An interpretable reliability-aware view selection framework for satellite-based deep MVS reconstruction

Reliable view selection is essential for accurate deep multi-view stereo (MVS) reconstruction from satellite imagery. However, existing strategies mainly rely on acquisition metadata or simplified geometric criteria and therefore cannot adequately characterize the heterogeneous reliability of satellite views under complex geometric, radiometric, and temporal conditions. This paper proposes a reliability-aware view selection framework that evaluates candidate source views by jointly considering geometric strength, radiometric consistency, RPC-derived geometric stability, and temporal reliability. Unlike conventional angle-based geometric criteria, the proposed framework characterizes the spatial variation in RPC-derived epipolar orientations across the image domain, thereby providing complementary geometric information for identifying more reliable multi-view configurations. Experiments on the US3D benchmark show that the proposed framework improves reconstruction performance in terms of MAE and PAG1 relative to conventional metadata-driven and simulation-based selection strategies, reducing MAE by up to 9.9% and increasing the Percentage of Accurate Grids at 1 m (PAG1) by up to 8.5%. RMSE exhibits region-dependent behavior, decreasing in JAX but increasing in OMA. Additional evaluations on the MVS3D benchmark and cross-backbone experiments using two satellite MVS architectures further support the transferability of the proposed framework across the evaluated urban environments and reconstruction architectures.