Hyperspectral Sensor Management for UAS: Sensor Context Based Band Selection for Anomaly Detection

Today, hyperspectral sensors on small tactical drones for reconnaissance are a highly recommended research topic due to their rich spectral information. However, data processing and assessment can be extremely challenging. Especially for advanced reconnaissance tasks such as the detection and identification of explosive devices and camouflaged objects in unknown environments, the selection of spectral information from specific image bands is crucial. To overcome these challenges a new automated approach to an environment based spectral band selection method for atmospheric uncorrected hyperspectral images of such missions is presented here. The method is based on a k-means clustering procedure which first extracts the environmental context of the sensor. Subsequently, the deviation of the targets to this context is predicted by a Random Forest Regressor and bands with the highest target deviation can be selected based on this. Low computational effort is achieved by purposefully reducing the spectral and spatial resolution. The results show that the method has a high accuracy of prediction for both various by the model known and unknown targets and environments under realistic conditions.