Ensemble of multiple CNN classifiers for HSI classification with Superpixel Smoothing

Abstract

Hyperspectral Image analysis has gained much attention due to the presence of rich spectral information. Hyperspectral Image (HSI) classification is being utilized in a wide range of applications. Convolutional Neural Networks (CNN) are popularly used in the image classification tasks due to their capability of extracting spatial features from the raw image data. Creating an ensemble of multiple classifiers generates more robust and reliable classification results. In this paper, we propose an ensemble of four CNN classifiers with superpixel smoothing for the task of HSI classification. Stacked Auto-encoder is utilized to reduce the dimensionality of the hyperspectral data. A new method is suggested to derive the optimal number of features by exploiting the diversity among the classifiers. The uniform Local Binary Patterns (ULBP) are extracted from the HSI and is used along with reduced HSI data for classification. The two single-channel models take reduced HSI cubes as input. The two dual-Channel CNN models explore both ULBP patterns and HSI data simultaneously. We explore various techniques for combining the predictions of individual classifiers and choose the best one for ensembling purpose. The obtained prediction map is made to undergo superpixel based smoothing to remove most of the misclassified pixels. Experimental results on standard data sets confirm the superiority of the proposed ensemble model over the state of the art models. The advantages of superpixel smoothing after CNN classifications are also validated through numerical results and corresponding classification maps.