Mapping surface soil organic carbon density by combining different soil sampling data sources and prediction models in Yangtze River Delta, China

Abstract

Soil organic carbon (SOC) is a crucial soil property governing various agricultural, ecological, and environmental processes in the terrestrial ecosystem. However, acquiring accurate and high resolution SOC density (SOCD) map remained a challenge in the current researches. Based on soil samples collected from soil survey and published literatures (952 samples), this study tested different approaches for generating high accuracy dataset of 0–20 cm SOCD with the spatial resolution of 90-m in the Yangtze River Delta. Two spatial prediction methods, including artificial neural network (ANN, represented non-linear), multiscale geographically weighted regression (MGWR, represented linear) were applied to generated the SOCD maps. Then three fusion models, including multiple linear regression fusion (MLR_F), ANN fusion (ANN_F) and a self-developed model (multiscale geographically normalize weighted fusion, MGNW_F), were applied to fuse the prediction results of spatial interpolation methods. Among spatial interpolation methods, the ANN performed better in capturing the SOCD spatial pattern, while the MGWR performed better in accuracy measures of mean square error, mean absolute error and coefficient of determination. After data fusing by the three models, improved accuracy measures than the ANN and MGWR were achieved. The predictions of fusion models also showed better reliability than the soil dataset of high resolution National Soil Information Grids of China, mainly due to that more soil samples were included in the prediction. Among these fusion models, the MGNW_F performed slightly better than the other two and had the strongest adaptability to the overall environment of the study area. Findings of this study highlighted the potentials of acquiring additional soil samples from existing literatures and fusing different methods for enhancing the accuracy in mapping SOCD.