Abstract

Sampled data are used to calibrate and determine the accuracy of both remotely sensed data and the products of remotely sensed data. This paper discusses the problems of achieving a balance between the size and the error of these samples, not at the level of the scene but at the level of the sites within that scene. Using an area of limestone grassland as an illustrative example, the sample size required to characterize a wide range of grassland fields was determined. With a maximum of 5% error at the 95% confidence level the minimum sample size per field was noted to vary between 1–58 for ground radiometric and airborne multispectral scanner measurements and 142–293 for green leaf area index measurements. The collection of such large sample sizes is unusual in remote sensing. This review concludes that there is a need for an increased awareness of the magnitude of the sampling error and an attempt should be made to use the known spatial autocorrelation in the data to reduce error for a given sample size.