Data-driven model development in environmental geography - Methodological advancements and scientific applications

One key task in environmental geography is obtaining information of geographic features in space or in space and time. For this purpose, modelling strategies are needed that allow a delineation of spatio-temporal information based on limited field data. In this context, the nonlinearity and complexity of environmental systems require modelling strategies that allow handling arbitrary relationships and large sets of potential predictor variables. These requirements provoke a paradigm-shift from a parametric towards a non-parametric and data-driven model development which is strengthened by an increasing availability of geographic data. In that respect, machine learning algorithms have been proven to be an important tool to learn patterns in nonlinear and complex systems. While the large number of machine learning applications in scientific journals as well as recent software developments nowadays feign a simplicity of these methods, their application is not a trivial task. This holds especially true for geographic data as they have certain characteristics, especially spatial dependency, that make them stand out against the mass of "ordinary" data. However, this is widely ignored in geographic machine learning applications. This thesis assesses the potential and the sensitivity of machine learning in environmental geography. In this context, a number of machine learning applications in a broad spectrum of environmental geography have been published, providing a collection of comprehensive knowledge about machine learning in environmental geography. The individual contributions are incorporated in the major hypothesis that, only if characteristics of geospatial data are considered, data-driven modelling strategies lead to a reliable gain of information and to robust spatio-temporal model results. Beside this superior methodological focus, each application aims at providing new insights in its respective field of research. In this thesis, a number of relevant environmental monitoring products have been developed. The results emphasize that a high expertise of the machine learning methods as well as of the scientific field is crucial to advance the environmental geography. The thesis is the first to raise awareness of spatial or spatio-temporal over-fitting in geographic machine learning applications and the significant consequences to the outcome. To approach this problem, a new method for model development is provided that is adapted for geographic data and allows for improved model results. The thesis is finally an appeal to think beyond the "standard machine learning way" as it proves that applying standard machine learning concepts on geographic data results in considerable over-fitting and misinterpretation of the results. Only when characteristics of geographic data are considered, machine learning provides a powerful tool to provide scientifically valuable results in environmental geography.