A two-step deep learning framework for mapping gapless all-weather land surface temperature using thermal infrared and passive microwave data

Abstract

Blending data from thermal infrared (TIR) and passive microwave (PMW) measurements is a promising solution for generating the all-weather land surface temperature (LST). However, owing to swath gaps in PMW data and the resolution inconsistence between TIR and PWM data, spatial details are often incomplete or considerable losses are generated in the all-weather LST using traditional methods. This study was conducted to develop a two-step deep learning framework (TDLF) for mapping gapless all-weather LST over the China's landmass using MODIS and AMSR-E LST data. In the TDLF, a multi-temporal feature connected convolutional neural network bidirectional reconstruction model was developed to obtain the spatially complete AMSR-E LST. A multi-scale multi-temporal feature connected generative adversarial network model was then designed to blend spatially complete AMSR-E LST and cloudy-sky MODIS LST, and generate gapless all-weather LST data. Gapless all-weather LST data were evaluated using six in-situ LST data from the Tibetan Plateau (TP) and the Heihe River Basin (HRB). The root mean squared errors (RMSEs) of the gapless all-weather LST were 1.71–2.0 K with determination coefficients (R2) of 0.94–0.98 under clear conditions, and RMSEs of 3.41–3.87 K and R2 of 0.88–0.94 were obtained under cloudy conditions. Compared to the existing PMW-based all-weather LSTs, the validation accuracy and image quality (such as spatial detail) of the generated gapless all-weather LSTs were superior. The TDLF does not require the use of any additional data and can potentially be implemented with other satellite TIR and PWM sensors to produce long-term, gapless, all-weather MODIS LST records on a global scale. Such a capability is beneficial for generating further gapless all-weather soil moisture and evapotranspiration datasets that can all be applied in global climate change research.