Comparison of regression methods for spatial downscaling of soil organic carbon stocks maps

This paper presents a refinement of the dissever algorithm, a framework for downscaling spatial information based on available environmental covariates proposed by Malone et al. (2012). While the original algorithm models the relationships between the target variable and the covariates using a general additive model (GAM), the modified procedure presented in this paper allows the user to choose between a wide range of regression methods. These developments have been implemented in an open-source package for the R statistical environment, and tested by downscaling soil organic carbon stocks (SOCS) maps available on two study sites in Australia and New Zealand using 4 different regression methods: linear model (LM), GAM, random forest (RF), and Cubist (CU). In this study, the spatial resolution of a set of reference maps were degraded to a coarser resolution, so to assess the performance of the different downscaling methods. On the Australian site, the 1-km SOCS coarse resolution map has been downscaled to a 90-m resolution. The best results were achieved using either CU or RF (R-2= 0.91 and 0.94 respectively). On the New Zealand site, the 250-m SOCS coarse resolution map has been downscaled to a 10-m resolution. The best results were achieved using GAM (R-2= 0.90). The results illustrate that the optimal regression methods for downscaling spatial information using dissever vary on a case-by-case basis. In particular, simpler approaches such as LM or GAM outperformed more complex approaches in cases where only a limited number of pixels are available to train the downscaling algorithm. This demonstrate the value of an implementation that facilitates testing of different regression strategies. (c) 2017 Elsevier B.V. All rights reserved.