The main purpose of the paper is to show that significant improvements in infrared land mine detectors can be achieved, by also considering visual wavelength images. A Bayesian approach, based on dual-band data, is presented that incorporates prior knowledge regarding external parameters such as weather, burial depth and soil moisture. By noting that most relevant backgrounds render rotationally invariant statistics, a low dimensional parameterization of the noise space is derived. Simulations show the performance of three different detectors; first the standard detector used, the matched filter which correlates the infrared image with the known mine shape; secondly a detector which models the spatial statistics of the infrared background while neglecting the visual wavelength data, and thirdly the proposed detector that exploits the full dual-band space. The second detector outperforms the matched filter, and is significantly improved by also utilizing the visual wavelength image.