An ensemble estimate of Australian soil organic carbon using machine learning and process-based modelling

Abstract. Spatially explicit prediction of soil organic carbon (SOC) serves as a crucial foundation for effective land management strategies aimed at mitigating soil degradation and assessing carbon sequestration potential. Here, using more than 1000 in-situ observations, we trained two machine learning models (random forest, and K-means coupled with multiple linear regression), and one process-based model (the vertically resolved MIcrobial-MIneral Carbon Stabilization (MIMICS)) to predict SOC content of the top 30 cm of soil in Australia. Parameters of MIMICS were optimized for different site groupings, using two distinct approaches, plant functional types (MIMICS-PFT), and the most influential environmental factors (MIMICS-ENV). We found that at the continental scale, soil bulk density and mean annual temperature are the dominant controls of SOC variation, and that dominant controls vary for different vegetation types. All models showed good performance in SOC predictions with R² greater than 0.8 during out-of-sample validation with random forest being the most accurate, and SOC in forests is more predictable than that in non-forest soils. Parameter optimization approaches made a notable difference in the performance of MIMICS SOC prediction with MIMICS-ENV performing better than MIMICS-PFT especially in non-forest soils. Digital maps of terrestrial SOC stocks generated using all the models showed similar spatial distribution with higher values in southeast and southwest Australia, but the magnitude of estimated SOC stocks varied. The mean ensemble estimate of SOC stocks was 30.08 t/ha with K-means coupled with multiple linear regression generating the highest estimate (mean SOC stocks at 38.15 t/ha) and MIMICS-PFT generating the lowest estimate (mean SOC stocks at 24.29 t/ha). We suggest that enhancing process-based models to incorporate newly identified drivers that significantly influence SOC variations in different environments could be key to reducing the discrepancies in these estimates. Our findings underscore the considerable uncertainty in SOC estimates derived from different modelling approaches and emphasize the importance of rigorous out-of-sample validation before applying any one approach in Australia.