Predictive ability of logistic regression, auto-logistic regression and neural network models in empirical land-use change modeling – a case study

Abstract

The objective of this study is to compare the abilities of logistic, auto-logistic and artificial neural network (ANN) models for quantifying the relationships between land uses and their drivers. In addition, the application of the results obtained by the three techniques is tested in a dynamic land-use change model (CLUE-s) for the Paochiao watershed region in Taiwan. Relative operating characteristic curves (ROCs), kappa statistics, multiple resolution validation and landscape metrics were used to assess the ability of the three techniques in estimating the relationship between driving factors and land use and its subsequent application in land-use change models. The validation results illustrate that for this case study ANNs constitute a powerful alternative for the use of logistic regression in empirical modeling of spatial land-use change processes. ANNs provide in this case a better fit between driving factors and land-use pattern. In addition, auto-logistic regression performs better than logistic regression and nearly as well as ANNs. Auto-logistic regression and ANNs are considered especially useful when the performance of more conventional models is not satisfactory or the underlying data relationships are unknown. The results indicate that an evaluation of alternative techniques to specify relationships between driving factors and land use can improve the performance of land-use change models.