance and works better than traditional feature fusion.