Rotation Equivariant Convolutional Neural Networks for Hyperspectral Image Classification

Detection of surface material based on hyperspectral imaging (HSI) analysis is an important and challenging task in remote sensing. It is widely known that spectral-spatial data exploitation performs better than traditional spectral pixel-wise procedures. Nowadays, convolutional neural networks (CNNs) have shown to be a powerful deep learning (DL) technique due their strong feature extraction ability. CNNs not only combine spectral-spatial information in a natural way, but have also shown to be able to learn translation-equivariant representations, i.e. a translation of input features into an equivalent internal CNN feature map. This provides great robustness to spatial feature locations. However, as far as we know, CNNs do not exhibit a natural way to exploit rotation equivariance, i.e. make use of the fact that data patches in a HSI data cube are observed in different orientations due to their orientation or on the varying paths/orbits of the airborne/spaceborne spectrometers. This article presents a rotation-equivariant CNN2D model for HSI analysis, where traditional convolution kernels have been replaced by circular harmonic filters (CHFs). The obtained results over three well-known HSI datasets showcase the potential of the approach.