Growing Correspondence Seeds for Efficient and Accurate Satellite DSM Extraction

Extracting digital surface model (DSM) from high-resolution satellite images remains a challenge in remote sensing and photogrammetry. In this article, a precise and efficient method for DSM extraction from satellite images, called SATellite-growing correspondence seeds (Sat-GCS), is proposed. The proposed method consists of the following six stages: 1) extracting tie-points using features from accelerated segment test detector, dense adaptive self-correlation dense descriptor, and local keypoint correspondence, 2) rational polynomial coefficients bias compensation, 3) epipolar image rectification, 4) dense matching using the GCS algorithm, 5) three-dimensional triangulation to generate ground point clouds, and 6) height interpolation of the point clouds to produce DSM. The main property of the Sat-GCS method is the generation of a set of precise and dense tie-points in stage (2), used as initial correspondence seeds to improve the GCS algorithm in stage (4). The proposed method is evaluated on 12 different datasets from four different satellite sensors including ZY3-01, CartoSat-1, ZY3-02, and Worldview-3, and the results are compared with the CATALYST, SAT-MVSF, and SS-DSM methods. The DSM extraction results show the superiority of the proposed method in terms of completeness, root-mean-square error (RMSE), and MEE compared to other methods.