The HybridSN is spectral-spatial 3D-CNN followed by spatial 2D-CNN. The 3D-CNN facilitates the joint spatial-spectral feature representation from a stack of spectral bands. The 2D-CNN on top of the 3D-CNN further learns more abstract level spatial representation.
Fig: Proposed HybridSpectralNet (HybridSN) Model with 3D and 2D convolutions for hyperspectral image (HSI) classification.
Fig.2 The IN dataset classification result (Overall Accuracy 99.81%) of Hybrid-SN using 30% samples for training. (a) False color image. (b) Ground truth labels. (c) Classification map. (d) Class legend.
Fig.3 The UP dataset classification result (Overall Accuracy 99.99%) of Hybrid-SN using 30% samples for training. (a) False color image. (b) Ground truth labels. (c) Classification map. (d) Class legend.
Fig.4 The UP dataset classification result (Overall Accuracy 100%) of Hybrid-SN using 30% samples for training. (a) False color image. (b) Ground truth labels. (c) Classification map.