A statistical protocol for describing global land -cover characterizations derived from remotely -sensed images.

Abstract

Statistical measures are evaluated and recommended for use in a protocol to describe land-cover characterizations (often with more than 100 classes), to identify classes needing further evaluation, and to support mapping the land-cover characterization classes into user-specified land-cover classes (often with fewer than 20 classes). The data sets studied are the Land-Cover Characterization database derived from AVHRR images (1992--1993) for the conterminous U.S., the USGS Land-Use Land-Cover database derived from aerial photography (1960s and 1970s), and the USFS coordinated ground-sampled Land-Cover Characterization Validation database (1993--1994). A set of overall, class-specific and class-pairs measures of association with distinct statistical properties is incorporated into the protocol. Three overall measures, Proportional Reduction in Error (PRE), Proportion of Explained Variance (PEV), and Cramer's V, quantify the relationship between land-cover characterizations and landcover classifications. Using a Monte Carlo simulation, at sample sizes less than 0.1% the population parameters are no longer contained within the interquartile range of the estimates. Standardized conditional variants of PEV and PRE are used to identify classes whose distributions of counts differ statistically from the marginal frequencies. The Chi-square, correlation coefficient, Yule's Q and deviance residuals quantify the associations between land-cover characterization and land-cover classification class-pairs. The Chi-square statistic identifies statistically independent class-pairs. The other three measures allow class-pairs to be ranked. The protocol was used to compare the three data sets for 2,523 sample plots (0.0327% of the population). The Land-Cover Characterization was slightly more closely associated with the USGS classification than was the USFS classification. The strongest associations for each Land-Cover Characterization class, in general, support the interpreted land-cover labels. A few Land-Cover Characterization classes with no, or only weak, associations with the interpreted class warrant further analysis. A comparison of the reference classifications identified potential sources of confusion. Techniques described in the remote-sensing literature for the analysis of contingency tables are designed to assess accuracy or agreement. The protocol and measures described in this dissertation provide a valuable set of tools for use by the resource scientist to efficiently and effectively handle the very large number of classes and relationships that exist in general global land-cover characterizations.