Classification of hyperspectral images using machine learning methods

Mixed pixels problem has significant effects on the application of remote sensing images. Spectral unmixing analysis has been extensively used to solve mixed pixels in hyperspectral images. This is based on the knowledge of a set of unidentified endmembers. This study used pixel purity index to extract endmembers from hyperspectral dataset of Washington DC mall. Generalized reduced gradient (GRG) a mathematical optimization method is used to estimate fractional abundances (FA) in the dataset. WEKA data mining tool is chosen to develop ensemble and non-ensemble classifiers using the set of the FA. Random forest (RF) and bagging represent ensemble methods while neural networks and C4.5 represent non-ensemble models for land cover classification (LCC). Experimental comparison between the classifiers shows that RF outperforms all other classifiers. The study resolves the problem associated with LCC by using GRG algorithm with supervised classifiers to improve overall classification accuracy. The accuracy comparison of the learners is important for decision makers in order to consider tradeoffs in accuracy and complexity of methods.