Towards reducing the cloud-induced sampling biases in MODIS LST data: a case study from Greenland

Satellite-driven Land Surface Temperature (LST) datasets are essential for characterizing climate change impacts on terrestrial ecosystems, as well as a wide range of surface-atmosphere studies. In the past one and a half decade, NASA's Moderate Resolution Imaging Spectroradiometer (MODIS) has provided the scientific community with LST estimates on a global scale with reasonable spatial resolution and revisit time. However, the use of MODIS LST for climate studies is complicated by the simple fact that the observations can only be made under clear-sky conditions. In regions with frequent overcast skies, this can result in the calculated climatic variables deviating from the actual surface conditions. In the present study, we propose and validate a framework based on model-driven downwelling radiation data from ERA-Interim and instantenous LST observations from both MODIS Terra and Aqua, in order to minimize the clear-sky sampling bias. The framework is validated on a cloud-affected MODIS scene covering parts of Greenland (h15v02), and by incorporating in-situ data from a number of monitoring stations in the area. The results indicate that the proposed method is able to increase the number of daily LST estimates by a factor of 2.07 and reduce the skewnewss of monthly distribution of the successful estimates by a factor of 0.22. Considering that these improvements are achieved mainly through introducing data from partially overcast days, the estimated climatic variables show better agreement with the ground truth. The overall accuracy of the model in estimating in-situ mean daily LST remained satisfactory even after incoprporating the daily downweling radiation from ERA-interim (RMSE=0.41 °K, R-squared=0.992). Nonetheless, since technical constraints are expected to continue limiting the use of high temporal resolution satellites in high latitudes, more research is required to quantify and deal with various types of cloud-induced biases present in the data from moderate resolution sensors.