A framework is proposed for selection of wavelength bands for multispectral sensors using hyperspectral reference data. Using results from detection theory we derive a cost function that is minimised by a set of spectral bands optimal in terms of detection performance for discrimination between a class of small rare targets and clutter with known spectral distributions. The method may be used, e.g., in the design of IRST and electro-optical missile warning sensors, where low false-alarm rate and high detection probability for detection of small targets against a clutter background are of critical importance, but the required high frame-rate prevents the use of hyperspectral sensors.