Clay Content Prediction Using on-the-Go Proximal Soil Sensor Fusion.

There is a growing demand for high quality and reliable data on different soil constituents and properties in different scales. Research in the past decade has shown that traditional soil sampling cannot supply for this demand. Modern methods such as visible-near infrared (VNIR) and mid-infrared (MIR) in the lab have already shown to be rapid, reliable, robust and low cost substitutes for traditional methods.

Previous researchers have tried to use methods such as VNIR spectroscopy for on-the-go modeling of several soil constituents. However, in spite of being a highly influential parameter on soil usability, very few studies so far have provided robust and accurate predictions for fields with high clay content variability.

An on-the-go multi-sensor platform was used to measure topsoil (25cm) VNIR spectra and temperature as well as electrical conductivity of top 30cm and top 90cm in 5 fields in different regions in Denmark. 125 calibration samples were collected from the points found by clustering the principal components (PC) of the spectra. Several pretreatments such as mean-centering, auto-scaling, spectral transformations and removal of faulty measurements were performed on the data. Partial least squares regression (PLSR) and support vector machines regression (SVMR) were performed using VNIR spectra, EC and soil temperature as predictors and clay content as the response variable. PLSR and SVMR models were validated using full and 20-segment cross-validation respectively.

The results were highly accurate with R² of 0.94 and 0.94, root mean square error (RMSE) of 0.99 and 0.95, and ratio of performance to interquartile range (RPIQ) of 5.56 and 5.79 for PLSR and SVMR respectively. This shows the high potential of on-the-go soil sensor fusion to predict soil clay content and automate the mapping process.