Semi-supervised Classification of Land Cover in Multi-spectral Images Using Spectral Slopes

We propose a spectral-slope based technique to classify land-cover using spectral angle mapper (SAM) classifier, which is a semi-supervised process of labeling the land surface. Initial reference samples for SAM are obtained from spectral-slopes based rules, which are tailored to classify multispectral images obtained by Indian Space Research Organization`s (ISRO) Linear Imaging Self-Scanning Sensor 3 (LISS-III) and Advanced Wide Field Sensor (AWiFS). All the pixels in the are classified by using these reference samples, which are subsequently used in a SAM algorithm. The rules for selecting the reference samples from the multi-spectral imageries are defined by using the properties of spectral-profiles. The land-cover is classified into five broad classes, namely, water, grass land, built-up, vegetation, and bare land. The classification results are validated using very high resolution (VHR) satellite images.