1 INTRODUCTION Soil organic carbon (SOC) sequestration has re- ceived much attention recently as the concentration of CO 2 rises in the atmosphere, intensifying climate change. Considering global issues of significance, such as food security, land degradation, and global climate change, more emphasis is needed on main- taining soil’s natural condition under the impact of disturbances and human activities (Grunwald et al., 2011). Characterizing organic carbon (C) pools across large regions is critical to understanding the dynamics of soil C in the context of climate change. Research on soil C has mainly focused on region- al and local scales, while continental scale studies at a spatial resolution that matches the underlying soil C variability is lacking. Comparison of regional digital soil models is hampered by the fact that stud- ies differ in terms of soil C measurement techniques, sampling densities, sample protocols, environmental co-variates (predictor variables), and statistical and geostatistical methods used to predict soil properties. But all these factors may play a role when continen- tal and global digital soil models are created (Grun- wald et al., 2011). To assess differences in the strength and magnitude in soil prediction models developed in regions that show contrasting topogra- phy, ecology, parent material, soils, land use and/or climate, soil and environmental data inputs and model setup need to be standardized. Understanding the effects of regional soil C predictions may then provide guidance to upscale to larger scales (e.g., whole U.S. or even larger extent). Our specific objective was to compare the predic- tion performance of soil C models in two contrasting regions in the United States. These two regions dif- fered in terms of topography, parent material, soils, geology, climate, and land use. 2 STUDY AREAS Two states in the U.S. were selected: Colorado and Florida, which are extremely different in terms of ecoregion, climate zone, topography and other eco- logical conditions. 2.1 Colorado Colorado is a U.S. state that encompasses most of the Southern Rocky Mountains as well as the north- eastern portion of the Colorado Plateau and the western edge of the Great Plains (Fig. 1). Colorado is the 8th most extensive (size: 269,601 km 2 ). Colo- rado is noted for its vivid landscape of mountains, forests, high plains, mesas, canyons, plateaus, rivers, and desert lands. The climate of Colorado is quite complex compared to most of the United States due to the complex orography (1,011~4401.2 m). North- east, east, and southeast Colorado are mostly the high plains, while Northern Colorado is a mix of high plains, foothills, and mountains. Northwest and west Colorado are predominantly mountainous, with some desert lands mixed in. Southwest and southern Colorado are a complex mixture of desert and moun- tain areas. 2.2 Florida Florida is a state in the southeastern U.S., located on the southern coastal plain. It is bordered to the west Cross-regional Digital Soil Carbon Modeling in Two Contrasting Soil- Ecological Regions in the U.S. B. Cao, S. Grunwald & X. Xiong University of Florida, Gainesville, United States of America ABSTRACT: The implications of transposing regional digital soil models to continental and global scales are still poorly understood. This paper presents a ‘controlled landscape experiment’ where soil organic carbon (SOC) stocks were predicted using a standardized set of nationally available environmental predictor varia- bles and sampling density of soil observations. Our specific objective was to compare the prediction perfor- mance of SOC in two contrasting regions in the United States. We used soil samples in the topsoil (0-20 cm) depth across Colorado and Florida from the U.S. National Soil Survey Database (Natural Resource and Con- servation Service, NRCS). Environmental covariate sets were assembled representing a subset of STEP- AWBH factors (S: soils, T: topography, E: ecology, P: parent material, A: atmosphere, W: water, B: biota, and H: human). We used single regression trees (RT) and support vector machines (SVMs) and various error metrics to assess the prediction performance of SOC stocks. Our results demonstrate that in ecologically con- trasting states, both RT and SVM could produce moderate good predictions with R 2 of 0.76 (SVMs) in Flori- da and R 2 of 0.62 (RT) in Colorado. The differences in model results elucidate on the contrasting relation- ships between SOC and environmental predictors that were climate, soil and vegetation driven in Colorado and soil and vegetation driven in Florida. These findings have implications for upscaling of regional digital soil models to continental and global scales, specifically for SOC modeling across the U.S.