Impact of multi-scale predictor selection for modeling soil propertiesby Bradley A. Miller, Sylvia Koszinski, Marc Wehrhan, Michael Sommer



Soil Science


Selection harvests in Amazonian rainforests: long-term impacts on soil properties

K.L. McNabb, M.S. Miller, B.G. Lockaby, B.J. Stokes, R.G. Clawson, J.A. Stanturf, J.N.M. Silva

Total and active thymoglobulin levels: effects of dose and sensitization on serum concentrations

Jeffrey F Regan, Corinne Lyonnais, Kevin Campbell, Le Van Smith, Roland Buelow, for the US Thymoglobulin™ Multi-C

A multi-scale disk dynamo model

R. Stepanov, P. Frick, D. Sokoloff

Soil compaction impact and modelling. A review

Muhammad Farrakh Nawaz, Guilhem Bourrié, Fabienne Trolard

Multi-Information Fusion for Scale Selection in Robot Tracking

Xiaoqin Zhang, Hong Qiao, Zhiyong Liu


ric arch used pem ften ly o rod cla ll as ol w or p bstitute other predictors to compensate for themissing variables. However, a better includ , and th e latte ection ledge. ata av predictor variables can be overlooked. Primar additional remote sensing data. The exclusion of these potential predic- the possibility of dismissing terrain variables as not useful predictors and references thereof resolutions — often

Geoderma 239–240 (2015) 97–106

Contents lists available at ScienceDirect

Geode l seproperty is known as the scale effect of themodifiable area unit problem (MAUP) (Openshaw, 1983; Armhein, 1995; Jelinski and Wu, 1996).

Because of MAUP, different analysis scales can be better suited for coarse, which can affect a modeler's decision to utilize that data (Woodcock and Strahler, 1987). Remote sensing data is also scale dependent, but it is most commonly used at the analysis scale of indi-et al., 2004; Roecker and Thompson, 2010). Therefore, different analysis scales have the potential to produce different results. This geographic able for digital soil mapping (Mulder et al., 2011 in). However, the data is available in a varietytors may be due to unawareness of their availability, belief in their similarity to included predictor variables, or belief in the superiority of the included predictor variables.

Land-surface derivatives are scale dependent (Wood, 1996; Albani of soil properties when the selected analysis scale does not match the scale of the phenomenon being modeled (Claessens et al., 2005;

Schoorl and Veldkamp, 2006; Drăgut et al., 2009).

Remote sensing has also been shown to be a useful predictor vari-representing different phenomena. Indeed, it ⁎ Corresponding author.

E-mail (B.A. Miller), skoszins (M. Wehrhan), (M. Som 0016-7061/© 2014 Elsevier B.V. All rights reserved.y examples of commonly s of terrain analysis and advertently by the best available resolution (Moore et al., 1993; Sharma et al., 2011). The limited scope of terrain variables considered leads tooverlooked variables include different scale1. Introduction

Modeling is a three part process that tor variables, the calibration of a model soil mapping research emphasizes th majority of the predictor variable sel making of the researcher's expert know dictor variables is always limited by dline subsets with the most commonly used predictors for digital soil mapping at a single scale, the use of multiscale predictor variables produced an improvement in model performance ranging from negligible to a 70% increase in the adjusted R2. Although the scale effect of the modifiable area unit problem is generally well known, this study suggests digital soil mapping efforts would be enhanced by the greater consideration of predictor variables at multiple analysis scales. © 2014 Elsevier B.V. All rights reserved. es the selection of predice validation. Often digital r two parts, leaving the process to the decisionAlthough the pool of preailability, many possible soil scientists in the field use different analysis scales for different terrain variables when describing hillslope position (Miller, 2014). In contrast, it has been the mode in digital soil mapping to use the digital elevation model (DEM) with the best available resolution and to calculate land-surface derivatives at a single analysis scale (e.g. Zhu et al., 2001; Florinsky et al., 2002; Shi et al., 2009; Yang et al., 2011; Lacoste et al., 2014), usually only with a 3 × 3 cell analysis window. This focus on cell resolution alone leads to the analysis scale being determined in-Remote sensing

Digital terrain analysis performing model was always found by considering predictor variables at multiple scales. Compared with basePredictor variables limited predictor pools can suImpact of multi-scale predictor selection fo

Bradley A. Miller ⁎, Sylvia Koszinski, Marc Wehrhan, M

Leibniz Centre for Agricultural Landscape Research (ZALF) e.V., Institute of Soil Landscape Rese a b s t r a c ta r t i c l e i n f o

Article history:

Received 19 June 2014

Received in revised form 22 September 2014

Accepted 24 September 2014

Available online xxxx


Digital soil mapping

Analysis scale


Applying a data mining tool of predictors for soil–landsca for digital soil mapping are o

Predictor variables common additional remote sensing p included qualitative location hydrologic indicators, as we tions). Subsets of the full po the different starting predict j ourna l homepage: www.ehas been observed that (S. Koszinski), mer).modeling soil properties hael Sommer , Eberswalder Straße 84, 15374 Müncheberg, Germany regularly in digital soil mapping, this research focuses on the optimal inclusion odeling by utilizing aswide of a pool of variables as possible. Predictor variables chosen on the basis of data availability and the researcher's expert knowledge. verlooked include alternative analysis scales for land-surface derivatives and ucts. For this study, a pool of 412 potential predictors was assembled, which sses, elevation, land-surface derivatives (with a wide range of analysis scales), proximal and remote sensing (from multiple sources with a variety of resoluere also examined for comparison. The performance for the models built from ools was analyzed for seven target variables. Results suggest that models with rma v ie r .com/ locate /geodermavidual pixels. Therefore, the analysis scale of remote sensing is generally directly linked to resolution, where the intensity value of the raster cell is an amalgamation of the properties in the instantaneous field of view as detected by the sensor (Campbell and Wynne, 2011). Depending on the scale of the phenomena that the researcher is trying to measure or model, different resolutions may have different efficacies for capturing the desired variability (Moellering and Tobler, 1972).With this concept in mind, the choice for inclusion of remote sensing predictor variables commonly depends on the availability of data that the researcher deems sufficient in resolution.