Application of unsupervised classification in NIR spectroscopy of soil

Abstract

Assessment of soil quality requires expensive and time-consuming measurements in the laboratory and the field. Many repetitions of the measurements are required to deal with high soil variability. As a result, scientists have been unable to measure and monitor soil quality and soil degradation over large areas. Sensing Soil Quality is a technological approach for rapid assessment and large area surveillance of soil condition. The technology is based on rapid screening of soil quality using a portable reflectance spectrometer. The resulting data set requires reliable statistical methods that can aid to fast extraction of information. Several clustering algorthim procedures has been used to find the optimal classes in the NIR spectroscopy data set for global soils. However, there were no significant soil classes obtained for the 697 topsoil samples. Therefore, other methods with potential of handling large dimensional data need to be adopted. For, example performing direct calibration of the data set or using both physical and chemical soil properties data. We are also recommending where possible to integrate clustering results with advanced graphical representation for example by using minimum spanning tree- MST.