A Local Temperature Unmixing-Based Fusion Model for Land Surface Temperature Spatiotemporal Enhancement

Spatiotemporal fusion algorithms have the ability to obtain high-resolution land surface temperature (LST). However, most existing fusion methods have difficulty obtaining LSTs with a high resolution corresponding to that of visible and near-infrared (VNIR) data and cannot adapt well to LSTs with complex nonlinear changes. Hence, this study proposes a local temperature unmixing-based fusion model (LTUBFM), which is designed for LSTs with complex variations at both spatial and temporal scales, to obtain high-resolution LSTs. The LTUBFM contains three parts. First, component temperatures are estimated based on a linear temperature mixing model (LTMM) from low-resolution images. Second, component temperatures are redistributed by neighboring similar pixels via a weighted function. Finally, model residuals are derived from redistributed component temperatures and interpolated using an inverse distance weighting (IDW) interpolator to obtain high-resolution LSTs. Compared to existing fusion algorithms, LTUBFM has the following strengths: 1) it can weaken the effect of land cover change on LST; 2) it can fuse LSTs to a high resolution corresponding to that of VNIR data; and 3) it is strongly robust even when the temperature changes greatly. The LTUBFM is tested with simulated data and satellite data and fuses 1000 m-resolution Moderate Resolution Imaging Spectroradiometer (MODIS) LSTs into 30 m. Compared with STARFM, ESTARFM, and SPSTFM, LTUBFM reduces the RMSE by up to 3.5 K in areas covered with vegetation. These results indicate that the LTUBFM has the potential to generate satisfying high-spatiotemporal-resolution LSTs from multisource datasets with large temperature changes.