An Adaptive Pseudo Sample Generation Approach for Unsupervised Multi-Class Change Detection in Hyperspectral Images

This paper presents a novel adaptive pseudo-sample generation approach for unsupervised multi-class change detection in multi-temporal multispectral/hyperspectral remote sensing images. The approach is developed based on the adaptive spectral change vector representation, and the original spectral difference image is first projected into a 2D polar domain. The polar domain is automatically divided into multiple pseudo-sample selection regions by expectation maximum algorithm based on the statistical distributions and characteristics of each class of change and no-change samples. Pseudo-samples are generated by selecting certain proportions of samples in the candidate region, which are used as input to the supervised change detector. Experimental results obtained on real multi-temporal hyperspectral dataset demonstrate the effectiveness of the proposed approach in addressing an unsupervised multi-class change detection task.