The agricultural land cover in a 263 km2 irrigation district was classified utilizing two Landsat 5 TM scenes. Manual and automatic selection of training areas for the classification of two single subscenes and a combined multitemporal subscene result in several differently classified images. The extent of each land cover class was first estimated by area frame sampling and further expansion of the ground data to the entire irrigation area. The regression of the sampled surface on the corresponding pixels in the classified images was used to improve the regression estimates of the areas of different land cover classes. To ascertain if there is any statistical difference between the relative efficiencies (RE) of the regression estimator using each one of the classifications, and being RE= 1/ (1 -r2) a test of equality between correlation coefficients was applied. When the correlation coefficients were significantly different the most precise estimation was indicated by the highest RE. Manual multitemporal classifications provided the most precise results, with the exception of the single image spring classification of rice.