JIA-2018-09

2097 WANG Di et al. Journal of Integrative Agriculture 2018, 17(9): 2096–2106 crop production, formulating agricultural policies and ensuring national food security (Reynolds et al. 2000; Tao et al. 2005; Song et al. 2017). For a long time, crop acreage data have been routinely surveyed and updated by China’s statistical department mainly using the traditional multilevel list sampling method, in which sample counties are drawn from the provinces, and then sample villages are drawn from sampled counties, and finally sample households are drawn from the sampled villages (NBSC 2002). However, the operational problems facing the list sampling survey include: slow update of the sampling frame, insufficient use of spatial information technology during the survey, and human influence on the survey results. Consequently, the accuracy and timeliness of the crop acreage survey remain poor. With the development of earth observation technology, satellite-based remotely sensed data have the unique advantage of providing continuous spatial and temporal information on crop cultivation and growth at various regional scales, owing to its wide real-time coverage. Therefore, it has often been combined with traditional sampling methods, to estimate the crop acreage over a large region. Compared with the traditional list sampling survey, spatial sampling transforms the investigation object from a farmer household into a cultivated land plot. It can, therefore, achieve more accurate crop acreage information. Moreover, using satellite-based remotely sensed data increases the timeliness of crop area surveys in conjunction with spatial sampling. The use of remotely sensed data in sampling surveys for crop acreage estimation began in the early 1970s. The Large Area Crop Inventory Experiment (LACIE), conducted jointly, in 1974, by the United States Department of Agriculture (USDA) and the National Aeronautics and Space Administration (NASA) is a typical example. Satellite images were used to derive a thematic map of wheat distribution, and this was the basis for a stratified sampling scheme (Macdonald and Hall 1980). More recent projects include the Agriculture and Resource Inventory Surveys through Aerospace Remote Sensing (AGRISARS) (Benedetti et al. 2010) and the Monitoring Agriculture with Remote Sensing (MARS) Project (now named CROP4CAST) sponsored by the European Union. The latter used stratified sampling to monitor and estimate the acreage of 17 crops. In that project, satellite images were employed to formulate a stratification scheme and to measure crop acreages within the sampled units (Gallego 2004; Carfagna and Gallego 2005). Previous studies on crop acreage estimation using the spatial sampling method have focused mainly on how to improve the survey accuracy and timeliness, and decrease the survey cost and workload, by combining traditional sampling methods with remotely sensed data. In these studies, simple random sampling, systematic sampling, stratified sampling and multistage sampling methods have often been used, and the crops involved included wheat, rice, maize and cotton. The deficiencies of the list sampling survey can thus be effectively addressed by combining traditional sampling methods and remotely sensed data. However, traditional sampling is based on classical statistical theory, which is appropriate when researching changes in purely random variables. These methods treat the sampling units, when used to estimate population parameters, as independent. In practice, the influence of natural conditions (climate, soil type, topography, and landforms), socioeconomic factors, and crop distributions mean the sampling units, at the regional scale, are not independent and exhibit a degree of spatial autocorrelation. Traditional sampling methods cannot address this spatial autocorrelation. Many previous studies have indicated that the spatial autocorrelation of such key variables should be considered in spatial sampling (Overmars et al. 2003; Gertner et al. 2007; Buarque et al. 2010; Le Rest et al. 2013; Holmberg and Lundevaller 2015). Neglect of spatial autocorrelation may lead to overestimation of variance and require sample size, and possibly a false conclusion (Lichstein et al. 2002; Betts et al. 2006; Frutos et al. 2007; Hoeting 2009; Kulkarni and Mohanty 2012; Melecky 2015). Spatial autocorrelation refers to correlations that exist between the observed values of a single attribute/variable at different locations across an area, and is a measure of the degree of spatial clustering of the attribute (Cliff and Ord 1981; Griffith 1988). Spatial autocorrelation is common within geographic data and is found in diverse spatial variables, contexts and structures (Legendre and Legendre 1998; Keitt et al. 2002). Since Tobler (1970) put forward his First Law of Geography, studies on spatial autocorrelation among geographical attributes have been undertaken by many researchers (Cliff and Ord 1981; Anselin 1988; Goodchild et al. 1993; Fisher et al. 1996; Haining 2003). The level of spatial autocorrelation can be estimated using a global or a local indicator. A global indicator is used to detect spatial clustering across a whole study territory, with the area’s spatial autocorrelation represented by a single value. To date, Moran’s I remains the most commonly used statistical indicator for global spatial autocorrelation (Holmberg and Lundevaller 2015; Melecky 2015). Using Moran’s I as the statistical indicator, the referenced studies have demonstrated the existence of spatial autocorrelation within various geographical variables, explored the characteristics and variations in spatial autocorrelation, and analyzed its influence on required sample size, population extrapolation and estimated error in a sampling survey. Although studies on spatial autocorrelation among sampling units have been conducted for more than a decade, these studies have focused mainly on animal conservation