Mid-infrared spectroscopy for prediction of soil health indicators in the United States

Abstract

Quantifying the holistic notion of soil health requires a large suite of measurements spanning the physical, chemical and biological properties of soil. The cost of measuring the full suite of soil health indicators via traditional methods is cost prohibitive for the spatial and temporal monitoring many land managers would like. Here we investigated the ability to make predictions of numerous soil health indicators from the large mid-infrared (MIR) spectral library developed by the National Soil Survey Center. We developed nationally applicable models with test-set validation coefficient of determination (R2) > 0.80 for water retention (WT), bulk density (BD), texture (as percentages of sand, silt and clay), cation exchange capacity (CEC), exchangeable bases, base saturation (BS), electrical conductivity (EC), pH, soil organic carbon (SOC) and total nitrogen (TN). Using the memory-based learning (MBL) model, a locally weighted approach, we found that this database could also predict aggregate stability, plant available P and K with an R2 > 0.70 on independent validation samples. Models for nitrate and micronutrients were less successful but these properties were only sparsely represented in the database. Two test cases were used to demonstrate the utility of MIR spectroscopy in providing a wealth of data for a fraction of the time and cost involved with traditional analyses. While MIR spectroscopy-based predictions may not always be appropriate for applications requiring very high levels of accuracy, results from this work and elsewhere suggest that this technology can replace, or at least supplement, traditional analyses for many purposes.