Thermal remote sensing is an important tool for monitoring regional climate and environment, including urban heat islands. However, it suffers from a relatively lower spatial resolution compared to optical remote sensing. To improve the spatial resolution, various "data-driven" image processing techniques (pan-sharpening, kernel-driven methods, and machine learning) have been developed in the previous decades. Such empirical super-resolution methods create visually appealing thermal images; however, they may sacrifice radiometric consistency because they are not necessarily sensitive to specific sensor features. In this paper, we evaluated a "sensor-driven" super-resolution approach that explicitly considers the sensor blurring process, to ensure radiometric consistency with the original thermal image during high-resolution thermal image retrieval. The sensor-driven algorithm was applied to a cloud-free Moderate Resolution Imaging Spectroradiometer (MODIS) scene of heterogeneous urban and suburban landscape that included built-up areas, low mountains with a forest, a lake, croplands, and river channels. Validation against the reference high-resolution thermal image obtained by the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) shows that the sensor-driven algorithm can downscale the MODIS image to 250-m resolution, while maintaining a high statistical consistency with the original MODIS and ASTER images. Part of our algorithm, such as radiometric offset correction based on the Mahalanobis distance, may be integrated with other existing approaches in the future.