The development and validation of the prediction models for soil properties using VIS/NIR spectra.

Abstract

In recent years, precision agriculture has been widely applied to farming practices in order to improve their productivity, and to reduce their environmental impacts. One of the important challenges in the precision agriculture is to determine the soil properties of an agricultural field accurately, and yet rapidly. Such evaluation is essential to executing well-timed, appropriate actions, which could improve farming efficiency. To this end, we have developed statistical models that are capable of predicting soil properties of a field, such as moisture, nitrogen, or carbon contents. These models are constructed from visible/near-infrared spectra data obtained by a real-time soil spectrometer. In this presentation we discuss a few approaches and methods to improve the predictive power of such models. Support vector regression (SVR) method allows us to obtain reasonably predictive models without explicit variable selection or wavelength selection models. Ensemble learning procedure could significantly improve the predictivity of partial least square (PLS) models. Furthermore the applicability domain of a model can be estimated through application of bootstrap technique in calculation of standard deviation of each predicted values.