Site specific nutrient management (SSNM) in corn production environments can increase nutrient use efficiency and reduce gaseous and leaching losses. To implement SSNM plans, farmers need methods to monitor and map the spatial and temporal trends of soil nutrients. High resolution electrical conductivity (EC) mapping is becoming more available and affordable. The hypothesis for this study is that EC of the soil, in conjunction with detailed terrain attributes, can be used to map soil nutrients and characteristics. To test this, we have used an extensive data set of EC measurements (EM38, n = 12,980) and soil samplings (n = 489) conducted in 13 corn fields in Illinois, U.S. during the years 2000-2003. Detailed digital elevation model (5m resolution) was generated for each field using ground measurements, and was subsequently used to calculate multiple terrain attributes such as slope, aspect, and curvature etc. The multiple fragmented layers were standardized and unified using Agmatix’s Axiom platform.

To predict soil nutrients, EC data and multiple terrain attributes were split into training (80%) and testing (20%) datasets and used as inputs for a Random Forest model. Moran’s I test confirmed spatial correlation exists for the all predictors and dependent variables. There are multiple techniques to implement spatial effects into digital soil mapping. Here, the modeling was repeated twice – first without accounting for geo-location data, and secondly when each sampling location coordinate was used directly as covariate for the ML model. Without accounting for geo-location, the model obtained Mean Absolute Prediction Error (MAPE) of 24%, 14%, 11% and 8%, and R² of 0.63, 0.68, 0.80, and 0.84 for soil phosphorus, potassium, organic matter and cation exchange capacity, respectively. Accounting explicitly for sampling locations significantly improved predictions, resulting in relatively small MAPE of 11%, 5%, 5% and 4%, and R² of 0.88, 0.93, 0.93, and 0.96 respectively.

Feature importance ranking, conducted using a standard feature permutation approach, found elevation and EC to be the most affecting factors on model predictions when the sampling locations were not accounted for. The location coordinates emerged, however, as the most affecting factor when spatial sampling locations were included. Higher prediction efficiency when spatial coordinates were accounted for could be explained by the strong spatial trends in all data layers. Model validation performance is also effected by the dense grid of soil sampling in each field, and the random splitting of train and test data, resulting in train data to potentially reside spatially close to test data. Altogether, the results of this study suggested that soil mapping using EC data can benefit from inclusion of geo-location information to improve prediction of soil nutrients and other properties. The models developed here perform well on fields used for calibration. Further research is needed however to better understand their generalization potential.