Comparison of Unsupervised Classification and Principal Component Analysis (PCA) Using Hyperspectral Remote Sensing Data for Classifying Mollisols

The purpose of this research is to differentiate Mollisol soils into discrete spectral regions using hyperspectral remote sensing data. The hyperspectral remote sensing data consisted of 120 bands in the visible and near infrared region (457-823 nm) with spectral and spatial resolutions of 3 nm and 1 m respectively. Fifty classes with nine iterations of the raw data were selected for unsupervised classification. These fifty classes were merged into three groups based on their spectral signatures and compared with the first order soil survey which consisted of three soil series (Drummer silty clay loam, Flanagan silt loam, and Elburn silt loam) for the study area. Ten principal components were selected for principal component analysis. The first three principal components which had 99.21% of the information were merged and an unsupervised classification was performed on the merged first three principal components similar to the unsupervised classification for the raw datatset. Separabilty analysis (SA) was also performed in both studies to distinguish between the three soils series. Results from SA indicated Drummer and Flanagan to be more distinct as compared with Elburn (SA value 54 for unsupervised classification of the raw dataset and 48 for the unsupervised classification of the first three merged PC bands). The accuracy assessment (100, 150, and 250 random sample points) compared with the first order soil survey for both procedures yielded higher overall accuracy for unsupervised classification of the dataset than the unsupervised classification of the first three principal component bands. Additional research involved establishing statistical spectral relationships between soil texture classes, selected soil nutrients to support the current study. The classifications developed in this study were found to be more detailed than the ones usually found in soil surveys, which facilitates their use for various potential applications requiring a higher level of detail.