Using Principal Component Regression to Predict Soil Nitrogen Supply Potential from Electromagnetic Induction

Precise nitrogen (N) application is vital to sustainable and profitable corn production. Surplus N can negatively impact air and water quality while inadequate amounts of N for crop growth will have adverse effects on corn yield and quality. Determining the correct amount of N to apply requires knowledge of the soil N supply potential (sNsp), which can be challenging. Electromagnetic induction (EMI) has become increasingly popular in precision agriculture due to its ability to rapidly and inexpensively provide spatially continuous soil characterization. EMI captures bulk soil properties, such as soil salinity, texture, cation exchange capacity, organic carbon, etc. Hence, EMI offers an opportunity to rapidly and inexpensively predict high-resolution sNsp. The objective of this research is to develop a strategy to predict high-resolution sNsp from EMI to improve N fertilizer recommendations.

We employ Principal Component Regression (PCR) to predict high-resolution sNsp from EMI. Specifically, we develop a sNsp predictive model from co-located EMI and sNsp sparse measurements and subsequently apply the model to predict high-resolution sNsp from EMI measurements. To develop a robust predictive model and to enable assessment of uncertainty in the estimation, we apply bootstrapping to draw replicates of the calibration data to perform rotation model calibration. We retain all calibrated models with reasonable cross-validation residuals to establish probability distributions for resampling the regression coefficients. PCR is a particularly appealing nonlinear, multivariate regression technique for our purpose due to its dimensionality reduction capability and its ability to create new uncorrelated independent variables. Using uncorrelated predictor variables in regression analysis eliminates redundancy in the variables (multicollinearity) to improve the predictive accuracy of the model. We illustrate the performance of the strategy on EMI data and 40 soil samples collected on a regular grid from a 40-acre corn field in Western New York. We demonstrate the potential to estimate high-resolution sNsp from EMI measurements to improve N fertilizer recommendations, for profitable corn production in an environmentally friendly manner.