Improving anomaly detection with Multinormal Mixture Models in shadow

Hyperspectral images are well suited for automatic target detection, but detection performance in shadow is often degraded due to effects such as low signal-to-noise ratio, high dynamic range and spectral distortions. This paper focuses on improving target detection performance for a specific anomaly detector based on a statistical Multinormal Mixture Model (MMM) that is trained on the entire image to produce a global model of the background. It is demonstrated that a simple square root transformation and a hyperspheric transformation may be applied to the radiance image to enhance detection performance. A balancing strategy for the training of the model with respect to light level is shown to be a further improvement.