Accurate understanding of the transportation environment is fundamental for enabling mobility. To support this need, image-based surveying techniques such as photogrammetry and remote sensing play a critical role in capturing detailed spatial information. In this study, we focus on the technical advancement of image data processing for such applications. Transformer-based neural networks have significantly advanced the field of computer vision, including aerial image analysis. However, their increasing computational and memory demands pose practical challenges, especially in large-scale photogrammetric workflows. This research explores the impact of transformer network compression on aerial photogrammetry and semantic segmentation performance. Using the ISPRS-EuroSDR Benchmark dataset, we apply various compression methods via the CompressAI framework to high-resolution aerial images. Standard photogrammetric pipelines are then used to generate 3D point clouds from both original and compressed image sets, allowing us to evaluate geometric reconstruction fidelity. Additionally, transformer models are trained to perform semantic segmentation across three land cover classes, assessing the influence of compression on classification accuracy. Key evaluation metrics—such as PSNR, SSIM, IoU, and point cloud completeness and accuracy—are used to quantify the trade-offs. The results provide practical insights into balancing compression efficiency with data quality, and inform the deployment of transformer models in bandwidth-constrained or resource-limited aerial mapping scenarios.