Decision fusion strategy for target recognition in hyperspectral images

Hyperspectral sensors allow a considerable improvement in the performance of a target recognition process to be achieved. This characteristic is particular interesting in a lot of military and civilian remote sensing applications, such as automatic target recognition (ATR) and surveillance of wide areas. In this framework, real time processing of the observed scenario is becoming a key issue, because it permits the operator to provide immediate assessment of the surveyed area. In the literature is presented a line-by-line real time implementation of the widely used Constrained Energy Minimization (CEM) target detector. However, experimental results show that sometimes the CEM filter produces False Alarms (FAs) corresponding to rare objects, whose spectra are angularly very different from the target signature and from the natural background classes in the image. A solution to such a problem is presented in this work: the proposed strategy is based on the decision fusion of the CEM and the SAM algorithms. Only those pixels that pass the CEM-stage are processed by the SAM algorithm. The second stage allows false alarms to be reduced by preserving most of target pixels. The fusion strategy is applied to an experimental hyperspectral data set to recognize a known green target. Detection performance is numerically evaluated and compared to the one of the classical CEM detector.