Abstract
<jats:p>Three-dimensional reconstruction of Earth’s surface from multi-view satellite and aerial imagery is critical for applications in terrain modeling, disaster assessment, and environmental monitoring. However, existing methods face two key challenges: achieving high-fidelity reconstruction at large scales and ensuring the security and authenticity of geospatial 3D models. In this paper, we propose TerraGS, a unified framework that integrates multi-scale 3D Gaussian Splatting with built-in provenance embedding and tamper detection. The multi-scale reconstruction module adaptively allocates Gaussian primitives based on terrain complexity, achieving state-of-the-art rendering quality on benchmark datasets. The provenance embedding module encodes geospatial metadata directly into Gaussian parameters via a security-aware optimization strategy, preserving high message recovery accuracy with minimal degradation to rendering fidelity. The hybrid tamper detection module employs dual watermarking to localize unauthorized modifications and verify model ownership. Extensive experiments on the DFC2019 and ISPRS Vaihingen datasets demonstrate that TerraGS achieves superior reconstruction quality (33.5 dB PSNR), high-capacity provenance embedding (4.5 bits/GS at</jats:p>