Comparisons of Statistical Approaches for Modelling Land-Use Change

Land-use and land-cover change (LUCC) can have local-to-global environment impacts such as loss of biodiversity and climate change as well as social-economic impacts such as social inequality. Models that are built to analyze LUCC can help us understand the causes and effects of LUCC, which can provide support and evidence to land-use planning and land-use policies to eliminate or alleviate potential negative outcomes. A variety of modelling approaches have been developed and implemented to represent LUCC, in which statistical methods are often used in the classification of land use and land cover as well as to test hypotheses about the significance of potential drivers of LUCC. The utility of statistical models is found in the ease of their implementation and application as well as their ability to provide a general representation of LUCC, given a limited amount of time, resources, and data. Despite the use of many different statistical methods for modelling LUCC (e.g., linear models and logistic regression), comparison among more than two statistical methods is rare and an evaluation of the performance of a combination of different statistical methods with the same dataset has not been done before. The presented research fills this gap in LUCC modelling literature using four statistical methods, Markov chain, logistic regression, generalized additive models and survival analysis, to quantify their ability to represent LUCC. The selection of these methods is based on criteria: 1) the popularity of a method, 2) the difficulty level of implementation, and 3) the ability of accounting for different scenarios. Results from this comparison show that generalized additive models outperformed Markov chain, logistic regression and survival analysis in overall accuracy of LUCC but logistic regression performed the best for industrial land-use change, and survival analysis performed the best for low-density residential land-use change. The superiority of generalized additive models is due to its ability to model non-linear LUCC predictors, but there is no absolute favor in generalized additive models over other methods in terms of classification accuracies of specific LU changes and the run time. Markov chain is not competitive with the other three methods in most of the LU change cases but it retains the meaning as a null model (i.e., a model without any predictors) in our study