Evolutionary Multi-task Ensemble Learning Model for Hyperspectral Image Classification

Recently, ensemble learning paradigm has shown great potential to achieve better prediction performance in the hyperspectral image classification. However, in the traditional methods, each classifier independently searches for the optimal spectral feature subspace in series and some important features are searched repeatedly, which leads to high computing redundancy and low effective utilization of features. In this article, an evolutionary multitask ensemble learning model (EMT_EL) for hyperspectral image classification is designed. First, the model formulates the spectral feature subspaces generation into a multitask optimization problem to concurrently search for optimal feature subspaces for multiple classifiers, which would be capable to select more informative and representative feature subspaces effectively. Second, seeking the optimal feature subspace for one base classifier can assist in the optima-seeking process for some other base classifiers via sharing the useful features, which can accelerate converge toward the direction of the optimal feature subspace, avoid trapping in local optimal subspace and improve searching capability. Third, randomization-enhanced genetic operators are designed for effective and reasonable feature selection, which can facilitate the exchange of information and improve the joint searching efficiency of the feature subspace. Eventually, the quality of generated spectral feature subspaces for each base classifier is improved and the feature sharing can parse HSI data by knowing which spectral features are important. Experimental results demonstrate that the proposed method can generate the appropriate feature subspace for each base classifier, thus it has outstanding classification performance on the different hyperspectral datasets.