Digital soil mapping techniques can reduce the costs associated with improving soil information at the farm scale. Many techniques exist for generating a digital soil mapping. In this study, we tested the use of a random forest (RF) and two model-based soil-sampling scheme, conditioned with Latin hypercube (cLHS) and fuzzy c-means sampling (FCMS) to predict soil properties for a 68-ha agricultural field. The predictor set included the use of inexpensive auxiliary information such as digital elevation models, yield maps, surface reflectance, and apparent soil electrical conductivity measurements. Our results support the underlying assumption that both cLHS and FCMS capture adequately the full predictor distribution in the Pampas conditions. Despite the complex local variation of soils and the instability of normalized difference vegetation index and yield maps, the RF models predicted soil properties at farm scale. The RF model with cLHS and FCMS explained about 65% (R2 = 0.648) and 57% (R2 = 0.571) of the variability of soil organic matter and about 71% (R2 = 0.714) and 68% (R2 = 0.686) of the variability of clay, respectively. Our results provide a significant improvement of information related to soil organic matter and clay content with respect to the soil survey available. Extrapolation and applicability of this study to other areas remain to be tested.
Address for correspondence: Mauricio Castro-Franco, PhD candidate, Department of Agronomy, Instituto Nacional de Tecnologia Agropecuaria, Km. 73.5 Ruta 226, Balcarce Buenos Aires, Argentina 7620. E-mail: email@example.com
Financial Disclosures/Conflicts of Interest: None reported.
Received November 24, 2014.
Accepted for publication April 13, 2015.
Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved.