Spatial downscaling of digital soil organic carbon map using Dissever algorithm

In most national and regional projects, in order to produce a digital map of soil properties, usually sampling density is low due to costly and time-consuming. Because of this, produced digital maps have a large spatial resolution (more than 90 meters) that can’t be used on a farm scale (spatial resolution less than 30 meters). One way to solve this problem is to downscale of digital maps with coarse spatial resolution using covariates with fine spatial resolution. The purpose of this study was to investigate the efficiency of the Dissever algorithm for producing an organic carbon map with a spatial resolution of 30 m from a carbon-organic digital map with a spatial resolution of 90 meters.

The study area is approximately 14084 hectares and formed a small part of the Karkhe catchment in Kermanshah province. Initially, using 110 random observations and block Kriging method, an organic carbon map was prepared with a pixel size of 90 m. Then a set of 23 environmental covariates associated with scorpan factors including climate, topography, living organisms and maternal materials were used as predictor variables. SRTM digital elevation model and Landsat 5 satellite imagery were used to provide climatic and topographic predictors and vegetation and geology indicators. The dissever algorithm is a repetitive process for approaching a mass balance solution. Multiple regression model, generalized additive model, cubist, random forest, and ensemble model used for production fine spatial resolution map. In order to evaluate the efficiency of different methods, restored maps (obtained by converting downscaled maps with 30 m resolution to maps with 90 m resolution using mean filter) were compared with base map (block kriging map) using validation criteria include Bias, R2, root mean square error, and concordance correlation coefficient. The spatial structure of the restored maps and the base map was also investigated using parameters of their experimental variogram.

The relationship between the covariates and the soil organic carbon using data mining methods in the framework of dissever algorithm resulted in the production of downscaled maps. The results showed that the probability density function of the restored map of cubist model is very close to the base map probability density function. Also, the downscaled map using the cubist model had the highest coefficient of determination (0.75) and concordance correlation coefficient (0.8) and the lowest root mean square error (0.06) and bias (0.001). Thus, cubist model have the highest efficiency of downscaling in compared to the rest of models. . It was also found that the use of ensemble model increases the accuracy and precision of downscaled map compared to single data mining models. The study on the spatial structure of restored maps indicates that the cubist restored maps captured more of the variance of the base map than others.

The modified dissever procedure due to the use of data mining methods and ensemble model is a practical option to downscale soil properties map with coarse resolution. Considering the high efficiency of dissever algorithm, this method can be used to prepare soil properties map at field scale from national and regional maps which can be used in farm management.