An optimal sampling method for multi-temporal land surface temperature validation over heterogeneous surfaces

Abstract

The development of ground-based sampling strategies is vital for validating medium or coarse-resolution satellite-derived land surface temperature (LST) products. Conventional LST sampling at the satellite pixel scale has been limited to the homogeneous surfaces. In this study, an optimal sampling strategy called spatial and diurnal temperature cycle-constrained sampling (SDCS) is proposed to extend the feasibility of LST validation over heterogeneous surfaces with dramatic diurnal LST changes. SDCS integrates a priori information including land cover, diurnal LSTs data, and spatial distribution characteristics of samples to improve the representativeness of multi-temporal measurements over heterogeneous surfaces. SDCS was applied to four varied study areas using simulated satellite data and was compared with four existing methods including random sampling, systematic sampling, land cover-based stratified sampling, and conditioned Latin hypercube (CLH) sampling. Results showed that SDCS could significantly improve the representativeness of samples with a limited sample size. In homogeneous surfaces with an LST standard deviation (SD) of less than 2 K, the root mean square error (RMSE) of diurnal LSTs estimated by SDCS was less than 0.3 K when using 0.20% of the total pixels. In moderate heterogeneous surfaces (LST SD less than 5 K), 0.32% of the total pixels were required to achieve RMSE less than 0.5 K. In extremely heterogeneous surfaces (LST SD > 6 K), 0.96% of the pixels were needed to achieve the same accuracy. Further, the representativeness of the samples selected by SDCS was stable in diurnal space with uncertainties of LST bias less than 0.27 K. Moreover, the samples exhibited a dispersed spatial distribution with a nearest neighbor index of 1.27–1.54. SDCS can generalize to various regions and dates and can be employed in field campaigns for diurnal LST validation over heterogeneous surfaces.