『Abstract
　Typically, regional assessment of the spatial variability and distribution of environmental properties are constrained by sparse field observations that are costly and labor intensive. We adopted a hybrid geospatial modeling approach that combined sparsely measured soil NO3-N observations collected in three seasons (Sept. 2003, Jan. and May 2004) with dense auxiliary environmental datasets to predict NO3-N within the Santa Fe River Watershed (3585 km²) in north-east Florida. Elevated nitrate-nitrogen concentrations have been found in this watershed in spring, surface and ground water. We collected soil samples at four depths (0-30, 30-60, 60-120, 120-180 cm) based on a random-stratified sampling design. Classification and regression trees were used to develop trend models for soil NO3-N predictions based on environmental correlation and predict values at the watershed scale. Residuals were spatially autocorrelated only for the Jan. 2004 sampling and regression kriging was used to combine the kriged residuals with tree-based trend estimates for this event. At each step of the upscaling process, error assessment provided important information about the uncertainty of predictions, which was lowest for the Jan. sampling event. Sites that showed consistently high NO3-N values throughout the cropping season (Jan-May 2004) with values ≦5μg g^-1 covered 95.7％ (3363.9 km²) of the watershed. Values in the 5-10μg g^-1 range covered 4.3％ (150.7 km²)) while values exceeding 10μg g^-1 covered only 0.59％ (20.7 km²) of the watershed. Elevated soil NO3-N on karst, unconfined areas with sand-rich soils, or in close proximity to streams and water bodies pose the greatest risk for accelerated nitrate leaching contributing to elevated nitrogen found in spring, surface and ground water in the watershed. This approach is transferable to other land resource problems that require the upscaling of sparse site-specific data to large watersheds.

Keywords: Nitrate-nitrogen; Geospatial; Classification and regression trees; Regression kriging』

1. Introduction
2. Materials and methods
　2.1. Study area
　2.2. Field data collection
　2.3. Spatial datasets
　2.4. Upscalling methods
3. Results and discussion
　3.1. The NO3-N dataset - exploratory data analysis
　3.2. Regional predictions of soil NO3-N
　　3.2.1. Classification tree model - Sept. 2003 sampling event
　　3.2.2. Regression tree model - Jan. and May 2004 sampling events
　　3.2.3. CART model error
　　3.2.4. Upscaling of the tree model predictions
4. Summary and conclusions
Acknowledgements
References