Downscaling of Landsat LST with HotSat-1 data and generative adversarial networks

Land Surface Temperature (LST) significantly affects the Earth's energy balance, making it vital for various environmental and scientific studies. Currently, the highest-resolution satellite-based LST products are derived from Landsat thermal infrared (TIR) data at 100 m resolution, which is insufficient for capturing fine-scale thermal patterns and in urban areas. Mid-wave infrared (MWIR) observations provide an alternative by being less sensitive to atmospheric effects and emissivity errors than TIR, enabling more accurate thermal characterization than other commonly used auxiliary data. HotSat-1, a recently launched satellite, carries MWIR sensor operating in the 3.7–5 μm range and provides thermal intensity data at approximately 5 m spatial resolution, suitable for high resolution thermal analysis. In this study, present a generative adversarial network (GAN) based approach to downscale LST information for three study areas (Old Fig Garden, Kherson, and Gijang), using Landsat-9 and HotSat-1 imagery across three study areas: Old Fig Garden, Kherson, and Gijang. Compared to the baseline Landsat LST, the downscaled results showed enhanced spatial coherence and clearer boundary delineation. Among the tested models, Super-Resolution Generative Adversarial Networks (SRGAN) incorporating HotSat-1 data outperformed other methods, achieving the highest accuracy with RMSE of 0.64 and PSNR of 54.65. These results highlight the feasibility of integrating MWIR observations and super-resolution techniques for LST downscaling in complex urban environments. • HotSat-1 MWIR imagery is fused with Landsat LST using a GAN-based super-resolution model • The method achieves strong performance across diverse areas with RMSE as low as 0.64 K. • The proposed method downscales 100 m Landsat LST to 30 m with reliable spatial detail • HotSat data better preserve fine-scale LST patterns than other spectral indices