Spatial Disaggregation and Harmonization of gSSURGO.

An increase in high performance computing resources and accessibility of high-resolution land data allows us to address many unresolved earth science challenges, such as the lack of high-resolution soil data at continental scales. This data would be helpful for agriculture, hydrologic modeling, and resource planning.

Current available soil datasets are mostly based on legacy polygon based datasets built from surveys (ground truth) and local expert knowledge. These products can be difficult to use at regional to continental scales due to surveyor biases (e.g. county and state boundary discontinuities). Determining how to disaggregate the local soil survey information (e.g. map units) to finer spatial resolutions is also a challenge. Using the DSMART algorithm, Odgers et al. (2014) shows how machine learning can be used to harmonize and spatially disaggregate these products by relating high resolution soil covariates to available observations.

In this study, an enhanced DSMART algorithm is applied over CONUS at a 1 arcsec spatial resolution. The gSSURGO database provides the ground truth and the USGS NED, MLRC NLCD, and USGS aeroradiometric datasets the soil covariates. Using a moving window approach, random forests are fit and used to estimate the 50 most probable soil classes and their associated probabilities at each 1 arcsec grid cell over CONUS (~9 billion grid cells). We will discuss the value and accessibility of the new dataset, its applications, and the potential for applying this method globally.

References: Odgers, N. P., W. Sun, A. B. McBratney, B. Minasny, and D. Clifford, 2014: Disaggregating and harmonising soil map units through resampled classification trees. Geoderma, 214, 91-100.