Vis-NIR and XRF data fusion for soil organic carbon inference: Insights from large-scale spectral libraries and fitness-for-purpose evaluation

Visible and near-infrared diffuse reflectance spectroscopy (vis-NIR) and X-ray fluorescence spectroscopy (XRF) are promising alternatives for rapid, reagent-free soil analysis that can help scale up soil organic carbon (SOC) monitoring for soil health and carbon credit reporting. We investigated whether combining laboratory-based vis-NIR and XRF spectra improves SOC prediction relative to vis-NIR alone, and whether the resulting performance is fit for purpose for monitoring SOC variability at farm scale. A national spectral library of 12,829 Brazilian soil samples with laboratory spectral readings was used to build global Cubist models, and eight local libraries (952 samples) were used to build local partial least squares (PLS) models. Two sensing scenarios were compared: vis-NIR alone and vis-NIR + XRF data fusion. Performance was evaluated using conventional accuracy metrics and a fitness-for-purpose framework based on analytical tolerances and percentile error profiles. The results support the added value of XRF as complementary information to vis-NIR for SOC prediction under both modelling strategies, with the largest relative improvement observed under the global-Cubist strategy (RMSE of 1.11 g kg –1 for vis-NIR+XRF, versus 1.52 g kg –1 for vis-NIR alone) and the best absolute performance achieved under the local-PLS strategy (RMSE of 0.97 g kg –1 with vis-NIR+XRF, versus 1.15 g kg –1 with vis-NIR alone). Using an application-specific analytical tolerance framework, sensor-based monitoring was fit-for-purpose mainly under medium and high within-farm SOC variability, but not under low variability. These results support sensor-based SOC inference as a practical tool for farm-scale monitoring.