The cost of accuracy in crop area estimation

Abstract

Accuracy and cost of direct crop area measurement are compared with those of farmers’ estimates after visual inspection, in this methodological study conducted in southern Mali. The observational error, the difference between the area measured and the area estimated, was first studied at the plot level. Average observational error or bias was −11% of the average area measured, indicating an average underestimation of plots by 11%. This observational error is strongly related to plot size, with smaller plots being overestimated and larger plots underestimated, in an approximately negative linear relationship. The observational error is also smaller for cotton fields than for cereals. The analysis was repeated at the farm level, where the bias in estimating the total area per farm was −8%. At this level, total error or accuracy was calculated by the relative total error (RTE), the square root of the mean square error, divided by the mean. The farmers’ estimate was found to be less accurate (RTE = 9.4% of the mean) than physical measurement (RTE = 6.6%), but at a cost of only $370 as compared with $2328 (for a sample of 96 farms in 11 villages). The coefficient of variation (CV) of most surface variables was found to lie between 60% and 100%, and their relative bias (average observational error divided by the mean area) between 2% and 10%. For crop area per farm, the physical measurement of plots resulted in a gain of accuracy of 2–4%, as compared with the farmers’ estimate after visual inspection. A general model was developed in which these calculated parameters are used to predict the accuracy in future surveys and to compare the accuracy with the survey’s cost. It is shown how the survey design can be optimized based on acceptable error, sample size and cost for each measurement technique. Simulations demonstrate that the total error for biased estimators, even for variables with small CVs, hardly decreases above sample sizes of 100–150 farmers.