Hyperspectral Image Classification Using Mixed Convolutions and Covariance Pooling

Fully Group Convolutional Neural Networks for Robust Spectral–Spatial Feature Learning

“…Here, we adopt another approach, where spectral-spatial features are extracted via 2D-CNNs. The representative methods of this kind [10,[40][41][42] typically apply first some form of dimensionality reduction like principal component analysis and use only several first principal components as input. This way, fewer learning parameters are needed and thus the computational cost is reduced.…”

“…This way, fewer learning parameters are needed and thus the computational cost is reduced. However, the spectral information is less well exploited [10,40] and the number of principal components is often inconsistent in different data sets [14,18,42,46]. We introduce here a novel multi-scale spectral feature extraction method based on 2D-CNN that requires no dimensionality reduction, making use of all spectral bands.…”

“…Our architecture is based on 2D-CNN. In contrast to most of the 2D-CNN methods, including [17,[40][41][42], which need to apply some form of dimensionality reduction (like PCA) prior to feature extraction, our learning architecture consumes the original raw 3D HSI data. We accomplish this through an efficient multiscale spectral feature extraction method.…”

Manifold-Based Nonlocal Second-Order Regularization for Hyperspectral Image Inpainting

“…yperspectral images are acquired with imaging spectrometer under continuous and narrow spectral bands, which results in rich details of object surface reflectance across quite a huge amount of bands ranging from the visible wavelength to the sub-infrared one. Therefore, HSIs have been widely studied and applied in the remote sensing domain [1,2,3], including urban layout, military surveillance, mineral exploration, and environmental monitoring, to just name a few. Nevertheless, the practical potential is usually compromised since the rich spectral information comes at the cost of greatly restricting the spatial resolutions.…”

Wide Sliding Window and Subsampling Network for Hyperspectral Image Classification

“…CNN was combined with other recently emerged methods to solve the HSI classification problem. The hybrid spectral CNN (HybridSN) [19] and mixed CNN [20] are spectral-spatial 3-D CNNs followed by spatial 2-D CNNs. Gong [21] proposed multiscale convolution (MS-CNNs) with diversified metric to obtain discriminative features.…”