Spatial variability of soil carbon across Mexico and the United States

Abstract:

Soil organic carbon (SOC) is directly linked to soil quality, food security, and land use/global environmental change. We use publicly available information on SOC and couple it with digital elevation models and derived terrain attributes using a machine learning approach. We found a strong spatial dependency of SOC across the United States, but less spatial dependency of SOC across Mexico. Using High Performance Computing (HPC) we derived a 1 km resolution map of SOC across Mexico and the United States. We tested different machine learning methods (e.g., kernel based, tree based and/or Geo-statistics approaches) for computational efficiency and statistical accuracy. Using random forest combined with geo-statistics we were able to explain >70% of SOC variance for Mexico and >40% in the case of the United States via cross validation. These results compare with other published estimates of SOC at 1km resolution that only explain <30% of SOC variance across the world. Topographic attributes derived from digital elevation models are freely available globally at fine spatial resolution (<100 m), and this information allowed us to make predictions of SOC at fine scales. We further tested this approach using SOC information from the International Soil Carbon Network to predict SOC in other regions of the world. We conclude that this approach (using public information and open source platforms for data analysis) could be implemented to predict detailed explicit information of SOC across different spatial scales.