Improved Kriging-based Spatial Interpolation Technique for Mapping Soil Organic Carbon using Categorical Auxiliary Information in Forest Ecosystems

Abstract

Accurate prediction of the soil organic carbon (SOC) distribution in forest ecosystems from limited georeferenced data has received considerable attention, as it underpins land-based climate mitigation strategies and supports sustainable forest management. Kriging-based interpolation methods incorporating auxiliary variables are widely applied, yet the use of categorical information, strongly associated with local trends in forest SOC spatial variability, remains relatively underexplored and deserves further exploration. This study compared conventional ordinary kriging (OK) with stratified kriging approaches, namely kriging with soil type (KST), forest type (KFT), and combined soil type-forest type pattern (KSTFT), using soil samples (0–30 cm) collected from 104 locations in forested areas of Danang city, Vietnam. The performance of the interpolation models was evaluated using mean error (ME), mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R2), while prediction uncertainty was assessed through the standard deviation (SD) of kriging estimates. Both soil type and forest type possessed significant influences on the spatial variability of SOC in forest landscapes. The stratified kriging approaches consistently outperformed OK in predicting forest SOC content, both in terms of estimation accuracy and reliability. Among these, KSTFT attained the greatest accuracy (R2 = 0.72, ME = -0.20 t.ha− 1, MAE = 9.46 t.ha− 1, RMSE = 12.03 t.ha− 1) and the lowest uncertainty (mean SD = 12.63 t.ha− 1). The kriged maps generated by KST, KFT, and KSTFT appeared more realistic, with fragmented polygon structures and abrupt transitions, compared to those produced by OK, indicating that these models more effectively captured the spatial variability of forest SOC. Our findings highlight the potential of kriging integrated with categorical information to alleviate the effects of the data variation issue and smoothing effect, resulting in enhanced accuracy and reliability of forest SOC prediction.