Geographically Weighted Regression Kriging Research Articles

A geographically weighted regression kriging approach for mapping soil organic carbon stock

Local variations in the model parameters can play an important explanatory role in the spatial modeling of soil organic carbon (SOC) stock. Linear regression models assume parameters to be spatially invariant and are unable to account for the spatially varying relationships in the variables. A recently developed approach, geographically weighted regression kriging (GWRK), was used in this study to examine the relationships between environmental variables and SOC stock for the state of Pennsylvania, USA. The specific objectives were to (i) estimate the SOC stock (kgCm−2) to 1.0-m depth, and (ii) compare the GWRK results with those obtained from regression kriging (RK). Data for 878 georeferenced soil profiles, extracted from National Soil Survey Center database, were divided into calibration (n=702) and validation (n=176) datasets. Environmental variables including temperature, precipitation, elevation, slope, geology, land use, and normalized difference vegetation index were explored and included as independent variables to establish the model for estimating the SOC stock. Results using Pennsylvania as a case study conclude that GWRK was the least biased and more accurate compared to RK for estimating the SOC stock based on the lowest root mean square error (2.61 vs. 4.61kgm−2), and high R2 (0.36 vs. 0.23) values. Higher stock was consistent with higher precipitation and cooler temperature of the region. Total SOC stock ranged from 1.12 to 1.18Pg for the soils of Pennsylvania. Forests store the highest SOC stock (64% of the total), followed by croplands (22%), wetlands (2.3%), and shrubs (2%). Results show that GWRK enhances the precision for estimating the SOC stock compared to the RK since the former takes into account the spatial non-stationarity coupled with spatial autocorrelation of the residuals.

Mapping the organic carbon stocks of surface soils using local spatial interpolator

The largest uncertainties are associated with estimating the soil organic carbon (SOC) stock because of natural soil variability and data scarcity. Thus, a local spatial geostatistical hybrid approach, the geographically weighted regression kriging (GWRK), was used in the present study to overcome some of these uncertainties. This study was designed to estimate the SOC stock (kg C m(-2)) for the surface 0 to 15 cm depth using the state of Pennsylvania as the study region. A total of 920 soil profiles were extracted from the National Soil Survey Center database and were divided into calibration (80%) and validation (20%) periods. Some soil parameters that include clay content, bulk density (ρ(b)), total nitrogen (TN) content, pH, Ca(2+), Na(+), extractable acidity (EXACID), and cation exchange capacity (CEC) were used as covariates for estimating the SOC stock. These covariates exhibited spatial autocorrelation (Moran's Index, I = 0.62 to 0.89). Further, residuals of geographically weighted regression were spatially autocorrelated, and hence support the use of the GWRK approach. Validation results concluded that the performance of the GWRK approach was the best with the lowest values of root mean square error, mean estimation error and mean absolute estimation error. The estimated SOC stock for the surface 0 to 15 cm depth ranged from 1.41 to 3.94 kg m(-2). Results from this study show that the GWRK captures spatial dependent relationships, and addresses spatial non-stationarity issues, hence this approach improves the estimations of SOC stock.