Mixture-of-algorithmic-experts: a robust and responsible framework for multinational-scale proximal sensor-based soil characterization

Accurate multinational soil prediction remains a fundamental challenge due to pedogenic heterogeneity, management-induced variability, and sample distribution imbalance that constrain the transferability of conventional global calibration models. To address these limitations, this study introduces the Mixture-of-Algorithmic-Experts (MoAE) framework, an architectural approach that decouples global coverage from local specialization by combining algorithmically specialised expert ensembles with a learned routing mechanism for conditional computation. The framework was evaluated using proximal sensor data (visible near infrared spectroscopy and portable X-ray fluorescence spectrometry) from four pedoclimatically distinct countries (Brazil, France, India, and the USA), MoAE was evaluated across 17 physicochemical soil properties encompassing carbon fractions, particle-size distribution, cation exchange capacity, electrical conductivity, and macro- and micronutrients. In the primary hold-out validation, the framework achieved high validation performance for structurally stable properties, including total carbon (Coefficient of Correlation (R2) = 0.99), total nitrogen (R2 = 0.99), and texture fractions (R2 = 0.95–0.98), with negligible systematic bias, substantially exceeding typical global spectroscopic calibration benchmarks reported in the literature. In contrast, lower performance for management-sensitive nutrients such as available S and P reflected signal limitations inherent to the sensing-feature space rather than architectural failure. Shapley Additive Explanations-based interpretability confirmed that predictions were driven by geochemically meaningful elemental proxies and diagnostically relevant spectral wavelengths, while routing-weight distributions provided an internal confidence indicator for deployment-aware decision support. Unlike monolithic calibration strategies, MoAE enables context-aware computation, dynamically weighting expert contributions according to sample-specific feature representations. By integrating conditional modelling, interpretability, and multi-sensor complementarity within a unified framework, MoAE offers a scalable and responsible foundation for multinational digital soil mapping and next-generation operational soil intelligence systems.