An Investigation into the Use of Maximum Posterior Probability Estimates for Assessing the Accuracy of Large Maps

• Activity Number: 261
• Type: Topic Contributed
• Date/Time: Tuesday, August 10, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Brian M. Steele*+
• Companies: University of Montana
• Keywords: land cover map ; classification ; cross-validation ; prediction rule ; accuracy assessment

Abstract

Assessing the accuracy of land cover maps using conventional methods is often prohibitively expensive because of the difficulty of adequately sampling the range of geographic and environmental variation. Often, probability sampling is not feasible, and conventional accuracy estimators based on the proportion of correctly classified observations among a sample of map units with known land cover type are unreliable. An alternative approach using maximum posterior probability (MPP) estimators was investigated in this research. This talk introduces MPP-based estimators and reports on a comparative study of conventional and MPP estimators for a land cover mapping problem involving nine Landsat TM scenes. We simulated conventional and MPP-based accuracy estimators derived from post classification samples and training sample cross-validation. The results show that substantial reductions in the mean square error can be obtained from MPP-based estimators compared to conventional estimators. Additionally, MPP-based estimators produced reasonably good accuracy estimates even when biased designs were used to draw the training samples.