Geographical information modelling for land resource survey

The increasing popularity of geographical information systems (GIS) has at least three major implications for land resources survey. Firstly, GIS allows alternative and richer representation of spatial phenomena than is possible with the traditional paper map. Secondly, digital technology has improved the accessibility of ancillary data, such as digital elevation models and remotely sensed imagery, and the possibilities of incorporating these into target database production. Thirdly, owing to the greater distance between data producers and consumers there is a greater need for uncertainty analysis. However, partly due to disciplinary gaps, the introduction of GIS has not resulted in a thorough adjustment of traditional survey methods. Against this background, the overall objective of this study was to explore and demonstrate the utility of new concepts and tools within the context of pedological and agronomical land surveys. To this end, research was conducted on the interface between five fields of study: geographic information theory, land resource survey, remote sensing, statistics and fuzzy set theory. A demonstration site was chosen around the village of Alora in southern Spain.Fuzzy set theory provides a formalism to deal with classes that are partly indistinct as a result of vague class intensions. Fuzzy sets are characterised by membership functions that assign real numbers from the interval [0, 1] to elements, thereby indicating the grade of membership in that set. When fuzzy membership functions are used to classify attribute data linked to geometrical elements, presence of spatial dependence among these elements ensures that they form spatially contiguous regions. These can be interpreted as objects with indeterminate boundaries or fuzzy objects. Fuzzy set theory thus adds to the conventional conceptual data models that assume either discrete spatial objects or continuous fields.This thesis includes two case studies that demonstrate the use of the fuzzy set theory in the acquisition and querying of geographical information. The first study explored the use of fuzzy c -means clustering of attribute data derived from a digital elevation model to represent transition zones in a soil-landscape model. Validity evaluation of the resulting terrain descriptions was based on the coefficient of determination of regressing topsoil clay data on membership grades. Vaguely bounded regions were more closely related to the observed variation of clay content () than crisply bounded units as used in a conventional soil survey ().The second case study involved the use of the fuzzy set theory in querying uncertain geographical data. It explains differences between fuzziness and stochastic uncertainty on the basis of an example query concerning loss of forest and ease of access. Relationships between probabilities and fuzzy set memberships were established using a linguistic probability qualifier (high probability) and the expectation of a membership function defined on a stochastic travel time. Fuzzy query processing was compared with crisp processing. The fuzzy query response contained more information because, unlike the crisp response, it indicated the degree to which individual locations matched the vague selection criteria.In a land resource survey, data acquisition typically involves collecting a small sample of precisely measured primary data as well as a larger or even exhaustive sample of related secondary data. Soil surveyors often rely on soil-landscape relationships and image interpretation to enable efficient mapping of soil properties. Yet, they generally fail to communicate about the knowledge and methods employed in deriving map units and statements about their content.In this thesis, a methodological framework is formulated and demonstrated that takes advantage of GIS to interactively formalise soil-landscape knowledge using stepwise image interpretation and inductive learning of soil-landscape relationships. It examines topology to record potential part of links between hierarchically nested terrain objects corresponding to distinct soil formation regimes. These relationships can be applied in similar areas to facilitate image interpretation by restricting possible lower level objects. GIS visualisation tools can be used to create images (e.g. perspective views) illustrating the landscape configuration of interpreted terrain objects. The framework is expected to support different methods for analysing and describing soil variation in relation to a terrain description, including those requiring alternative conceptual data models. In this thesis, though, it is only demonstrated with the discrete object model.Satellite remote sensing has become an important tool in land cover mapping, providing an attractive supplement to relatively inefficient ground surveys. A common approach to extract land cover data from remotely sensed imagery is by probabilistic classification of multispectral data. Additional information can be incorporated into such classification, for example by translating it into Bayesian prior probabilities for each land cover type. This is particularly advantageous in the case of spectral overlap among target classes, i.e. when unequivocal class assignment based on spectral data alone is impossible.This thesis demonstrates a procedure to iteratively estimate regional prior class probabilities pertaining to areas resulting from image stratification. This method thus allows the incorporation of additional information into the classification process without the requirement of known prior class probabilities. The demonstration project involved Landsat TM imagery from 1984 and 1995. Image stratification was based on a geological map of the study area. Overall classification accuracy improved from 76% to 90% (1984) and from 64% to 69% (1995) when employing iteratively estimated prior probabilities.The fact that any landscape description is a model based on a limited sample of measured target attribute data implies that it is never completely certain. The presence of error or inaccuracy in the data contributes significantly to such uncertainty. Usually, the accuracy of land survey datasets is indicated using global indices (e.g. see above). Error modelling, on the other hand, allows an indication of the spatial distribution of possible map inaccuracies to be given. This study explored two approaches to error modelling, which are demonstrated within the context of land cover analysis using remotely sensed imagery.The first approach involves the use of local class probabilities conditional to the pixels' spectral data. These probabilities are intermediate results of probabilistic image classification and indicate the magnitude and distribution of classification uncertainty. A case study demonstrated the implication of such uncertainty on change detection by comparing independently classified images. A major shortcoming of this approach is that it implicitly assumes data in neighbouring pixels to be independent. Moreover, it does not make full use of available reference data as it ignores their spatial component. It does not consider data locations nor does it use spatial dependence models that can be derived from the reference data.The assumption of independent pixels obviously impedes proper assessment of spatial uncertainty, such as joint uncertainty about the land cover class at several pixels taken together. Therefore, the second approach was based on geostatistical methods, which exploit spatial dependence rather than ignoring it. It is demonstrated how the above conditional probabilities can be updated by conditioning on sampled reference data at their locations. Stochastic simulation was used to generate a set of 500 equally probable maps, from which uncertainties regarding the spatial extent of contiguous olive orchards could be inferred.Future challenges include studies on other quality aspects of land survey datasets. The present research was limited to uncertainty analysis, so that, for example, data precision and fitness for use were not addressed. Other potential extensions to this work concern full inclusion of the third spatial dimension and modelling of temporal aspects.