地面站点叶面积指数观测的空间代表性评价——以CERN站网观测为例

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

在叶面积指数LAI(Leaf Area Index)产品真实性检验中,地面站点的多时相连续观测LAI数据是重要的验证数据来源。当站点观测范围与产品像元尺度不一致时,站点观测LAI直接用于产品验证可能为验证结果带来误差。因此,在验证之前需要分析站点观测对像元尺度的空间代表性,选择空间代表性好的观测来验证产品,从而减小尺度效应带来的验证误差。以往的研究只是简单的定性说明研究区域,并直接用站点测量数据对产品进行验证,缺少一套系统的站点观测在产品像元尺度内空间代表性评价的方法体系。本文提出了站点LAI观测的空间代表性评价方法,建立了评价指标DVTP(Dominant Vegetation Type Percent)、RSSE(Relative Spatial Sampling Error)和CS(Coefficient of Sill),构建了空间代表性评价分级体系。以中国生态系统研究网络CERN(Chinese Ecosystem Research Network)农田站和森林站LAI观测为例,对站点观测在1 km产品像元尺度内空间代表性进行评价,并分析评价前后站点观测对MODIS LAI产品验证精度的影响。结果显示,本文提出的方法能够有效地对不同站点LAI观测在产品像元尺度内空间代表性进行质量分级,且年际间的站点观测空间代表性较为一致。评价方法能够去掉在特定产品像元尺度下空间代表性不好的观测数据,一定程度上提高验证数据集对产品验证精度的可靠性。

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 representa- tiveness 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 Speetroradiometer （MODIS） LAI product validation before and after grading station observations were compared to demonstrate the importance of spatial representativeness evaluation. The spatial repr- esentativeness 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 data- set 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 com- prehensively and reliably validated.