Ensemble modelling-based pedotransfer functions for predicting soil bulk density in China

Abstract

Understanding and managing soil organic carbon stocks (SOCS) are integral to ensuring environmental sustainability and the health of terrestrial ecosystems. The information of soil bulk density (BD) is important in accurately determining SOCS while it is often missing in the soil database. Using 3,504 soil profiles (14,170 soil samples) that represented diverse regions across China, we investigated the effectiveness of various pedotransfer functions (PTFs), including traditional PTFs, machine learning (ML), and ensemble model (EM), in predicting BD. The results showed that refitting the parameter(s) in traditional PTFs was essential for BD prediction (coefficient of determination (R2) of 0.299–0.432, root mean squared error (RMSE) of 0.156–0.162 g cm−3, Lin’s concordance coefficient (LCCC) of 0.428–0.605). Compared to traditional PTFs, ML can greatly improve the model performance for BD prediction with R2 of 0.425–0.616, RMSE of 0.129–0.158 g cm−3 and LCCC of 0.622–0.765. Our results also showed that EM can further improve BD prediction by ensembling four ML models (R2 = 0.630, RMSE = 0.126 g cm−3, LCCC = 0.775). Using the EM model, we filled the missing BD (1207 soil profiles with 3,112 soil samples) in our database and built the SOC stock database (4,275 soil profiles with 17,282 soil samples). This study can be a good reference for gap-filling the missing BD depending on the data availability, thus contribute to a deeper understanding in soil C related climate change mitigation, ecological balance preservation and environmental sustainability promotion.

Keywords

• Soil organic carbon stock
• Variable selection
• Machine learning
• Land cover
• National scale
• Soil database