High-resolution thermal sharpener supported by generative techniques: case study in Perth, Australia

Land surface temperature (LST) is an important variable in understanding climate change, detecting forest fires, and monitoring the environment. Satellite Earth observation of LST represents a cost-effective tool for detecting thermal variations across large scenes. To provide downscaling of the LST images, the High-resolution Urban Thermal Sharpener (HUTS) algorithm is conventionally applied to higher-resolution bands of the satellite image. Since the NIR and two SWIR channels in high resolution are needed for this computation but are not available in general, we propose a method for its retrieval using generative techniques. We utilize InfraGAN, a state-of-the-art method for infrared image generation, to train a model, which was then applied on suitably pre-processed high-resolution images. Furthermore, we compute LST using the HUTS method and the generated NIR and SWIR images. We validate the accuracy using ground measurements by recording the temperature at different locations of an Australian urban environment. Finally, in the absence of ground-truth downscaled data for NIR and SWIR, a detailed evaluation of the proper network performance is provided to assess the results of DL-based methods.