Assessing the Performance of Linear Feature Models

Abstract

A methodology for judging the performance of simulation models that involve linear features is presented taking as its test case a model that produces logging road networks. The approach combines existing methods of accuracy assessment, based on the so-called epsilon band, with the principle of statistical inference. This requires adapting neutral models from landscape ecology to generate a bootstrapped probability distribution reflecting the probabilistic characteristics of stochastic networks. The distribution is then used for inferential hypothesis testing, taking as its “null” hypothesis that a network has been produced randomly. The assessment methodology is used to evaluate the road network simulation model under six different parameterizations. Results are compared to findings obtained by the use of fuzzy similarity metrics. The paper argues that the epsilon band method provides confi-dence that model performance departs significantly from a randomized process, and that conclusions are free of at least one class of spurious results.