Comparing two area estimation techniques for established surveys using remotely sensed and ground data

Estimates of the area of forest using remotely sensed and ground data can be derived using two distinctly different paradigms. The first, which will be referred to as the "tessellated population paradigm," relies on the assumption that the area is divided into contiguous blocks that constitute an area frame. Inference is drawn using traditional finite population sampling methods. The other approach will be referred to as the point paradigm, which assumes that the sampling unit is a point on the ground. The estimators are derived assuming the number of ground plots in each remotely sensed category is random. The mechanisms that determine the variance of the estimators are different for both paradigms, with the variance of the point paradigm depending on the accuracy and size of each remotely sensed class, while the variance of the tessellated population paradigm depends predominantly on the fragmentation of the landscape in comparison with the size of the sampling unit. An analytical comparison of the estimators is not possible because of the difficulty associated with relating landscape fragmentation to the accuracy of the remotely sensed data. Thus, both estimators were tested by drawing samples from a series of artificial landscapes and satellite images using a plot configuration similar to the one used by the Forest Inventory and Analysis program of the United States. It is concluded that if the landscape is heavily fragmented and the accuracy is reasonably high, the point paradigm estimator is likely to have a smaller variance. The tessellated population estimators will usually have a smaller variance when the landscape has only a small degree of fragmentation. For the majority of the artificial populations studied, the estimator based on the tessellated population paradigm had the smallest variance. In situations where there would not be sufficient prior knowledge to determine which paradigm is superior or when the variance of both is expected to be about equal, the point paradigm is recommended because it is theoretically more tractable and easier to implement in the field.