Application of Machine Learning Approaches to Predict Soil Inorganic Carbon in Drylands

The importance of soil inorganic carbon (SIC) in drylands carbon sequestration and global carbon is well known, but very few modeling studies in cold deserts have focused on SIC formation, change, and distribution. Predicting SIC distribution and storage can pave the way for enhancing the accuracy of climate and hydrological models. However, high variability in the distribution of SIC presents a challenge to mapping and spatio-temporal modeling of SIC since uncertainty increases with distance from the sampled area. Prior studies measured and mapped SIC content temporally and spatially; however, the majority of these works applied one-step models for SIC prediction. One-step modeling methods are less accurate in predicting environmental properties that are zero-inflated such as SIC datasets. Using a two-step machine learning (ML) model approach can manage the high number of samples with zero SIC values and increase the accuracy of the model. Considering the availability of satellite imagery and ML methods, this study aims to develop a predictive ML SIC regressor model for better estimation of SIC distribution and storage in a native sage-steppe ecosystem across the Reynolds Creek Experimental Watershed (RCEW) and Critical Zone Observatory in southwestern Idaho. ML methods will decrease the uncertainty in unsampled regions compared to spatial interpolation methods like Kriging.

Overcoming the challenging aspect of this research requires extracting land surface features to train the model and reduce the uncertainty in unsampled regions. To see the impact of vegetation on SIC formation, we will use surface reflectance Landsat 5, 7, and 8 multispectral imagery to drive vegetation indices such as Normalized Difference Vegetation Index. A high-resolution USGS DEM dataset will be used to extract the relationship between elevation and formation of SIC. Other characteristics that can inform the model include MODIS Land Cover dataset, land use dataset, and precipitation average. Different ML models will be tested using these datasets and SIC data from previous studies in RCEW to produce an accurate SIC model and a spatially distributed SIC map for RCEW. Results of this study, along with process-based models for predicting soil organic carbon in RCEW, can improve the accuracy of our estimates of carbon storage in drylands.