Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Doctorat → Belgique → Digital soil mapping in a selected arid landscape in south-eastern Iran

Universiteit Gent (2012)

Digital soil mapping in a selected arid landscape in south-eastern Iran


Titre : Digital soil mapping in a selected arid landscape in south-eastern Iran


Université de soutenance : Universiteit Gent

Grade : PhD Physical Land Resources 2012

Mapping of soil types and soil properties is expensive, time consuming and may be subjective. So-called Digital Soil Mapping (DSM) methods identify statistical relations between (existing) fullcover landscape information and local soil information, and thus provide an objective and reproducible alternative to traditional survey and mapping. As DSM-methods allow quantification of the quality of the resulting map, costs, time and quality can be managed. In this thesis, a number of DSM methods for mapping of soil types were implemented, tested and compared with the purpose to improve mapping at scale 1:50,000 in an arid landscape in Iran. The case study area is located in the Zarand region in south-eastern Iran, and is characterized by soils rich in soluble salts, gypsum and calcite, variable relief of limestone, dolomite and shale mountains, bajadas and playas. Existing maps of geology and geomorphology were combined to obtain sampling strata. In total, 126 soil profiles were sampled, analysed and classified using the USDA Soil Taxonomy. The common soil type definition at 1:50,000 scale in Iran is based on the USDA Soil Taxonomic level “great group”. In the Soil Taxonomy, the classification at the great group level is based on the occurrence of diagnostic surface and subsurface horizons. DSM-methods could therefore aim at mapping the occurrence of these diagnostic horizons followed by a pixel-wise classification into great group (“indirect method”). Alternatively, DSM-methods could also aim at directly mapping the great group (“direct method”). In a first case study, the indirect method was implemented by binary logistic regression, and the direct method by multinomial logistic regression. Validation results showed that the direct approach gave slightly better maps (66% purity) than the indirect approach (59%), and that this difference is probably related to the decision tree implemented to mimic the Soil Taxonomy classification. In either case, geomorphology maps proved to be the strongest ancillary variable in mapping. In a second case study, the method of boosted regression trees was applied, both in an indirect and direct context. Results again indicate a better performance of the direct approach (58% purity) relative to the indirect approach (49%). Both in the first and the second case study, the geomorphology maps provide useful ancillary variables, with the worst performance in the more developed soils (with a calcic horizon). Probable explanation is that these soils are older than the current landscape, while soils with accumulations of salt and gypsum are younger and relate better to the current landscape. In a third case study the question was raised and answered on how to determine the optimal taxonomic level for mapping at a chosen mapping scale and sampling density. The optimal level results in a map of good quality which also displays detailed spatial patterns. Artificial neural networks were trained to map soil types at the order, suborder, great group and subgroup level. The purity P was calculated to represent map quality. The Shannon’s entropy index S was calculated to represent the diversity of the produced maps. A combined index was defined as P*S and calculated at the four taxonomic levels. Results showed that the optimal taxonomic level, for the sample size, its spatial configuration and the map scale, was the great group level.


Version intégrale

Page publiée le 28 octobre 2015, mise à jour le 28 novembre 2018