Application of Logistic-Regression and Classification Trees to Prediction of Soil Classes at a Regional Scale.
Inakwu Ominy A. Odeh and Nathan Odgers. The University of Sydney, Faculty of Agriculture, Food & Natural Resources, McMillan Building A05, Camperdown, Sydney, Australia
Soil classes, created by the so-called natural classification systems, are as a result of five factors of soil formation, namely climate, lithology, organisms, parent materials and time, jointly termed CLORPT. The CLORPT are generally well described by surrogate or ancillary variables which can be remotely observed. As the soil classes are factor variables generally coded in a nominal alphanumeric format, it is illogical to apply classical regression methods to predicting them from the ancillary variables. In order to take advantage of the increasingly available ancillary variables, new quantitative methods for rapidly predicting soil classes would enhance the quality of soil survey maps. Logistic regression and classification trees are such methods which are based on logical sequence of decisions. While logistic regression is designed specifically for situations in which dichotomous dependant variables are used as the predictants, classification trees are used to predict classes at unknown locations through a series of rules formulated from ancillary data at known locations. The aim of this study is to use a variety of ancillary information in association with the two methods logistic regression and classification trees, to predict the soil classes (created with the Australian Soil Classification (Isbell, 1996) at a regional scale in the Namoi catchment of New South Wales, Australia. The study region, located in northern NSW, Australia, has a diverse geomorphology and geology. Consequently, a wide range of soil classes exists. Different combinations of ancillary variables, generated either from remote sensing, DEM and existing lithologic and old soil maps, were used to spatially predict soil classes at the sub-order level of Australian Soil Classification. The performance of the two prediction methods were also compared in terms of proportions of misclassification and probabilistic uncertainty at each pixel. The results display soil class maps which show spatial continuity of the soil class (mapping) units that are well contiguous with each other. Both methods performed well when compared using the misclassification scores and probabilistic uncertainty, although logistic regression is slightly better than the classification tree model. The methods could be used to rapidly predict soil types for making soil classification maps in areas with similar patterns of landscape.