Remote sensing assisted multi-level crop acreage estimation — An application of small area estimation in Heilongjiang province

The planted acreage estimation for major crops by using remote sensing is typically combine the sample data from ground survey with information derived from image classification, and the common applied approaches are regression estimator by using linear model and calibration estimator by using confusion matrix. In general, the crop acreage estimation for provincial level in China only satisfied the precision for the target population and could not disaggregate to small areas, such as county and town level statistics. In recent years, the small area estimation is extended its application to agricultural statistics, so that by building a small area model it is applicable to estimate the sub-population or domains. This paper is adopted small area estimation approach to estimate crop acreage at county level in Heilongjiang province, China by combining image classification and ground survey data of year 2011. First, the scheme of sample selection for ground survey and method of remote sensing classification for crops in Heilongjiang province is introduced. Historical Landsat TM images in recent years are used to extract cropland to construct area frame, and then a stratified two stage sampling is adopted to select samples for ground survey. Some real Landsat TM images in early August of year 2011 as a key phenological period are used to discriminate major crops(corn, rice and soybean) by ML for entire province. Second, for the purpose of multi-level crop acreage estimation, we expect that the aggregation of county level estimates are equal to the estimate for entire province based on regression estimation. To meet this constraint, this paper is adopted a basic level small area model with fixed effects so that the sum of county estimates could be added up to the estimate of provincial total. The fitted model in the case of Heilongjiang province is in the form of a multi-response multiple regression. Third, the precision in terms of MSE of the estimates for corn, rice and soybean derived from small area model are illustrated, the coefficient of variation of estimates for these three major crops on county level average are relatively small and acceptable in practice. Finally, we discuss the model fitness of small area model, which are relevant to sample design for ground survey and form of model setting. In conclusion, it is efficient and robust to estimate sub-provincial crop acreage by adopting small area model if model itself is statistical sounding.