Uncertainty : types and applications in spatial predictive models

Uncertainty is one of the most essential and fundamental issues that requires full attention in almost all spatial models and applications. Evidently, the quality of uncertainty modelling plays a critical role in resultant outcomes of geographical models and applications with an inevitable effect on decision- making processes. Therefore, up to now, uncertainty assessment and modelling has gained extensive attention in the field of spatial sciences. Considering the growing importance of this issue, this thesis investigates uncertainty modelling that applies in spatial science along with practical strategies to deal with them. To this end, three definitions of uncertainty are adopted, including Type A in which the uncertainties are derived from series of repeated observations, Type B , such as ambiguity and/or vagueness, with the uncertainties calculated by means other than the statistical analysis of series of observations (i.e. general knowledge of the behaviour and properties of phenomena) and Type C that is uncertainties in form of randomness. Proposed strategies to deal with each type of uncertainty is also exemplified in this thesis. In terms of Type A uncertainty, in this thesis, entropy is the key term representing uncertainty. Considering the fact that repeated observations are more often aimed at implementation of predictive spatial models, three examples are described (1) assessing uncertainty of several machine-learning classification algorithms that are repeatedly applied in implementation of spatial predictive models, (2) improving performance of a classification scheme using uncertainty assessment and (3) applying an optimised algorithm using uncertainty assessment in land-use change simulation as a popular example. To this end, in Chapter 2, a strategy is proposed to evaluate the uncertainty of two machine-learning algorithms (i.e. random forest vs deep neural network) that are applied for land-use classification. These two algorithms are highly popular in implementation of predictive spatial models. In Chapter 3, a strategy to improve the performance of predictive algorithms using uncertainty models is proposed in the context of image classification. A comprehensive practical illustration is then applied for improving the quality of a land-use change model in Chapter 4. With reference to Type B uncertainty, two different examples are implemented for drought sensitivity and landslide susceptibility mapping, where fuzzy If-Then rules and fuzzy sets are used to improve the quality of spatial models in Chapters 5 and 6, respectively. Finally, Chapter 7 is focused on applying randomness for modelling Type C uncertainty that is frequently encountered in spatial models, especially land-use change models. Although randomness is an inevitable component of land-use change models, it is sometimes overlooked. In this chapter, randomness is applied as a primary component of land-use change models, but it is also applied to achieve the optimised neighbourhood setting as a key requirement of the model calibration phase. This thesis, therefore, is a comprehensive illustration of different types of uncertainty in a range of important spatial models. Here, the thesis case studies are selected, so that they can be either applied to model the same spatial phenomenon within different case studies or can be generalised to new applications in the spatial domain