Field-scale digital mapping of soil organic matter using spatially enhanced quantile machine-learning models

Accurate field-scale mapping of soil organic matter (SOM) is essential for implementing precision agriculture strategies that enhance productivity and sustainability by supporting site-specific management. This study assessed three quantile-based machine learning (ML) algorithms—Quantile Regression Forest (QRF), Stochastic Gradient Boosting (SGB), and Deep Learning (DL)—in terms of predictive accuracy, uncertainty quantification, and spatial coherence. The models were trained using 7807 georeferenced SOM samples collected from 2052 fields, together with remote sensing and topographic covariates. To explicitly account for spatial autocorrelation, an additional covariate was derived from ordinary block kriging of SOM. Model performance was evaluated using root mean squared error (RMSE), mean error (ME), prediction interval coverage probability (PICP), and local standard deviation (LSD) as an indicator of spatial smoothness. Spatial validation was used to reduce potential bias arising from spatial autocorrelation. QRF consistently achieved the best balance among accuracy, uncertainty representation, and spatial coherence. Although SGB reached slightly higher accuracy, it underestimated uncertainty and produced noisier spatial patterns. DL generated the smoothest maps but tended to underestimate SOM and provided less reliable uncertainty estimates. Notably, QRF performance remained stable across fields with different sampling intensities, highlighting its robustness and practical relevance in data-limited scenarios. Overall, QRF models enhanced with spatially informed covariates provide a reliable framework for field-scale SOM prediction and uncertainty quantification—critical inputs for optimizing agricultural practices, guiding nutrient management, and supporting sustainable land management. • Spatially enhanced QRF improves field-scale SOM prediction and uncertainty. • A kriging-based covariate captures spatial dependence in ML models. • QRF shows robust performance across varying field sample sizes. • Approach supports site-specific management and sustainable soil practices.