Area estimation from a sample of satellite images: The impact of stratification on the clustering efficiency

Several projects dealing with land cover area estimation in large regions consider samples of sites to be analysed with high or very high resolution satellite images. This paper analyses the impact of stratification on the efficiency of sampling schemes of large-support units or clusters with a size between 5 km × 5 km and 30 km × 30 km. Cluster sampling schemes are compared with samples of unclustered points, both without and with stratification. The correlograms of land cover classes provide a useful tool to assess the sampling value of clusters in terms of variance; this sampling value is expressed as “equivalent number of points” of a cluster. We show that the “equivalent number of points” is generally higher for stratified cluster sampling than for non-stratified cluster sampling, whose values remain however moderate. When land cover data are acquired by photo-interpretation of tiles extracted from larger images, such as Landsat TM, a sampling plan based on a larger number of smaller clusters might be more efficient.