Continuous evaluation of the spatial representativeness of land surface temperature validation sites

Highlights

• A spatial representativeness evaluation method was proposed on T-based validation.

• The method (TVM) relies on the physical mechanism of LST temporal variation.

• TVM describes a ground site's daily spatial representativeness for satellite LST.

• The method is tested over 16 Chinese measurement sites with MODIS/AATSR LST products.

• The uncertainty caused by spatial representativeness in LST validation is quantified.

Abstract

A reliable accuracy is essential for the application of land surface temperature (LST) products. Current satellite retrieved LSTs are mainly validated over a few homogeneous sites. However, most of the existing ground sites are located in inhomogeneous areas: thus, their spatial representativeness on satellite pixel scales is unknown. In this situation, how to evaluate the spatial representativeness of these inhomogeneous sites, quantify the influence introduced by the spatial representativeness and describe the variation of the site's spatial representativeness are critical questions for satellite LST validation. In an attempt to answer those questions, a so-called temporal variation method (TVM) is proposed for evaluating a ground site's spatial representativeness. The method defines a spatial representativeness indicator (SRI), which is the LST difference between a ground radiometer's field-of-view (FOV) and a satellite pixel, and describes a site's spatial representativeness. Based on the temporal variation of LST, the SRI time series consists of three temporal components: ∆ATC, ∆DTC_F-P, and ∆USC, which describe the annual, diurnal, and instantaneous variations of SRI, respectively. Associated with the Landsat TM/ETM+ data and weather parameters, the method is implemented and tested at 16 Chinese ground sites for the validation of MODIS and AATSR LST products. Results show that the temporally continuous SRI (SRI_TPR) shows high correlations with the original SRI (SRI_ORI). The variation of SRI_TPR is mainly determined by changes in surface coverage (i.e. NDVI difference on the two scales) and affected by weather conditions (e.g. near-surface air temperature, accumulative downward solar radiation, and wind speed). Since the SRI is defined as the LST difference between the two scales, it can be used as a bridge to convert the in-situ LST to pixel scale to address the spatial scale mismatch in LST validation. With this idea, the in-situ LST at daytime was converted to pixel scale associated with the SRI_TPR, and the corresponding MODIS and AATSR LST were validated at the 16 ground sites. Results for MODIS and AATSR LST show that the effect of spatial representativeness on the validation results over the sites is large, with mean biases between −1.95 K and 5.60 K and standard deviations between 0.07 K and 3.72 K. Since the TVM method does not rely on a specific satellite or land surface product, it is readily applied to other LST products (e.g. Sentinel-3 SLSTR LST, NOAA VIIRS LST) and surface parameters (e.g. surface longwave radiation).

Introduction

Land surface temperature (LST) is a vital indicator of the interaction between the Earth's surface and atmosphere, and it is an essential input parameter for associated models (Kalma et al., 2008; Liang et al., 2010; Tomlinson et al., 2011; Zhan et al., 2014). Compared with LST measured at ground sites, LST retrieved from satellite remote sensing has the advantage of spatial high-density coverage, continuity, and low cost. Therefore, it is efficient to use remote sensing to obtain LST on regional and global scales. Accordingly, the validation of satellite-retrieved LST is important: on the one hand, it is essential for developers of LST products, since it allows them to improve their retrieval algorithms and parameterization schemes (Ghent et al., 2017; Wan, 2014; Liang et al., 2021); on the other hand, satellite LST accuracy is important information for many applications, e.g. the target accuracy of satellite LST is 1 K for climate-related studies (Guillevic et al., 2018; WMO, 2020). Therefore, it is crucial to quantify the accuracy of a satellite LST product before its application.

Four LST validation methods are recommended by the Committee on Earth Observation Satellite (CEOS) working group on calibration and validation: temperature-based (T-based), radiance-based, cross-validation/intercomparison, and time-series analysis (Guillevic et al., 2018). The T-based method, which compares satellite LST with in-situ LST directly, is the most widely used method in satellite LST validation and the only method that qualifies as validation in the classical, metrological sense (Coll et al., 2012; Duan et al., 2019; Göttsche et al., 2016; Guillevic et al., 2014; Li et al., 2013; Martin et al., 2019; Ouyang et al., 2017; Simó et al., 2016; Zhang et al., 2019; Zhang et al., 2021). The T-based method makes the hypothesis that in-situ LST is representative on the spatial scale of the satellite. Theoretically, only ground sites with perfectly homogeneous surfaces (e.g. lakes, grassland, and desert) on the satellite pixel scale as well as on the in-situ measurement scale are appropriate for LST validation, that is, the in-situ LST must be spatially representative on satellite pixel scale. In practice, most ground sites are located over inhomogeneous landscapes and are not dedicated to satellite LST validation (e.g. FluxNET, ChinaFlux, and HiWATER), i.e., the in-situ LST does not represent ground truth on satellite pixel scale. Can we evaluate the spatial representativeness of such sites and use them for satellite LST validation? How should we evaluate their spatial representativeness and quantify its influence on the LST validation results?

Several methods have been used to quantify a ground site's spatial representativeness in the context of validating LST and other surface parameters. Generally, those methods can be divided into two categories. The first category is based on analyses of the surface heterogeneity within a pixel. A more heterogeneous surface yields lower spatial representativeness, e.g. spatial heterogeneity was analyzed with semi-variograms in albedo validation (Román et al., 2009; Wang et al., 2014); the site was classified into different heterogeneity levels by combining semi-variogram, coefficient of variation, and dominant land cover type in LAI (leaf area index) and LST validation (Xu et al., 2016; Yu et al., 2017); other indicators, e.g. standard deviation (Coll et al., 2012; Duan et al., 2019; Guillevic et al., 2014; Li et al., 2014), and coefficient of variation (Zhou et al., 2015), were also used to analyze spatial heterogeneity. Li et al. (2020) proposed a spatial and diurnal temperature cycle-constrained sampling strategy, which provides an effective way to validate LST over heterogeneous surfaces with dramatic diurnal LST changes. The second category analyses the relationship between various parameters on the in-situ and pixel scale, e.g. Huang et al. (2016) compared multi-point observed surface solar radiation with their averages to quantify the representativeness error; Hakuba et al. (2013) assessed the spatial sampling error of a site on the grid-scale with the corresponding collocated pixel and surrounding area averages; Martin et al. (2019) selected the representativeness period based on the annual variation of multi-year monthly averaged daytime differences between AATSR LST and in-situ LST; Liu et al. (2013) use the footprint model to assess the representativeness area of evapotranspiration measured at a ground site.

In general, in LST validation one cannot use those methods directly to evaluate the spatial representativeness. On the one hand, spatial heterogeneity characterizes the complexity of the surface, and spatial representativeness characterizes the degree of difference between ground measured and pixel observed parameters. However, satellite observations reflect the averaged value within a pixel: therefore, heterogeneity cannot fully characterize representativeness. On the other hand, compared to land surface parameters with low temporal variation, e.g. LAI, LST varies strongly on diurnal and intra-annual scales (Göttsche and Olesen, 2001; Zhan et al., 2014). In addition, due to differences in thermal inertia, LST variations in the ground radiometer's field-of-view (FOV) and on satellite pixel scale also differ in their responses to the variations of weather conditions. Therefore, it is reasonable to speculate that a ground site's spatial representativeness also varies on the temporal scale and, thus, the temporal evaluation of spatial representativeness is an important issue for validating satellite LST. Overall, the current methods do not provide a temporally continuous evaluation of spatial representativeness.

Therefore, this study proposes a temporal variation method (TVM) for evaluating the spatial representativeness of LST validation sites. The basic idea of the method is that a ground site's spatial representativeness is a function of time and can be described by a spatial representativeness indicator (SRI), which is defined as the LST difference between the ground radiometer's FOV and satellite pixel. Under ideal conditions, LST variation depends exclusively on the revolution of Earth around its axis and the Sun: the temporally continuous SRI introduced here describes this ideal scenario as a smooth curve. In reality, the SRI curve fluctuates around this ideal smooth curve as it is affected by changing surface characteristics (e.g. soil moisture, vegetation, land cover) and weather conditions. The variation of SRI can be mainly attributed to spatial variations of LST that follow typical intra-annual variations, e.g. due to vegetation changes, but are also affected significantly by weather conditions. Since the indicator describes the LST difference between the two scales, it can also be used as a bridge to convert the in-situ LST to pixel scale, and the satellite LST can be directly compared with the scale-converted LST to address the uncertainty caused by the spatial representativeness. The remainder of this article introduces and demonstrates applications of the TVM method and is structured into the sections Datasets, Methods, Results, Discussion, and Conclusions.