Distinguishing soil horizons using proximal sensing and machine learning

Authors: GÖZÜKARA, GAFUR - University Of Wisconsin; HARTEMINK, ALFRED - University Of Wisconsin; Young, Eric; ZHANG, YAKUN - Oregon State University

Submitted to: European Journal of Soil Science
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/15/2026
Publication Date: 8/20/2026
Citation: Gözükara, G., Hartemink, A.E., Young, E.O., Zhang, Y. 2026. Distinguishing soil horizons using proximal sensing and machine learning. European Journal of Soil Science. https://doi.org/10.1016/j.catena.2026.110523.
DOI: https://doi.org/10.1016/j.catena.2026.110523

Interpretive Summary: Soil types have a major impact on crop yield potential, water movement, and nutrient cycling. Accurate soils maps are important for crop selection, fertilizer and manure nutrient application rates, and determining nutrient loss risk from surface runoff and leaching through the soil profile. Traditional soil mapping techniques rely on extensive field surveys and provide limited information on physicochemical properties important for cropping system sustainability. Proximal sensing is a rapid analytical method (based on spectroscopic theory) that can be coupled with traditional soil survey methods to provide detailed soil chemical and mineralogical information. Few studies have assessed the ability of proximal sensing with machine learning to predict the occurrence of soil horizons, or distinct vertical soil layers. We sampled 150 locations at the Marshfield Agricultural Research Station during 2024 to assess the efficacy of using proximal sensing (Vis-NIR, MIR, and XRF) and machine learning to predict soil horizon occurrence in the field. Surface horizons were best predicted by MIR or XRF. Model accuracy decreased with horizon depth beneath the surface. The first subsurface horizon was best predicted by MIR, while the deeper horizon was less reliably predicted by any combination. Leveraging machine learning and proximal sensing data shows promise for large scale mapping of soil properties critical for agricultural nutrient use efficiency.

Technical Abstract: Accurate soil data are critical for crop input efficiency and environmental assessments. Combining proximal sensing with machine learning to predict soil horizon properties may offer an effective alternative to traditional mapping. In 2024,150 pedons covering 350 ha at the Marshfield Agricultural Research Station in central Wisconsin were described and sampled for a range of properties. Soil horizon data were collected in the field and samples were taken in 10 cm depth increments to a depth of 1 m. In addition to a range of standard properties (pH, organic carbon contents, and elemental cations), proximal sensing (near infrared [Vis-NIR], mid-infrared [MIR], and x-ray fluorescence spectroscopy [XRF]) quantified mineral phases. Random forest models were developed to predict horizons based on physicochemical properties and/or spectra. Soil surface horizons (Ap) were best predicted by MIR or XRF (R² = 0.98) whereas Glossic horizons were well predicted by combining Vis–NIR and XRF (R² = 0.98). Model accuracy decreased with horizon depth beneath the surface. Pedon Bt horizons were best predicted by MIR (R² = 0.90) while deeper horizons (2Bt) were less reliably predicted (R² = 0.50; combined Vis-NIR, MIR, and XRF). Surface soil horizons were accurately predicted by leveraging machine learning and proximal sensing data, showing promise for large scale mapping of soil properties critical for agricultural nutrient use efficiency.