A three layer spatial-spectral hyperspectral image classification model using guided median filters

Hyperspectral images (HSI) contain rich spectral information from a large portion of the electromagnetic spectrum. Using these images, it is possible to make pixel-level classification as each pixel holds hundreds of features. In this paper, we propose an efficient, three-layer hyperspectral image classification model by utilizing spectral/spatial features. The first layer of the system includes two classifiers that work in parallel. These classifiers generate probability scores that form the "new feature set" of the original dataset. The second layer is an ensemble classifier that combines the new features to generate the initial region classification. The third layer introduces a novel approach for enhancing the initial region classification's accuracy from the second layer by utilizing the spatial characteristics of the dataset. A new proximity-based 2D edge preserving order-statistic filtering called Guided Median Filter (GMF) is introduced with weights assigned to each neighboring pixel. Experimental results show that the proposed system improves our previously published results and reaches over 96% overall accuracy on Indian Pines dataset by exceeding some well-known traditional classifiers. Moreover, our GMF based system produced comparable results with the state-of-the-art neural network based methodologies without complex training stage and lack of interpretability of classification model.