Hyperspectral remote sensing image classification based on dense residual three-dimensional convolutional neural network

Data-driven deep learning techniques set the current state of the art in image classification for hyperspectral remote sensing images. The lack of labeled training data and high dimensionality of hyperspectral images, results in these techniques being far from satisfactory with respect to accuracy and efficiency. To address the deficiencies of the existing approaches, we proposed a novel neural network technique, namely, dense residual three-dimensional convolutional neural network (DR-3D-CNN). Tailored for hyperspectral images, this network used 3D convolution instead of the conventional 2D convolution for more effective spectral feature extraction. It also employed dense residual connections to alleviate the problem of gradient dispersion. After the initial classification by the network, the proposed technique further refined the result using multi-label conditional random field optimization. Experimental results on various hyperspectral image datasets showed that the proposed model outperforms existing deep learning techniques with respect to accuracy by a large margin while requiring less training time.

Automated summary

The study introduces a novel technique based on a dense residual three-dimensional convolutional neural network to address the issues of accuracy and efficiency in hyperspectral image classification. Experimental results show that the proposed technique significantly outperforms existing deep learning techniques in accuracy and training time.