Multi-scale Neighborhood Information Fusion Network for Classification of Remote Sensing LiDAR Images

In the field of remote sensing data processing, light detection and ranging (LiDAR) derived Digital Surface Model (DSM), which is capable of effectively reflecting elevation information, has significant value in earth observation applications. In this work, we proposed a multi-scale neighborhood information fusion (MNIF) network for classification of remote sensing LiDAR images. Specifically, capture the details and overall characteristics of the data through multi-scale patches simultaneously, with the information from different scales complementing one another. Subsequently, a spatial-aware region (SR) attention module is utilized to emphasize distinctive features specific to each category. Furthermore, since the input feature maps from different scales encompass distinct neighborhood information, the cross-scale convolutional kernels are incorporated into the designed optimized feature extractor (OFE). Experimental results based on two DSM datasets named Houston and Trento in the 2013 GRSS Data Fusion Competition demonstrate that the proposed MNIF network can achieve superior classification performance compared with several existing methods.