FITTING SOIL PROPERTY SPATIAL DISTRIBUTIONMODELS IN THE MOJAVE DESERT FOR DIGITAL SOILMAPPING

摘要

We developed models from soil profile descriptions and GIS landscape analysis to es-timate the spatial distribution of soil properties to assist soil scientists with soil-landscape information. Soil profile descriptions were obtained within soil survey projects in the Mojave Desert of southeastern California, USA. Sites were located on broad alluvial fans. Soil development varied from young soils with little or no soil development to well-developed soils on older alluvial fan remnants. We obtained a set of profile descriptions (n = 264) from the traditional ongoing field-work. The location of these sample points was determined by soil scientist judgment of com-binations of soil-forming factors. The project area is sparsely vegetated and access is relatively unimpaired in most areas. We feel that these purposive samples represent the range of the soil-forming factors and that sample location bias will be low. Although this bias is not measurable. We wanted to see if we could make use of these data. We developed models from these data and evaluated the performance of the models using the measured values at randomly located sites not used to fit the models. The models estimated selected soil characteristics continuously in a 30-m raster over the project area. The response varia-bles that we modelled were soil genetic features that are used as diagnostic properties in USDA Soil Taxonomy, for example particle-size class, presence or absence of argillic horizon. Soil profiles and landscape features were described at 97 randomly located field sites within a portion of the active soil survey project. Explanatory variable information was developed for each of these sites through GIS extraction from digital elevation model data, landform derivatives, band-ratio satellite images and geomorphologic data. Model estimates for particle-size class were correct or within one class of the correct class for 73% of sample points. Models for depth to soil features had a range of per-formance. The best fitting model estimated the depth to secondary carbonates within 20 cm of actual depths for 71% of sample points, which contained carbonates. The model for depth to calcic horizon performed less well, while the model for depth to argillic was slightly less reliable. The model for presence or absence of calcic horizon was the most reliable logistic model. Soils on millions of hectares will be mapped in this general area in the future and we are trying to increase mapping efficiency and depth of understanding of soil-landscape