A stepwise framework for interpolating land surface temperature under cloudy conditions based on the solar-cloud-satellite geometry

Abstract

Thermal infrared land surface temperature (LST) data from satellites often contain extensive missing values due to high cloudiness degree, which severely hinders their use in applications. Despite the many methods developed, common methods, such as fusion with microwave or reanalysis data and the surface energy budget approach, still remain subject to important limitations and uncertainties, such as dependency on coarse resolution data and difficulty in interpolation for large-scale missing LST data. In this study, we proposed a stepwise framework for estimating missing cloudy-sky LST values of Moderate Resolution Imaging Spectroradiometer (MODIS) from informative samples owing to the solar-cloud-satellite geometry (SCSG) effect, by which satellite imagery records the cloudy-sky LST values of a portion of pixels. We first estimated the clear-sky LST equivalents for all cloud-affected pixels via a similarity-based approach and then determined unknown LSTs for cloudy pixels by training a machine-learning model on cloudy-sky LST values observed owing to the SCSG effect. We demonstrated the utility of this approach by using MODIS/Aqua daytime LST data over Qinghai-Tibet Plateau (QTP) and validated the interpolation results against representative in-situ LST observations and two recently published all-weather LST datasets. When compared to the corresponding in-situ measurements, the interpolated cloudy-sky LST values showed satisfactory accuracy with a mean absolute error (MAE) value of 3.99 °C and a coefficient of determination (R2) value of 0.74, while the MODIS/Aqua clear-sky LST values led to an MAE value of 2.66 °C and an R2 value of 0.86. Compared to the two all-weather LST datasets, results of this study showed the highest accuracy over the data-gap-filled regions in terms of all quantitative performance metrics, more natural transition textures, and better representation of seasonal characteristics. The proposed framework has the advantage of relying on the MODIS family data and handling extensive missing data as well as triggers opportunities to leverage the SCSG effect to produce high-quality all-weather LST data.