Submitted to: ASA-CSSA-SSSA Annual Meeting Abstracts
Publication Type: Abstract Only
Publication Acceptance Date: July 19, 2004
Publication Date: October 31, 2004
Citation: Pierson Jr, F.B., Pachepsky, Y.A., Weltz, M.A. 2004. Explorative analysis of the database on rangeland runoff and erosion experiments.[Abstract]. ASA-CSSA-SSSA Annual Meeting Abstracts, October 31 - November 4, 2004, Seattle, Washington. CD-ROM. Technical Abstract: Estimates of surface runoff and soil erosion are needed for both rangeland management and rangeland productivity evaluation. Such estimates are made using various models ranging from purely empirical to mechanistic. Our hypothesis was that the use of more easily obtainable data in machine-learning predictive tools can be a viable option for large scale estimates of soil erosion and runoff. Regression trees were selected as the machine-learning tool because of their ability to discover relationship structures specific for subsets of the whole database, transparency of results, and ability to select the most influential input variables. Data from 442 erosion events from 26 locations in the Western United States were used. Regression trees were developed (a) with input variables on weather and soil that could be obtained from public sources and (b) with both data from public sources and input variables that had to be measured on-site. The jackknife cross-validation was used to trim the trees. About 60 % of the variability in runoff could be explained using soil basic properties. Adding on-site data measured ground cover and surface roughness helped to explain about 70% of the variability. Predictions of the sediment yield were worse as only 40% of variation could be explained. Regression trees were easy to interpret, and grouping had a clear physical meaning in most cases.