Emissivity spectra recovered from spectral radiance images may have lowered spectral contrast due to irradiance from nearby surface elements ('cavity effect'). For analysis based only on photointerpretation or Reststrahlen band identification, it is not always necessary to account for cavity effects, but for full spectral analysis it may be desirable. We present an approach to compensate thermal infrared (TIR) images for cavity radiation. This approach is based on optical estimates of subpixel surface roughness and estimation of cavity contribution for different natural surfaces using a TIR radiosity model. It was tested using tripod-mounted Hyper-Cam (Telops, Inc., Quebec City, Canada) hyperspectral TIR images of natural targets from the Mojave Desert, California, USA, along with centimetre-scale digital elevation models of similar targets measured by ground lidar. For remote subpixel roughness estimation, Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) nadir- and aft-looking (27.6°) near-infrared (NIR) brightness ratios, as well as synthetic aperture radar (SAR) images calibrated to roughness root mean square (RMS), were used. The TIR compensation approach is adaptable for different spectral resolutions, including hyperspectral.