Generating 5 km resolution 1981–2018 daily global land surface longwave radiation products from AVHRR shortwave and longwave observations using densely connected convolutional neural networks

Surface longwave radiation (SLWR) components, including downward longwave radiation (DLR), upward longwave radiation (ULR), and net longwave radiation (NLR), are major contributors to the Earth's surface radiation budget and play important roles in ecological, hydrological, and atmospheric processes. Previous SLWR products have different drawbacks, such as being temporally short (after 2000), spatially coarse (≥ 25 km), and instantaneous values, which hinder their in-depth applications in land surface process modeling and climate trends analysis. Here, we reported the Advanced Very High-Resolution Radiometer (AVHRR)-based Global LAnd Surface Satellites (GLASS-AVHRR) SLWR products over the global land surface at a 5 km spatial resolution and 1 day temporal resolution between 1981 and 2018. These products were generated using multiple densely connected convolutional neural networks (DesCNNs) from the AVHRR top-of-atmosphere (TOA) reflected and emitted observations and European Centre for Medium-Range Weather Forecasts (ECMWF) fifth generation reanalysis (ERA5) near-surface meteorological data. DesCNNs were trained using integrated SLWR samples derived from the Moderate Resolution Imaging Spectroradiometer (MODIS)-based GLASS, Clouds and the Earth's Radiant Energy System Synoptic (CERES-SYN), and ERA5 SLWR products. In situ measurements from 231 globally distributed sites were used to evaluate the GLASS-AVHRR SLWR estimates. The results illustrated the overall high accuracies of GLASS-AVHRR SLWR products with root-mean-square-errors (RMSEs) of 18.66, 14.92, and 16.29 Wm−2, and mean bias errors (MBEs) of −2.69, −3.77, and 0.49 Wm−2 for all-sky DLR, ULR, and NLR, respectively. We found good correlation and consistency between GLASS-AVHRR and both CERES-SYN and ERA5 in terms of spatial patterns, latitudinal gradient, and temporal evolution. Our results revealed the significant contribution of shortwave observations to SLWR estimation owing to the high amounts of clouds over polar regions and water vapor and clouds in tropical areas, which was not previously widely recognized by the remote sensing community. GLASS-AVHRR SLWR products were updated, documented, and made available to the public at www.glass.umd.edu and www.geodata.cn.