Multiwavelength thermometry without a priori emissivity information: from promise to disillusionment

Abstract

Infrared (IR) thermography provides 2D radiance maps of the IR radiation leaving the surfaces of a scene, based on preliminary calibration. Then, to convert radiance maps into temperature maps, we need to know the emissivity of each element of the scene conjugated to each of the detector elements in the camera's focal plane. Compared to the single-band approach, can multispectral thermography help solve the inverse temperature-emissivity separation problem? Multiwavelength thermometry (MWT) is known to be an underdetermined problem having a continuous infinity of solutions. For this reason, and right from the origin of MWT, it appeared necessary to introduce information on emissivity to assess temperature, for example, by means of an analytical model. Overlooking these recommendations, a number of papers appeared in the early 2000s exploring the idea that we could do without any a priori knowledge of emissivity. Growth became exponential from 2020 onward, with the allegedly successful application of neural networks, genetic algorithms, and other novel optimization methods. The aim of the present work is to recall the consequences of MWT as an underdetermined inverse problem, to highlight the errors made by ignoring them, and to bring us back to harsh reality: we must introduce information on emissivity in order to evaluate temperature accurately. Furthermore, this information has to be fully consistent with the true spectral emissivity. To this end, we propose a new method for optimal selection of the emissivity model. A series of blind tests was set up to benchmark different inversion algorithms. The results confirm the failure of the temperature-emissivity separation when no emissivity information is available, despite the multiwavelength approach, which in any case comes as no surprise.