Optimization based manifold embedding for hyperspectral image classification and visualization

Remote sensing and interpretation of hyperspectral images are becoming an increasingly important field of research. High dimensional hyperspectral images consist of hundreds of bands and reflect the properties of different materials. The need for more detail about objects and the improvement of sensor resolutions have resulted in the generation of higher size hyperspectral data. Many years of research have shown that there are many difficulties in the pre-processing of these data due to their high dimensionality. Recent studies have revealed that manifold learning techniques are a very important solution to this problem. However, as the complexity of the data increases, the performance of these methods cannot reach a sufficient level. This letter proposes a particle swarm-based multidimensional field embedding method inspired by the force field formulation to increase the performance. Detailed comparative analyses of the proposed method were made for Botswana and Kennedy Space Center (KSC) data. It is also shared in the results of other popular datasets. Experimental results show that the proposed method is superior to existing manifold embedding methods in classification accuracy and visualization of hyperspectral data. In addition, an optimization-based solution is presented to the problem of parameter determination of existing embedding methods.