Unsupervised geometrical feature learning from hyperspectral data

Hyperspectral technology has made significant advancements in the past two decades. Current sensors onboard airborne and space-borne platforms cover large areas of the Earth surface with unprecedented spectral resolutions. These characteristics enable a myriad of applications requiring fine identification of materials. Quite often, these applications rely on complicated methods of data analysis. In essence, the challenges include high dimensionality, spectral mixing, and atmospheric effects. This paper presents a robust unsupervised method to efficiently overcome this issue. The proposed algorithm performs three core tasks to acquire good results: i) optimizing the weights within a fixed threshold value for pure pixel estimation, ii) finding the best-averaged weighted endmember signatures with similarity error below the threshold value, and iii) iterating until a fixed number of average weighted endmembers is chosen. The experimental results on both real and synthetic data demonstrate that the proposed method is more robust and accurate then other geometrical methods.