Landsat LST Product Gap Filling Using the Multiple Features Ensemble Learning Model Caused by the Missing Emissivity Dataset

Land surface temperature (LST) is an important surface physical parameter that has received wide attention in many fields, and thermal infrared (TIR) remote sensing technology can efficiently obtain large-scale LST information based on the LST retrieval algorithms. Various LST remote sensing products are available, among which Landsat data has formed an important data accumulation with its long time series and high spatial resolution. Since it is formed by a single-channel algorithm relying on the ASTER Global Emissivity Dataset (GED), which is not spatially seamless, the LST product also suffers from missing data in the corresponding emissivity aperture region. In this study, a multiple feature reconstruction (MFR) method, which establishes the regression relationship between three types of feature parameters, namely, surface elevation, reflectance spectral, solar radiation, and LST, is developed to fill the gaps due to missing emissivity to improve the quality and availability of the Landsat LST product using the LightGBM ensemble learning algorithm. The proposed method is applied to the missing regions of the LST product in two study areas, and the spatial distribution of the existing LST products is utilized to match the filling results to obtain the corrected gap-filling LST product based on the preliminary results. Besides, the split-window algorithm retrieves the LST using the thermal radiance observed at the top of the atmosphere (TOA), and the gap-filling results of both study areas are cross-validated. The results of the accuracy validation indicate that the MFR method can fill the vacant positions in the LST product well and maintain a smooth spatial distribution. The development of an end-to-end LST retrieval algorithm directly from TOA thermal infrared observations to LST can directly avoid the occurrence of incomplete LST remote sensing products caused by insufficient auxiliary parameters. In addition, the MFR method can be also applied to under-cloud LST estimation and LST downscaling utilizing its ability to characterize LST based on other remote sensing data sources.