Illustrating Approaches to Uncertainty Estimation for Map-Based Estimation Problems

Mark A. Hatfield and Ronald E. McRoberts

USDA Forest Service North Central Research Station

Traditionally natural resource managers and users of natural resource data have asked the question “How much?” and have received sample-based estimates of totals or means for large areas such as counties, regions, or states.  Increasingly, however, the same managers and users are now asking the additional question “Where?” and are expecting spatially explicit answers in the form of maps.  Fortunately, the recent development and widespread availability of natural resource databases, moderate resolution satellite imagery, image classification techniques, statistical software, and geographic information systems (GIS) have facilitated construction of the required maps.  Unfortunately, the relevance of the uncertainty associated with these maps has not been sufficiently emphasized, techniques for estimating it are either not known or are not taught, and systems for portraying and analyzing it are generally not available.   When scientists who estimate and analyze uncertainty decry the lack of map uncertainty estimates, managers and users often respond that it doesn’t matter because the map itself is still useful and, besides, there are no other alternatives.  Thus, it behooves uncertainty scientists to articulate the consequences of ignoring uncertainty and to describe and demonstrate to user communities the techniques for estimating and analyzing it.

This presentation illustrates the use of analytical, Monte Carlo, and sensitivity techniques for assessing uncertainty in three categories:  (1) individual maps, (2) maps constructed by combining underlying maps using GIS techniques, and (3)  maps constructed using models that accept input from underlying maps.  In the first category, the emphasis is on estimating the precision of predictions for individual mapping units using analytical techniques associated with the particular map construction technique or Monte Carlo techniques when the analytical techniques are intractable or unavailable.  In the second category, the emphasis is on estimating the precision of mapping unit predictions resulting from the GIS intersection of underlying maps.  In the third category, the emphasis is on estimating the precision of mapping unit predictions when underlying map information is integrated via an application-based model.  Examples and illustrations are drawn from map-based analyses of the potential threat to forest land in Michigan, USA, from the Emerald Ash Borer.

Statistical Methods Session - Wednesday Afternoon

corresponding author:

Mark Hatfield
Forest Inventory and Analysis
North Central Research Station
1992 Folwell Avenue
St. Paul, MN
651-649-5169
mahatfield@fs.fed.us