A Geospatial Transformer-based Network in Engineering Surveying and Mapping

Nowadays, engineering surveying and mapping involved in collecting and analyzing spatial data to support infrastructure development as well as environmental planning. Thus, the demand of high-precision and Deep Learning (DL) for automated solutions in terrain analysis has grown significantly. However, the existing DL models struggled to generalize across various geospatial environments. Hence, this research proposed a generalizable framework namely Geospatial Transformer-based Network (GeoTransNet) for semantic understanding and 3-Dimensional (3D) reconstruction of complex terrains. Initially, the input data is collected by using Unmanned Aerial Vehicle (UAV) mounted cameras as well as Light Detection And Ranging (LiDAR) sensors. Then, the data is preprocessed with Bilateral Filtering (BF) to improve spatial clarity by removing noise while preserving critical edge features. After that, Vision Transformer (ViT) is used to extract features by learning global contextual relationships from image patches for superior terrain-invariant feature learning. Then, Transformer-based Semantic Segmentation (SegFormer) is employed to provide precise pixel-wise classification of landscape elements. Finally, the Multi-View Attention Fusion (MVAF)-based module combines segmentation outputs with LiDAR point clouds to reconstruct high-resolution 3D terrain models. From the results, the proposed GeoTransNet achieved superior results in overall accuracy (97.31%) when compared to existing Boundary Enhancement and Multiscale Refinement Fusion Network (BEMRF-Net) model.