PROJECT TITLE :

Composite Kernel of Mutual Learning on Mid-Level Features for Hyperspectral Image Classification

ABSTRACT:

The effectiveness of the algorithm for machine learning can be enhanced by training multiple models and then taking the average of the predictions generated by these models. It is anticipated that the performance optimization of multiple models will generalize further data in a satisfactory manner. To accomplish this, it is necessary to transfer generalization information between the different models. For the purpose of hyperspectral classification, a multiple kernel mutual learning method that is based on transfer learning of combined mid-level features has been proposed in this article. On the basis of the image produced by principal component analysis (PCA), which is utilized for the purpose of computing mid-level features, three-layer homogenous superpixels are computed. The sparse reconstructed feature, the combined mean feature, and the uniqueness feature are the three characteristics that make up the mid-level features. The sparse reconstruction feature is achieved through the utilization of a joint sparse representation model while adhering to the limitations imposed by the regions and boundaries of three-scale superpixels. The combined mean features are computed using the average values of spectra in multilayer superpixels, and the uniqueness is obtained by superposing the manifold ranking values of each of the multilayer superpixels. After that, three kernels of samples are computed for mutual learning in different feature spaces, with the goal of minimizing the divergence between them. After that, a combined kernel is built in order to optimize the sample distance measurement, and it is applied by using SVM training in order to construct classifiers. Experiments were run on real hyperspectral datasets, and the results showed that the proposed method can perform significantly better than several state-of-the-art competitive algorithms based on MKL and deep learning. This was demonstrated by the fact that the method outperformed the competitive algorithms by a significant margin.