Representation-based classification for hyperspectral imagery: An elastic net regularization approach

In hyerspectral image analysis, representation-based classification is a novel concept- A testing pixel is linearly represented by using the labeled samples. The weight coefficients can be solved by an ℓ₁-norm penalty for sparse representation, or solved by an ℓ₂-norm penalty for collaborative representation. In this work, a convex combination of these two representations using the elastic net model is proposed. The objective is to produce more robust weight coefficients for better classification performance. Experimental results on two hyperspectral data prove that our proposed method outperforms the traditional sparse representation-based classifier and collaborative representation-based classifier.