Deep Cube-Pair Network for Hyperspectral Imagery Classification

Code

Abstract: Hyperspectral image classification benefits from models that exploit 3D spatial-spectral structure, but limited labeled samples make deep models difficult to train reliably. This work proposes a cube-pair-based CNN framework that augments training samples while preserving local 3D HSI structure. Within this framework, a 3D fully convolutional network named DCPN is designed with fewer parameters than conventional CNNs, allowing deeper modeling under the same amount of training data and improving generalization. Experiments on several HSI datasets show stronger classification performance than competing methods.