Utilizing soil polypedons to improve model performance for digital soil mapping

Authors: Levi, Matthew; Bestelmeyer, Brandon; WHITE II, DAVID - Natural Resources Conservation Service (NRCS, USDA)

Submitted to: Agronomy Society of America, Crop Science Society of America, Soil Science Society of America Meeting
Publication Type: Abstract Only
Publication Acceptance Date: 7/21/2016
Publication Date: 11/6/2016
Citation: Levi, M.R., Bestelmeyer, B.T., White Ii, D.A. 2016. Utilizing soil polypedons to improve model performance for digital soil mapping. Agronomy Society of America, Crop Science Society of America [abstract], Soil Science Society of America Meeting. November 6-9, 2016, Phoenix, Arizona.

Interpretive Summary:

Technical Abstract: Most digital soil mapping approaches that use point data to develop relationships with covariate data intersect sample locations with one raster pixel regardless of pixel size. Resulting models are subject to spurious values in covariate data which may limit model performance. An alternative approach is to aggregate covariate data for discernable soil bodies of similar soils, called polypedons, as training data for sample locations. Our objective was to explore the utility of different aggregation schemes of covariate data for soil mapping efforts and disaggregation of existing soil map units. Two study areas in semiarid rangelands of southeastern Arizona and southwestern New Mexico were the focus of this work. We present three examples that represent recent developments in the use of polypedon concepts for spatial predictions of soil components: 1) the effect of aggregating covariate data with a circular buffer around training data, 2) hand digitizing of polypedons to bolster sample size and overcome limitations of class imbalance for model development within a soil survey update, and 3) object-based image analysis for a data-driven approach of model development with polypedon units. A comparison of eight machine learning algorithms showed that a circular buffer size of 150 meters around training data had the best model performance (Kappa = 0.43) in the Arizona watershed. Radial support vector machine and random forest models performed best. In the New Mexico example, random forest models had with a Kappa of 0.16 for sampled points that improved to 0.95 with an additional 50 polypedon points for each class. Image segmentation of covariate data in the Arizona watershed using eCognition generated polypedons for model development without hand digitizing. Using polypedons as training data will improve spatial predictions of soil properties by maximizing information provided by covariate data and linking concepts of traditional soil survey and digital soil mapping.