Spectral Feature Extraction and Classification of Soil Types Using EO-1 Hyperion and Field Spectroradiometer Data Based on PCA and SVM

Abstract

This paper reports application of Hyperspectral Remote Sensing (HRS) datasets to the soil taxonomy in Phulambri Taluka of Aurangabad district of Maharashtra, India. The preprocessing of imaging HRS dataset were carried out in three steps, the first removal of noisy and unwanted bands, second conversion of radiance value to reflectance value and finally atmospheric correction through Quick Atmospheric Correction (QUAC) algorithm. Principal Component Analysis (PCA) algorithm was implemented to reduce the dimensionality of huge Hyperion data. First three PCs were valuable to preserve 98% of the variance creation. The soil spectra of 74 samples obtained from Analytical Spectral Device (ASD) non-imaging spectroradiometer which was used as input reference spectra for imaging Hyperion data for soil feature extraction, classification of surface soil type’s and its mapping. Gaussian Radial Basis Function (RBF) kernel of Support Vector Machine (SVM) classifier with very less training pixels was computed after dimensionality reduction of data. The overall accuracy of SVM classifier was 92.76% with kappa value 0.90. The identified soil types were black cotton soil, lateritic soil, and sand dunes. The results are significant for soil analysis and its mapping of the complex region.