TL;DR
This paper presents a method for extracting Digital Surface Models (DSMs) from Gaofen-6 (GF-6) satellite cross-track images, which have a wide field of view. The authors demonstrate that by carefully selecting image pairs with sufficient overlap (20-40%) and employing an error-compensated rational function model (RFM) combined with semi-global block matching (SGBM), accurate DSMs can be generated, retaining detailed topographic features.
What Problem Does This Paper Address?
Generating accurate Digital Surface Models (DSMs) is crucial for various applications like urban planning, natural resource management, and disaster analysis. While satellite stereo imagery offers broad coverage, extracting high-resolution DSMs from specific sensor types, like the Gaofen-6 (GF-6) satellite's Wide Field of View (WFV) cross-track cameras, presents challenges. These challenges include ensuring sufficient intersection angles between images for accurate stereo matching, dealing with potential mismatches due to varying imaging conditions, and achieving high elevation accuracy. This paper aims to address these issues by proposing and validating a method for DSM extraction from GF-6 WFV cross-track imagery.
Methods
The proposed DSM extraction method involves several key steps:
- Stereo Image Orientation: An error-compensated Rational Function Model (RFM) is used for precise orientation of the stereo images. This approach aims to correct systematic errors introduced by satellite position and sensor attitude, improving geometric accuracy.
- Epipolar Image Generation: A projection trajectory method is employed to generate epipolar images. This step rectifies the stereo images so that corresponding points lie on epipolar lines, reducing the search space for matching and improving efficiency and accuracy.
- Dense Image Matching: The Semi-Global Block Matching (SGBM) algorithm is utilized to generate a disparity map. SGBM is an extension of the Semi-Global Matching (SGM) algorithm that incorporates block matching and pixel-cost aggregation to enhance efficiency and accuracy, particularly in challenging regions.
- DSM Generation: Finally, the generated disparity map is used to compute three-dimensional coordinates through forward intersection, creating a point cloud. This point cloud is then filtered and rasterized to produce the final DSM.
Datasets
The study utilizes Gaofen-6 (GF-6) satellite panchromatic cross-track images with approximately 30% overlap, acquired over Taian, Shandong Province, China. For accuracy evaluation, the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model (GDEM) with a 30m spatial resolution served as the reference DEM. Additionally, a Digital Orthophoto Map (DOM) with a 2m spatial resolution was used for adjustment control points and checkpoints.
Key Insights and Ideas
- Overlap is Crucial: The study highlights that for GF-6 cross-track images, a specific overlap range (20-40%) is critical for achieving a good balance between intersection accuracy and data utilization. Insufficient overlap leads to poor elevation accuracy, while excessive overlap might not be necessary or efficient.
- WFV Cameras for DSM: The research demonstrates that even though GF-6 WFV cameras are not primarily designed for stereo mapping, they can still produce valuable DSMs when appropriate processing techniques are applied.
- SGBM's Effectiveness: The SGBM algorithm proves effective in generating disparity maps from the GF-6 cross-track imagery, even in complex terrain, outperforming commercial software (ENVI) in terms of smoothness and noise reduction.
- Terrain-Dependent Accuracy: The accuracy of the extracted DSM varies with terrain type, with valleys showing the highest accuracy and mountain peak areas exhibiting lower accuracy due to challenges like light and shadow effects causing mismatches.
Key Results
- The proposed method successfully extracted DSMs from GF-6 WFV HR cross-track images, retaining detailed topographic features, including mountainous terrain, houses, and roads.
- The spatial resolution of the extracted DSM is 10m, with elevation ranging from 37.3m to 1582.8m.
- The Root Mean Square Error (RMSE) of the extracted DSM was found to be:
- Valley areas: 6.303 m
- Ridge areas: 12.879 m
- Urban areas: 14.929 m
- Peak areas: 19.043 m
- The DSM generated by the proposed method was visually closer to the reference DEM and smoother than the DSM generated by ENVI, which was noted to be rough and noisy.
Limitations and Future Work
- Intersection Angle Limitation: The paper acknowledges that the intersection angle of GF-6 cross-track images inherently limits elevation accuracy.
- Future Work: The authors suggest incorporating external data, such as laser points from ICESat-2, to further improve DSM accuracy by providing additional elevation constraints. This could help mitigate issues in challenging areas like mountain shadows.
Summary updated