383 24 Analysing uncertainty B Minasny, TFA Bishop Introduction It is not sufficient for land resource surveys to produce just maps showing predictions of soil classes, attributes or various interpretations – the uncertainty of each prediction should also be shown. This additional requirement becomes indispensable when the data are used for simulation modelling (see Chapter 28). Uncertainty analysis provides answers to the following questions: ฀ how good is the prediction? ฀ which variables are the most sensitive? ฀ where we can spend the available resources to reduce the uncertainty of the results? Despite its importance, uncertainty has been seldom quantified in routine survey and land evaluation. Partly, this is because survey agencies are unaware of what to do in a fairly complex field. This chapter introduces the topic and guides agencies in the steps they can take. Analysis of uncertainty in laboratory measurement is well documented (e.g. Allmaras and Kempthorne 2002). However, there are differences between the uncertainties encountered during assessment of land resources and laboratory data, the main difference being the source of error. In the laboratory, the measure of uncertainty is obtained from replicated measurements under controlled conditions, and the variation is attributed to random error. In models, identi- cal outputs are expected when the same inputs are fed into a deterministic model. The uncer- tainty of the output can be quantified by treating the inputs as random variables. Thus, the outputs of the model will be random because they are transformations of random inputs (McKay 1988). This chapter deals with uncertainty in models, whether they be pedotransfer functions, statistical models for spatial prediction, environmental predictors or simulation models. Several terms such as error , deviation, uncertainty , sensitivity , risk and reliability have been used interchangeably and, it has to be said, carelessly. Each has a specific meaning and to prevent further confusion and misunderstanding they are defined formally (Table 24.1). McBratney (1992) recognised three types of uncertainty in soil information: stochastic, deter- ministic and semantic. Stochastic uncertainty has been the focus in statistics and probability theory, deterministic in chaos theory and semantic in fuzzy theory. This chapter will only deal with the first type of uncertainty, stochastic. Uncertainty is a major topic in the spatial information sciences, and there are good mono- graphs on it. Heuvelink (1998) provides a theoretical description and supplies applications of uncertainty analysis in geographical information systems (GISs). Zhang and Goodchild (2002) discuss the theoretical aspects of uncertainties in geographical information and how to deal with various types of error in modelling them. Foody and Atkinson (2002) review the theory and practical applications of uncertainty analysis in remote sensing and GISs. The same reference