Classification of ground targets has been carried out using simulated IR and radar data in combination. The observations are obtained by a missile approaching one of six different, already detected, targets. IR images are generated from textured models of the scene and targets using a 3-D tool. Radar High Resolution Range profiles (HRR) are calculated from detailed CAD models based on physical optics. To these profiles is added clutter using a model in coherent mode. IR features are primarily based on geometric invariants extracted from the IR image. Radar features comprise the entire range profile. A preliminary study of "super resolved" radar profiles has been done based on the so-called MUSIC algorithm. The targets are classified using an artificial neural net applied to the extracted features, both for individual sensors and in combination. Two different data fusion methods have been studied: feature fusion, using a common feature vector, and decision fusion, where individual sensor decisions are weighed together supported by confidence weights from earlier experiments. The results show that feature fusion equals or surpasses the best sensor given that one knows the uncertainty of data. Decision fusion, in turn, surpasses feature fusion as far as a priori data can be trusted. Local decisions have been updated time-wise by using Bayes rule in recursive mode.