Jointing Adjacent Environmental Variation Into a Deep Learning Model for Gap-Filling Passive Microwave-Based Land Surface Temperature

Passive microwave-based land surface temperature (PMW LST) serves as a significant source for complementary thermal infrared LST, whereas the orbit gaps frequently result in missing data. Up to now, many studies have proposed methods to fill these gaps in PMW LST. However, most of these methods depend on the assumption that the missing LST is similar to that of adjacent days, yet the natural environment changes may lead to this assumption not being established. To address this, we proposed a comprehensive deep-learning model that incorporates three groups of natural variables, including atmosphere, land environment, and radiation, from both the target and adjacent days. Simultaneously, we employ two advanced microwave scanning radiometer (AMSR) LST-based simulated validations and six in-situ measurements to evaluate the model's gap-filling performance. According to the results, the proposed model achieves root mean squared error (RMSE) of 1.87 K/1.89 K and 1.69 K/1.71 K for the two AMSR LST-based validations during the daytime/nighttime. Compared with the inverse distance weighted method and an advanced deep learning model, the proposed approach improves 0.27–0.5 K (12.6% –22.6%) and 0.14–0.3 K (6.9% –14.9%) during daytime and nighttime, respectively. Furthermore, based on the results of six in-situ measurements, the gap-filled results gain the average RMSE of 3.7 K and 3.21 K during the daytime and nighttime, respectively. In addition, we find that the land environment and radiation conditions have a stronger impact during the daytime, while atmospheric conditions are more sensitive at night. These findings present a more scientific and effective gap-filling method, potentially enhancing the accuracy of land thermal environment research.