Land Cover Area Estimation with a Sample of Very High Resolution Images

This paper examines the potential accuracy of land cover area estimations based on samples of units with a size that is suitable for data collection through the analysis or interpretation of Very High Resolution (VHR) images. Many VHR sensors produce images with a size between 10x10 km and 20x20 km. The feasibility of area estimates based on a sample of images of 10x10 km is being assessed in the framework of the Geoland2 project (SATCHMO core mapping service). A simple assessment can be carried out using a square grid as sampling frame and a land cover map, such as CORINE Land Cover, as pseudo-truth, but the estimated efficiency is biased because the real land cover has a different spatial structure. To estimate the sampling efficiency we propose a method based on intra-cluster correlations computed from correlograms. The results can be expressed in terms of �equivalent number of points� of each site. We obtained poor efficiency results for area estimation of major land cover types in the European Union (EU) with simple random sampling; for example for the estimation of the total arable land area, each site of 10 x 10 km is equivalent to 3 unclustered points. However there are indications that the sampling scheme can be much more efficient for land cover change. Only sampling error is analyzed in the paper. This implicitly assumes that the observations in each sampled site have no error; the impact of non-sampling error still needs to be analysed. ; JRC.H.4 - Monitoring Agricultural Resources