Tackling Spatial Heterogeneity in Global Soil Total Carbon Mapping using a Mixture of Localised Experts

Accurate, high-resolution mapping of soil total carbon (TC) stocks on a global scale is fundamental to global carbon cycle modeling, international climate policy (e.g., IPCC inventories), and sustainable land management. Current Digital Soil Mapping (DSM) efforts often rely on monolithic global machine learning models that frequently fail to capture fine-scale local variability and are prone to significant regional biases. These biases stemmed from spatial non-stationarity, disjointed calibration datasets from varied sources, and instrumentation mismatches, leading to poor predictive performance and high uncertainty in under-sampled regions. To address this critical challenge, the Mixture of Localised Experts (MoLE) framework was introduced as a novel deep learning architecture designed for robust and responsible soil property prediction. The MoLE framework overcame the limitations of traditional GLOBAL–LOCAL approaches by employing a dynamic gating network (router) that learned to partition the problem space. This router intelligently directed input data comprising multiple proximal soil sensor features from a multinational dataset to one of several specialised “expert” sub-models. Each “localised expert” was trained to become highly proficient within a specific geographical or data-driven domain, effectively creating a single, cohesive model that “thinks globally but acts locally.” This framework was developed using a large, harmonised proximal sensor dataset (n = 1443) from six countries across five continents to predict TC. When assessed against an independent hold-out validation set, the MoLE framework demonstrated outstanding precision for TC prediction, achieving a coefficient of determination (R²) of 0.98 and a root mean squared error (RMSE) of 0.06%. Crucially, the results indicated that the MoLE architecture substantially reduced regional prediction bias. The interpretable routing mechanism offered fresh perspectives on model decision-making, revealing the experts activated for various ecoregions and boosting the transparency of the model. The MoLE framework offered a scalable, resilient, and comprehensible framework for the advancement of next-generation global soil information systems. By adeptly addressing spatial heterogeneity and reducing regional bias, this methodology represented a substantial advancement in the precise quantification of global TC stocks.