Abstract
Aiming at the problems of complex RPC geometric modeling in multi-view stereo matching (MVS) of satellite remote-sensing images, significant feature noise caused by cross-view occlusion and inconsistent lighting, and blurred edge details of height maps, this paper proposes a satellite MVS height estimation network based on a Mamba-enhanced feature pyramid and deformable RPC sampling, MambaDPSMVS. In the feature extraction stage, the cross-view Mamba module and the single-view self-Mamba enhancement module are designed to realize explicit feature interaction between and within views through direction prediction, direction routing, geometric gating fusion and state-space sequence modeling, and suppress view noise from the source. A deformable sampler for an RPC geometry, RPC-DPSampler, is proposed to achieve accurate alignment of source-view features with the reference view through 3D deformable coarse alignment, block sampling, 2D sub-pixel fine-tuning and view-level attention-weighted aggregation. Then, the cascaded cost is constructed, and the loop encoder–decoder regularization is used to complete the height estimation from coarse to fine. Experiments on the WHU-TLC and US3D datasets show that the MAE/RMSE of the proposed method reaches 1.688 m/2.867 m and 1.744 m/3.847 m, respectively, and the PAG<1 m reaches 61.81% and 62.18%, respectively, which values are better than those of baseline methods such as Sat (Red-Net) and Sat (CasMVS-Net). The ablation experiment further validated the complementarity of Mamba enhancement features and the RPC-DPSampler module in feature expression and geometric alignment.