Beyond single receptive field: A receptive field fusion-and-stratification network for airborne laser scanning point cloud classification

The classification of airborne laser scanning (ALS) point clouds is a critical task remote sensing and photogrammetry fields. Although recent deep learning-based methods have achieved satisfactory performance, they ignored the unicity receptive field, which makes ALS cloud remain challenging for distinguishment areas with complex structures extreme scale variations. In this article, objective configuring multi-receptive field features, we propose novel fusion-and-stratification network (RFFS-Net). With dilated graph convolution (DGConv) its extension annular (ADConv) as basic building blocks, fusion process implemented (DAGFusion) module, obtains feature representation through capturing graphs various regions. stratification fields sets different resolutions calculation bases performed Multi-level Decoders nested in RFFS-Net driven by multi-level aggregation loss (MRFALoss) to drive learn direction supervision labels resolutions. fusion-and-stratification, more adaptable regions variations large-scale clouds. Evaluated on ISPRS Vaihingen 3D dataset, our significantly outperforms baseline approach 5.3% mF1 5.4% mIoU, accomplishing an overall accuracy 82.1%, 71.6%, mIoU 58.2%. Furthermore, experiments LASDU dataset 2019 IEEE-GRSS Data Fusion Contest show that achieves new state-of-the-art performance.