Spatial representativeness estimation of station observation in validation of LAI products: A case study with CERN insitu data

Abstract

The continuously observed Leaf Area Index (LAI) dataset from the ground station network is an important data source for the validation of remote sensing products. However, direct comparison introduces errors to the validation results if the station-observed LAI cannot represent the pixel because of the scale mismatch between station and pixel observations. This study aims to present an approach to evaluate the spatial representativeness of station LAI observations. This proposed approach will be used to validate LAI products. Three evaluation indicators, including the Dominant Vegetation Type Percent (DVTP), the Relative Spatial Sampling Error (RSSE), and the Coefficient of Sill (CS), were established to determine the different levels of spatial representativeness for station observations. DVTP calculated by land-cover maps can evaluate the vegetation-type representativeness in the product pixel. RSSE and CS were calculated from LAI/normalized difference vegetation index high-resolution reference maps, which were used to describe the degree of representativeness for vegetation density in the pixel. The approach was applied to 25 stations from the Chinese Ecosystem Research Network (CERN), which includes croplands and forest in China. The threshold was set as 60% for DVTP and 20% for both RSSE and CS to determine the level of spatial representativenessat different observed dates and stations. Then, the variation between seasonal and inter-annual spatial representativeness was evaluated by comparing 2010 and 2011.Finally, the results of Moderate-Resolution Imaging Spectroradiometer (MODIS) LAI product validation before and after grading station observations were compared to demonstrate the importance of spatial representativeness evaluation. The spatial representativeness level of station-observed LAI data with different dates was first determined on the basis of grading criterion. The seasonal level varied at different growth stages of vegetation, whereas the inter-annual level was consistent because the structure and pattern of vegetation were stable for the adjacent years. The root mean squared error between the MOD15A2 and observed LAI with the good spatial representativeness reduced from 1.67 to 1.16 compared with that of all observed LAI data. The combination of DVTP, RSSE, and CS is an effective approach to assess the spatial representativeness of station-observed LAI dataset. Moreover, the uncertainty of MOD15A2 validation significantly differsat different levels of spatial representativeness. Thus, the level of station-observed LAI data at the product pixel scale should be determined, and high-level LAI observation should be chosen to reduce the error for validating LAI products. However, the station LAI observations that can represent the product pixel were not sufficient because of the influence of spatial heterogeneity. For example, the percentages of levels 0 to 3 for CERN station-observed LAI dataset were70.9% and 8.9% in 2010, respectively. Therefore, further studies should focus on increasing the number of validation dataset by two ways: collecting station LAI observations over many years at the global scale for various biomes and rectifying scale errors between station and pixel observations to fully utilize station-observed LAI data. Consequently, the LAI products can be comprehensively and reliably validated.