Field-scale digital soil mapping in the era of tabular foundation models

Digital soil mapping at the field-scale faces a fundamental challenge of building accurate predictive models from small, high-dimensional tabular datasets where training sample sizes are limited by cost and labor constraints. Traditional machine learning methods like Random Forest have long dominated pedometrics, but recent advances in artificial neural network architectures challenge this status. To investigate this, we develop a comprehensive evaluation framework built upon LimeSoDa, our diverse and fully open-access collection of field-scale digital soil mapping datasets. This allows us to assess the application of modern neural networks in pedometrics under realistic conditions of data scarcity. Our results demonstrate that contemporary architectures consistently outperform classical methods when coupled with specific methodological enhancements that address training instability. In-context learning tabular foundation models, such as TabPFN, show particular promise and surpass established baselines even on very small datasets. We go further and investigate the application of tabular foundation models on datasets with unfavorable feature-to-sample ratios typical in soil spectroscopy. Building upon principal component analysis and partial least squares, we propose hybrid strategies that effectively address the challenges posed by soil spectroscopy datasets. Going beyond purely tabular regression modeling, we extend our framework to incorporate spatial information through Kriging prior Regression, integrating geostatistical features into tabular machine learning predictions and further improving accuracy when sensor data alone provide limited information. Our findings establish a new baseline for field-scale digital soil mapping and offer methodological insights applicable to any precision agriculture domain constrained by small tabular datasets.