Deep Graph Network for Multispectral Point Cloud Classification with Adaptive Multi-kernel Graph

As a multidimensional form of remote sensing data, multispectral point clouds provide the data basis for more refined land cover classification. But the challenge remains in modelling and representing the rich spatial-spectral information. With better performance on non-Euclidean data, graph neural networks have been used on multispectral point clouds. However, most of the current studies use a fixed graph structure, which means that the feature aggregation pattern is limited. Thus, this paper proposes a novel Multi-Kernel deep graph network with Initial residual and Identity mapping (MKII). The multi-kernel mechanism is used to obtain adaptive graph structure and the deep graph neural network is designed to better adapt to the complex aggregation patterns introduced by changing graph structure. Specifically, we measure the similarity between each point by multiple base kernels and form base graphs. Then the feature aggregation of different patterns is achieved by learning combinations of base graphs. And a deep graph neural network is used to implement the classification of multispectral point clouds, with initial residual and identity mapping introduced to overcome the over-smoothing of deep features. Experimental results demonstrate that the proposed MKII outperforms several state-of-the-art methods on a real multispectral point cloud.