Spatial prediction and mapping of soil pH across a tropical afro-montane landscape

Abstract

Soil pH is an indispensable part of the soil bionetwork, but so vulnerable to the dynamics of land-use change. Understanding the spatial and temporal distribution of soil pH is thus important for sustainable land use planning and management. In this study, we developed an approach based on geographically weighted regression-kriging (GWRK) to predict the continuous variation of soil pH up to 15 cm depth. The approach is compared to multiple linear regression-kriging (MLRK) using a total of 220 soil observations and 28 candidate auxiliary data, including climatic, topographic, and remotely sensed data. Results revealed that organic matter, sand, silt, temperature, and Landsat 8 operational land imager band 7 can explain the spatial variability of soil pH. GWR-based (local) models remarkably improved the prediction accuracy of soil pH compared to the MLR-based (global) models considering the root mean square error (i.e., 0.33 vs. 0.15; 0.37 vs. 0.17, respectively) and the coefficient of multiple determination (R2) (i.e., 0.59 vs. 0.74; 0.43 vs. 0.55, respectively). In conclusion, the use of GWR-based models revealed that the correlations between soil pH and environmental covariates were not stationary in space. Forested areas tended to be very strongly to moderately acidic, while croplands were slightly acidic to neutral. The GWR-based models demonstrated their utility in improving predictive soil mapping. These findings will support spatially-targeted soil pH management for improved food security and ecosystem health.