Abstract: The fusion of airborne laser point clouds and multi-spectral images is of great importance for applications in remote sensing image processing, environmental monitoring and urban planning. Aiming at the low efficiency and robustness of the existing fusion methods, this paper proposes a non-rigid probabilistic matching-based deep fusion method for laser point clouds and multi-spectral images. Line-CNN deep learning network is used to extract straight line segment features, which are sampled as 2D scatter points, and non-rigid CPD algorithm is utilized to match discrete points in different scale images, and then the airborne laser point cloud with multi-spectral information is obtained by fusing two modal images with high accuracy through the improved single response matrix. To comprehensively validate the performance of the method, airborne laser point clouds and multi-spectral images from a variety of scenarios are used as experimental data in this paper. Experiments show that the accuracy of discrete point matching under multiple complex scenes is as high as 90%, the fused image can well retain the features and information of the original image, the fusion correlation coefficient under multiple scenes is as high as 90% or more, and the algorithm is more efficient, which is conducive to the subsequent applications such as vegetation monitoring, environmental monitoring, and land analysis.

Key words: deep learning, non-rigid CPD algorithms, single response matrix, high-precision fusion, cross-modal remote sensing data