TL;DR
This paper introduces a novel bundle block adjustment approach for airborne three-line array imagery that directly uses rotation angles (ω-φ-κ) to model and compensate for systematic errors in GPS/IMU data. This method improves geo-referencing accuracy and is particularly effective for large-scale mapping projects.
What Problem Does This Paper Address?
Airborne three-line array sensors are increasingly used for high-resolution Earth observation and mapping. While these sensors, coupled with GPS/IMU systems, provide rich data, the exterior orientation parameters (EOPs) derived from these systems often contain systematic errors. Direct geo-referencing using these raw EOPs is not accurate enough for surveying and mapping. Traditional methods require ground control points (GCPs) to refine EOPs, but this paper aims to improve the accuracy and automation of this process by developing better mathematical models for systematic error compensation. Specifically, the paper addresses the challenge of accurately modeling and eliminating IMU-derived systematic errors, which are often represented using quaternions or other angle systems that may not directly align with the IMU's native ω-φ-κ output.
Methods
The core contribution of this paper is the development of two bundle block adjustment models for airborne three-line array imagery:
- Systematic Error Compensation Model: This model directly incorporates the ω-φ-κ rotation angle system (roll, pitch, yaw) to represent the sensor's attitude. It explicitly models and compensates for systematic errors in GPS and IMU observations, including GPS antenna offsets, IMU primary axes misalignment, and strip-related constant and drift errors. This is achieved by introducing additional unknown parameters into the bundle adjustment.
- Orientation Image Model: This model uses a set of "orientation images" (key points in time with interpolated EOPs) to represent the sensor's trajectory. EOPs for individual scanning lines are then interpolated between these orientation images. The paper shows how this model can be integrated with the systematic error compensation approach.
A key novelty is the direct use of the ω-φ-κ rotation angles, which aligns with the standard output of IMU devices, making it more intuitive to model and eliminate systematic errors compared to methods relying on quaternions. The paper also proposes a strategy to resolve potential issues with yaw angle interpolation due to its periodic nature.
Datasets
The proposed models were evaluated using real datasets from three test areas:
- Taigu test area, China: Captured in May 2007, with 13 strips, ~30 km² area, 0.06 m GSD, and 600 m flying height. 105 GCPs were measured, with 4 used for adjustment and 101 for checking.
- Pingyao test area, China: Captured in October 2006, with 9 strips, ~240 km² area, 0.2 m GSD, and 2000 m flying height. 84 GCPs were measured, with 4 for adjustment and 80 for checking.
- Waldkirch test area, Switzerland: Captured in May 2002, with 7 strips (4 east-west, 3 north-south), ~80 km² area, 0.2 m GSD, and 2000 m flying height. 30 GCPs were measured, with 4 for adjustment and 26 for checking.
Additionally, a large-scale project in Henan Province, China, involving 41 ADS40 strips covering over 7,000 km² was used to demonstrate the scalability and robustness of the approach.
Key Insights and Ideas
- Direct ω-φ-κ modeling is advantageous: Using the native ω-φ-κ rotation angles from IMUs directly in bundle adjustment simplifies the modeling of systematic errors compared to methods that convert to or use quaternions. This is because the IMU's systematic errors are more naturally expressed within this angular system.
- Yaw angle interpolation can be tricky: The periodic nature of the yaw angle (κ) can lead to interpolation errors, especially during sharp turns. The paper proposes a practical method to ensure consistent sign conventions for yaw angles to resolve this.
- Combining models yields best results: The paper demonstrates that combining the systematic error compensation model with the orientation image model leads to the most accurate results, achieving significant improvements over using either model alone.
- Scalability for large datasets: The developed software and approach are capable of handling very large datasets (e.g., 41 strips, 300,000+ tie points) efficiently, which is crucial for modern photogrammetric projects.
Key Results
The proposed models, particularly the combined systematic error compensation and orientation image model, demonstrated significant improvements in accuracy:
- Taigu test area: RMSE for check points was smaller than 0.7 GSD in planimetry and less than 1.0 GSD in height.
- Pingyao test area: Similar improvements were observed, with RMSEs for check points generally below 1.0 GSD.
- Waldkirch test area: The orientation image model substantially reduced height residuals, with RMSEs decreasing to less than 0.75 GSD.
- Overall: The combined model showed 5-10% improvements compared to the orientation image model alone. The systematic errors of check points were significantly compensated, and residuals showed a more random distribution. The approach also proved effective for large-scale projects, handling extensive datasets with reasonable memory requirements.
Limitations and Future Work
- Limitations:
- While systematic errors are significantly reduced, some residual systematic errors may remain, particularly if they cannot be fully modeled by the chosen parameters.
- The accuracy of the results is still dependent on the quality of the input GPS/IMU data and the automatically matched tie points.
- The paper mentions that for the Henan project, precise measurement of image points under stereo view was not possible, which might have impacted height accuracy compared to other test fields.
- Future Work:
- The paper implicitly suggests that further refinement of systematic error modeling could lead to even higher accuracies.
- The successful handling of large datasets opens avenues for applying these methods to even more extensive mapping projects.
- Further investigation into optimizing the interpolation strategies for rotation angles could be beneficial.
Summary updated