Predicting land use change with data-driven models

Abstract

One of the main tasks of data-driven modelling methods is to induce arepresentative model of underlying spatial - temporal processes using past dataand data mining and machine learning approach. As relatively new methods,known to be capable of solving complex nonlinear problems, data-driven methodsare insufficiently researched in the field of land use. The main objective of thisdissertation is to develop a methodology for predictive urban land use changemodels using data-driven approach together with evaluation of the performance ofdifferent data-driven methods, which in the stage of finding patterns of land usechanges use three different machine learning techniques: Decision Trees, NeuralNetworks and Support Vector Machines. The proposed methodology of data-drivenmethods was presented and special attention was paid to different datarepresentation, data sampling and the selection of attributes by four methods (χ2,Info Gain, Gain Ratio and Correlation-based Feature Subset) that best describe theprocess of land use change. Additionally, a sensitivity analysis of the SupportVector Machines -based models was performed with regards to attribute selectionand parameter changes. Development and evaluation of the methodology wasperformed using data on three Belgrade municipalities (Zemun, New Belgrade andSurčin), which are represented as 10×10 m grid cells in four different moments intime (2001, 2003, 2007 and 2010).The obtained results indicate that the proposed data-driven methodology providespredictive models which could be successfully used for creation of possiblescenarios of urban land use changes in the future. All three examined machinelearning techniques are suitable for modeling land use change. Accuracy andperformance of models can be improved using proposed balanced data sampling,including the information about neighbourhood and history in datarepresentations and relevant attribute selections. Additionally, using selectedsubset of attributes resulted in a simple model and with less possibility to beoverfitted with higher values of Support Vector Machines parameters.

Један од главних задатака моделирања метода вођених подацима (Data-driven methods) је проналажење репрезентативног модела испитивног просторно временског процеса, применом података из прошлости и Data Mining и Machine Learning приступа...