A Multi-Source Convolutional Neural Network for Lidar Bathymetry Data Classification
Yiqiang Zhao, Xuemin Yu, Bin Hu, Rui Chen
Abstract
Airborne Lidar bathymetry (ALB) has been widely applied in coastal hydrological research due to outstanding advantages in integrated sea-land mapping. This study aims to investigate the classification capability of convolutional neural networks (CNN) for land echoes, shallow water echoes and deep water echoes in multichannel ALB systems. First, the raw data and the response function after deconvolution were input into the network via different channels. The proposed multi-source CNN (MS-CNN) was designed with a one-dimensional (1 D) squeeze-and-excitation module (SEM) and a calibrated reference module (CRM). The classification results were then output by the SoftMax layer. Finally, the accuracy of MS-CNN was validated on the test sets of land, shallow water and deep water. The results show that more than 99.5% have been correctly classified. Besides, it has suggested the best robustness of the proposed MS-CNN compared with other advanced classification algorithms. The results indicate that CNN is a promising candidate for the classification of Lidar bathymetry data.