<p>Two-dimensional building maps are essential representations of spatial information, which are critical for understanding the built environment and urban planning. Since cities are dynamic and changing rapidly, traditional methods for constructing 2D building maps such as manual measurements, remote sensing techniques, and LiDAR-based techniques face challenges related to high financial costs and human resource consumption. Recently, global navigation satellite system (GNSS)-based city mapping has garnered attention due to its advantages, including wide coverage, high update efficiency, and low cost. However, current research on GNSS mapping primarily relies on signal power as a singular feature, which leads to suboptimal mapping performance. Moreover, existing studies lack a quantitative evaluation of 2D mapping accuracy. This article proposes a method for constructing 2D building maps using GNSS signals collected from smartphones and introduces precision evaluation metrics commonly employed in professional cartography to quantitatively assess the mapping outcomes. Supervised machine learning method is employed to classify GNSS line-of-sight (LOS)/non-line-of-sight (NLOS) signals based on multiple GNSS signal features extracted from smartphone raw data. The 3D environment is decomposed into voxels, and ray tracing is applied to determine the intersection of GNSS signals with each voxel. The probability of voxel occupancy is computed, for generating a 2D pixel map by projecting the voxels onto the horizontal plane. Building footprints are extracted from the 2D pixel map using an edge detection method. Experimental results show that the proposed 2D building map construction method based on GNSS LOS/NLOS signal classification, effectively recognizes common rectangular buildings, achieving an F1-score of 88.5% and an intersection over union of 80.2%. Additionally, the method demonstrates the capability to detect irregular building footprints which are more common in practical environments.</p>