Urban Feature Extraction within a Complex Urban Area with an Improved 3D-CNN Using Airborne Hyperspectral Data

Abstract

Airborne hyperspectral data has high spectral-spatial information. However, how to mine and use this information effectively is still a great challenge. Recently, a three-dimensional convolutional neural network (3D-CNN) provides a new effective way of hyperspectral classification. However, its capability of data mining in complex urban areas, especially in cloud shadow areas has not been validated. Therefore, a 3D-1D-CNN model was proposed for feature extraction in complex urban with hyperspectral images affected by cloud shadows. Firstly, spectral composition parameters, vegetation index, and texture characteristics were extracted from hyperspectral data. Secondly, the parameters were fused and segmented into many S × S × B patches which would be input into a 3D-CNN classifier for feature extraction in complex urban areas. Thirdly, Support Vector Machine (SVM), Random Forest (RF),1D-CNN, 3D-CNN, and 3D-2D-CNN classifiers were also carried out for comparison. Finally, a confusion matrix and Kappa coefficient were calculated for accuracy assessment. The overall accuracy of the proposed 3D-1D-CNN is 96.32%, which is 23.96%, 11.02%, 5.22%, and 0.42%, much higher than that of SVM, RF, 1D-CNN, or 3D-CNN, respectively. The results indicated that 3D-1D-CNN could mine spatial-spectral information from hyperspectral data effectively, especially that of grass and highway in cloud shadow areas with missing spectral information. In the future, 3D-1D-CNN could also be used for the extraction of urban green spaces.

Keywords

• airborne hyperspectral image
• deep learning
• 3D-CNN
• pixel-based
• 3D-1D-CNN