Unmanned Aerial Vehicle Hyperspectral Remote Sensing Image Classification Based on Parallel Convolutional Neural Network

Unmanned aerial vehicle (UAV) hyperspectral images (HSI) are widely used in remote sensing fine-grained monitoring and classification owing to the abundant spatial- spectral information. Deep learning (DL) for HSI classification is a popular method because of the strong information extraction ability. The existing DL methods mainly include pixel-level convolutional neural network (CNN) methods and semantic segmentation algorithms. However, the HSI of UAV has rich spatial-spectral information, and it also shows the characteristics of severe shadow, poor signal-to-noise ratio and redundant spectral information. Pixel-level CNN shows a large amount of salt and pepper noise in classification results for these HSI, while the semantic segmentation algorithm shows inaccurate boundaries and small fragments, and requires a large number of training samples. Therefore, to address this problem, we hereby propose a parallel neural network model based on ICA-2D-CNN branch and 3D-CNN branch for UAV HSI classification. Among them, the ICA-2D-CNN branch extracts HSI spatial features using ICA and 2D-CNN, and the 3D-CNN branch is designed to extract spatial and spectral information of HSI to alleviate the classification problem of salt noise and boundary inaccuracy. Taking the 5cm spatial resolution of UAV HSI as an example, the overall accuracy (OA) of the proposed model is 71.87%, the average accuracy reaches 72.9%, and the Kappa coefficient is 0.639. Compared with 1D-CNN, 2D-CNN and 3D-CNN, OA increased by 3.6%, 9.3% and 6.5%, respectively. The parallel neural network proposed by this paper can achieve satisfactory classification results, which showing robust classification performance and significant advantages for UAV HSI classification.