Evaluation of Spatio-Temporal Variogram Models for Mapping Xco2Using Satellite Observations: A Case Study in China

Abstract

Greenhouse Gases Observing Satellite (GOSAT), which measures column-averaged carbon dioxide dry air mole fractions (Xco <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> ) from space, provides new data sources to improve our understanding of carbon cycle. The available GOSAT data, however, have many gaps and are irregularly positioned, which make it difficult to directly interpret their scientific significance without further data analysis. Spatio-temporal geostatistical prediction approach can be used to fill the gaps for global and regional Xco <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> mapping. It is important to choose a suitable spatio-temporal variogram model since modeling spatio-temporal correlation structure using variogram model is a critical step in the geostatistical prediction. In this study, three different flexible spatio-temporal variogram models, including the product-sum model, Cressie-Huang model, and Gneiting model, are used to model the spatio-temporal correlation structure of Xco <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> over China, using the Atmospheric CO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> Observations from Space retrievals of the GOSAT (ACOS-GOSAT) Xco <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> (v3.3) data products. The three models are compared and evaluated using the weighted mean square errors (WMSE) indicating the fitness between the empirical variogram surface and the theoretical variogram model, cross-validation for quantifying prediction accuracies, and the performance of the three models when used to fill the spatial gaps and generate Xco <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> maps in 3-day temporal interval. The results indicate that 1) the model fitness of the commonly used product-sum model is slightly better than Cressie-Huang model and Gneiting model as indicated from WMSE, and 2) all the three models present similar summary statistics in cross-validation, all with a significantly high correlation coefficient of 0.92, and about 83% of prediction error within 2 ppm and about 53% within 1 ppm, and (3) differences between the mapping results using the three models are generally less than 0.1 ppm, and no significant differences can be identified. As a conclusion from the above results, all the three variogram models can precisely catch the empirical characteristics of the spatio-temporal correlation structure of Xco <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> over China, and the precision and effectiveness of predicting and mapping Xco <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> using the three models are almost the same.