Classifying hyperspectral remotely sensed images is difficult due to the scarcity of annotated examples. Self-training is a semi-supervised process to develop a classifier by providing a limited number of labelled and a vast collection of samples that have not been labelled. However, the basic self-training method is prone to incorporate misclassified examples into the training set. A new semi-supervised approach has been presented in this study which produces correctly classified samples of diverse nature. The proposed approach uses spectral and spatial structural information of the hyperspectral images to generate representative samples for the target classes. The structural information also provides support in the final classification phase of hyperspectral images. The proposed method has been found to be both effective and efficient by the experimental results using well-known data sets as a reference.