Fusion of Spectral and Electrochemical Sensor Data for Estimating Soil Macronutrients

Rapid and efficient quantification of plant-available soil phosphorus (P) and potassium (K) is needed to support variable-rate fertilization strategies. Two methods that have been used for estimating these soil macronutrients are diffuse reflectance spectroscopy in visible and near-infrared (VNIR) wavelength ranges and electrochemical sensing using an automated ion-selective electrode (ISE) system. The goal of this research was to compare P and K estimates with VNIR spectroscopy, ISE analysis, and a data fusion approach that combined the two methods. Diffuse reflectance spectra of air-dried, sieved samples were obtained in the laboratory. Estimates of P and K based on ISE measurements and standard laboratory analysis methods were available for the same samples from a previous study. Calibrations relating spectra to P, K, and other soil properties determined by standard laboratory methods were developed using partial least squares regression with leave-one-out cross-validation. Although good models (R2 = 0.83 to 0.92, residual prediction deviation (RPD) = 2.45 to 3.60) for soil texture fractions, organic matter, and CEC were obtained using the spectral data, models for pH, P, and K were not good (R2 ≤ 0.7, RPD ≤ 1.81). Models for P and K based on ISE data fit better (R2 ≥ 0.88, RPD ≥ 2.96), but both P and K were systematically underestimated by over 40% due to a large slope bias in both models. Including both spectral and ISE information in P and K calibrations improved model fits (R2 ≥ 0.90, RPD ≥ 3.19) and reduced the systematic slope bias underestimation to less than 10%, resulting in a considerable improvement in accuracy over ISE data alone. In P and K models including ISE data and spectrally estimated soil properties, significant variables included texture and cation exchange capacity, suggesting that these soil properties might affect the relationship of ISE-measured P and K to values determined by standard laboratory methods. These improved results obtained with a data fusion approach using laboratory-collected data are promising and warrant further investigation with additional datasets.