A polynomial regression approach to subpixel temperature extraction from a single-band thermal infrared image

Target temperature estimation from thermal infrared (TIR) imagery is a complex task that becomes increasingly more difficult as the target size approaches the size of a projected pixel. At that point the assumption of pixel homogeneity is invalid as the radiance value recorded at the sensor is the result of energy contributions from the target material and any other background material that falls within a pixel boundary. More often than not, thermal infrared pixels are heterogeneous and therefore subpixel temperature extraction becomes an important capability. Typical subpixel estimation approaches make use of data from multispectral or hyperspectral sensors. These technologies are expensive and data collected by a multispectral or hyperspectral thermal imagery might not be readily available for a target of interest. A methodology has been developed to retrieve the temperature of an object that is smaller than a projected pixel of a single-band TIR image using physics-based modeling. The process can be broken into two distinct pieces. In the first part, the Digital Imaging and Remote Sensing Image Generation (DIRSIG) tool will be used to replicate a collected TIR image based on parameter estimates from the collected image. This is done many times to build a multi-dimensional lookup table (LUT). For the second part, a regression model is built from the data in the LUT and is used to perform the temperature retrieval. The results presented are from synthetic imagery.