| Predictive mapping of organic carbon content in soils of Russia using ensemble machine learning |
| Paper ID : 1070-ICRSSSA-FULL |
| Authors |
|
Andrey Chinilin *1, Igor Savin2 1FRC "V.V. Dokuchaev Soil Science Institute" 2FRC “V.V. Dokuchaev Soil Science Institute” |
| Abstract |
| The study reflects an understanding of individual factors regulating and controlling the content of organic carbon of soils, and shows a modern quantitative assessment of the content of organic carbon of soils in Russia, taking into account a huge variety. Article presents the results of three-dimensional modeling of the organic carbon content of soils with 500 m spatial resolution at several standard depths (0-5, 5-15) to the territory of the Russian Federation using ensemble machine learning. Automated predictive mapping was based on 4 961 soil horizons from 863 soil profiles, as well as on an extensive set of spatial information, including bioclimatic variables, a digital elevation model and its derivatives, and long-term averaged time series of MODIS data. An ensemble machine learning algorithm (stacking, stacked generalization, stacked regression) was used to build models of spatial and vertical distribution. The accuracy of the obtained cartographic models was assessed using spatial cross-validation. The results of spatial cross-validation show lower accuracy: the coefficient of determination is 0.46, CCC 0.63, logRMSE 0.88 (RMSE 1.41 g/kg) compared to randomize (R2cv - 0.68, CCC - 0.81, logRMSE - 0.68 (RMSE 0.97 g/kg)). The proposed quantitative assessment is fully automated and allows to reproduce the modeling and refine the results as new soil data are obtained. |
| Keywords |
| soil carbon, spatial modeling, stacked regression, spatial cross-validation, Russia |
| Status: Accepted (Oral Presentation) |
