Data Mining Algorithms for Prediction of Soil Organic Matter and Clay Based on Vis-NIR Spectroscopy

Abstract

Organic matter (OM) amount and clay content in the soil are important constituents in the sustainability of agricultural systems. The methods used for OM and clay analyses in laboratories are laborious, time consuming and use require reagents that pollute the environment. The use of reflectance in the visible and near infrared (Vis-NIR) can be highly viable in soil analysis identifying the attributes contents in a cleaner and quicker way. There is still no general model specifying the wavelengths to be used for neither each variable being analyzed nor a well-defined methodology to be applied. The aim of this study was to apply all the classification algorithms available in the Weka software trying to find the best correlations between spectral data in the Vis and NIR spectrums, separately, and OM and clay content in the soil. As result, the clay prediction had a strong correlation with both Vis and NIR spectrum. OM prediction presented a determination coefficient greater than 0.7 but brought an error that cannot be overlooked. Lazy KStar algorithm showed to be more adequate to mine the data presenting the higher determination coefficients and the lower errors. The best results for both OM and clay were obtained when correlated with the Vis spectrum. This suggests that it is possible to predict OM and clay using only the Vis spectrum.