LFSR: Low-resolution Filling then Super-resolution Reconstruction framework for gapless all-weather MODIS-like land surface temperature generation

Abstract

Due to the great advancements in land surface models (LSMs), integrating data from thermal infrared (TIR) and LSMs is a promising way for obtaining gapless all-weather land surface temperature (LST). However, the differences of spatial resolution and discrepancy of data acquisition ways between TIR LST and model-simulated LST usually brought great challenges to traditional methods in terms of accuracy and texture details. This study proposes a low-resolution filling then super-resolution reconstruction (LFSR) framework for generating gapless all-weather LST using Moderate Resolution Imaging Spectroradiometer (MODIS) LST and China Meteorological Administration Land Data Assimilation System (CLDAS) LST. For the LFSR, a multi-source multi-temporal low-resolution filling (MSMTLF) network is first designed to alleviate the discrepancy of data acquisition ways between the MODIS LST and CLDAS LST, and generate gapless low-resolution degraded LSTs. A multi-scale multi-temporal super-resolution reconstruction (MSMTSR) network is then used to reconstruct the gapless low-resolution degraded LSTs into gapless high-resolution MODIS-like LSTs with rich-texture, which is mainly used to deal with resolution differences between the two LSTs. The experiments suggested that the LFSR achieved satisfactory results, and the maximal RMSE is less 2.5 K in the simulated experiments. When validated against the in-situ LST data under clear and cloudy skies, the small difference of the overall average bias (−0.91 K for clear skies VS -0.88 K for cloudy skies) and overall average RMSE (4.15 K for clear skies VS 5.68 K for cloudy skies) were obtained. Compared with results from the different input data, the other strategies and the other methods, the generated gapless all-weather MODIS-like LSTs from the LFSR were closer to the actual labels or have better consistency and spatial details. These results indicated the LFSR achieves impressive performance for fusing MODIS and CLDAS data. The LFSR actually provides a new framework for fusing TIR LST and simulation-based LST with considerable data inconsistency, and has the potential for generating gapless all-weather TIR LST records.