Simulating land use land cover change using data mining and machine learning algorithms

Amin Tayyebi, Purdue University

Abstract

The objectives of this dissertation are to: (1) review the breadth and depth of land use land cover (LUCC) issues that are being addressed by the land change science community by discussing how an existing model, Purdue's Land Transformation Model (LTM), has been used to better understand these very important issues; (2) summarize the current state-of-the-art in LUCC modeling in an attempt to provide a context for the advances in LUCC modeling presented here; (3) use a variety of statistical, data mining and machine learning algorithms to model single LUCC transitions in diverse regions of the world (e.g. United States and Africa) in order to determine which tools are most effective in modeling common LUCC patterns that are nonlinear; (4) develop new techniques for modeling multiple class (MC) transitions at the same time using existing LUCC models as these models are rare and in great demand; (5) reconfigure the existing LTM for urban growth boundary (UGB) simulation because UGB modeling has been ignored by the LUCC modeling community, and (6) compare two rule based models for urban growth boundary simulation for use in UGB land use planning. The review of LTM applications during the last decade indicates that a model like the LTM has addressed a majority of land change science issues although it has not explicitly been used to study terrestrial biodiversity issues. The review of the existing LUCC models indicates that there is no unique typology to differentiate between LUCC model structures and no models exist for UGB. Simulations designed to compare multiple models show that ANN-based LTM results are similar to Multivariate Adaptive Regression Spline (MARS)-based models and both ANN and MARS-based models outperform Classification and Regression Tree (CART)-based models for modeling single LULC transition; however, for modeling MC, an ANN-based LTM-MC is similar in goodness of fit to CART and both models outperform MARS in different regions of the world. In simulations across three regions (two in United States and one in Africa), the LTM had better goodness of fit measures while the outcome of CART and MARS were more interpretable and understandable than the ANN-based LTM. Modeling MC LUCC require the examination of several class separation rules and is thus more complicated than single LULC transition modeling; more research is clearly needed in this area. One of the greatest challenges identified with MC modeling is evaluating error distributions and map accuracies for multiple classes. A modified ANN-based LTM and a simple rule based UGBM outperformed a null model in all cardinal directions. For UGBM model to be useful for planning, other factors need to be considered including a separate routine that would determine urban quantity over time.

Degree

Ph.D.

Advisors

Pijanowski, Purdue University.

Subject Area

Civil engineering|Computer science|Urban planning

COinS