Automatic detection of time-critical mobile targets using nonimaging, spectral infrared radiometric target signatures is explored. A novel set of classification features is developed for the spectral data and utilized in a Bayesian classifier. The processing results are presented, and sensitivity of the class separability to target set, target configuration, diurnal variations, mean contrast, and ambient temperature estimation errors is explored. This work introduces the concept of atmospheric normalization of classification features, in which feature values are normalized using an estimate of the ambient temperature in the vicinity of the target. The probability of detection, false alarm rate, and total error rate associated with this detection process is presented. Testing on an array of U.S. and foreign military assets reveals a total error rate near 5% with a 95% probability of detection and a concurrent false alarm rate of 4% when a single classification feature is employed. Sensitivity analysis indicates that the probability of detection is reduced to 70 to 75% in the hours preceding daylight, and that for the total error rate to be less than 10%, the target-to-background mean contrast must be greater than 0.025. Analysis of the atmospheric normalization technique reveals that to keep the total error rate less than 10%, the ambient temperature must be estimated with less than 3 K absolute accuracy.