Mapping Soil Organic Matter Concentration at Different Scales Using a Mixed Geographically Weighted Regression Method
Canying Zeng,Lin Yang,A-Xing Zhu,David G. Rossiter,Jing Liu,Junzhi Liu,Chengzhi Qin,Desheng Wang
DOI: https://doi.org/10.1016/j.geoderma.2016.06.033
IF: 6.1
2016-01-01
Geoderma
Abstract:The present regression models in digital soil mapping usually assume that relationships between soil properties and environmental variables are always fixed (as in MLR) or varying (as in GWR) in geographical space. In reality, some of the environmental variables may be fixed in affecting soil property variation and some are local varying. In this study, a mixed geographically weighted regression (MGWR) method which can deal with fixed and varying spatial relationships between a target variable and its environmental variables were proposed and used to predict topsoil soil organic matter (SOM) concentration in two study areas (Heshan, Heilongjiang province and Xuancheng, Anhui province, China) at two scales. Three groups of sample sets were created based on the total samples in the study areas to evaluate the robustness and stability of the model. Multiple linear regression (MLR), geographically weighted regression (GWR), GWR-kriging (GWRK), local regression-kriging (LRK), kriging with an external drift (KED), and ordinary kriging (OK) were used for comparison with MGWR. The validation results showed that the use of MGWR reduced the RMSE of GWR by 10.5% and 7.6% on average, reduced the RMSE of MLR by 12.8% and 9.9% on average for Heshan and Xuancheng study areas respectively. MGWR also showed a good competitiveness when compared with GWRK, LRK, KED and OK. In Heshan study area, the influence of flow length, relative position index, foot slope and distance to the nearest drainage were constant, whereas the elevation, topographic wetness index and valley index showed different influence in different regions. In Xuancheng study area, the fixed environmental variables were profile curvature, topographic wetness index and slope, whereas the varying environmental variables were precipitation, temperature, elevation, and limestone. The results indicate that the accuracy of predictions can be improved by adaptive coefficient according to the variation of environmental variables as implemented in MGWR compared with others considering only the local or global relationships. It was concluded that mixed geographically weighted regression model could be a potential method for digital soil mapping.