A Branching Convolutional Encoder-based Spatio-spectral Dimensionality Reduction Model for Hyperspectral Image Classification and Change Detection

Abstract

Hyperspectral images are characterized by having high spectral resolution allowing the representation of a reflectance of a given scene in hundreds of bands. This property however transformative it might be in different applications; the curse of dimensionality is also inevitably present when trying to extract information and analyze hyperspectral data. Several dimensionality reduction techniques have been used to reduce the dimension of hyperspectral images such as principal component analysis, and numerous autoencoder variants. More recently, deep learning architecture-based reduction schemes have been devised for tacking the curse of dimensionality in hyperspectral images. Nonetheless, these architectures are not cognizant of both the spatial and spectral information available in hyperspectral images when reduction the original dimension. In this thesis, a branching convolutional encoder-based spatio-spectral hyperspectral image dimensionality reduction technique called “BCE (BCE)” is proposed. The branching architecture consists of a pointwise separable convolution to extract spectral features, and a two-dimensional convolution network to filter spectral feature. Later, these two features are fused and fed into a decoder network which attempts to reconstruct the original image as accurately as possible. This network is trained in a similar fashion to autoencoders, using a loss function to track the similarity between the original and the reconstructed image. Classification and change detection are important applications of hyperspectral images. The branching convolutional encoder is used together with two classification and change detection models to demonstrate its feature representation performance – the raw image has redundant features and poor interclass separability. The performance of the proposed dimensionality reduction model is compared with a spatial convolutional encoder and a densely-connected encoder. The L1 loss and L2 losses were down to about two digits after decimal point which is lower than the other two methods. Classification accuracy reaches over 90% on all the datasets which outperforms the other comparative methods. Moreover, the branching encoder’s representation power is observed with the change detection model; as the rate of accuracy reaches over 99% which for the Hermiston City data. This research demonstrably presents the success of a branching convolutional dimensionality encoder for classification and change detection applications