Surface Depth Estimation from Remote Sensing Images Based on Cross-Level Feature Representation

Remote sensing image surface depth estimation aims to construct 3D structure of geographic environment through 2D remote sensing images, which is a major topic in Computer Vision. Surface depth estimation providing 3D geospatial data for geohazard monitoring, digital twin and etc. While deep learning has become mainstream approach in surface depth estimation, surface features reconstruction remains the major challanges, which leads to the insufficient accuracy, inadequate multi-scale feature fusion, and high time consuming. To address the mentioned problems, this paper proposes a surface depth estimation model RT-SatMVSF which innovatively employes RMT network and Manhattan Attention Mechanism to imporve the capability of surface features reconstruction in weak texture regions. RT-SatMVSF achieves a reduction of 1.06% and 3.53% in MAE and RMSE on WHU-TLC dataset. Experimental results show that this proposed method effectively improves the surface features reconstruction perfomance in weak texture region.