A Novel Bias Correction Method for Replacing Ground Control Points in the Extracted DSM from Cartosat-1 Stereo Images

In order to reduce the effect of the systematic error (bias) in the RPC data in the generated DEM (relative model) from the satellite stereo images, an existing elevation model as an absolute model can replace the need for ground control points. In this research, a DEM matching strategy was introduced based on the development of the slope-based approach.
Unlike other existing DEM matching methods that first apply a projection system, the proposed mathematical model, based on the coordinate system of the input data was developed. In this way, the parameters of the transformation were obtained in this coordinate system. As a result, it is possible to directly improve the RPC data of the stereo images, which are also expressed in the same coordinate system.
Inspiring from the classical absolute orientation of the aerial images, the two-stage transformation was carried out separately. In order to evaluate the proposed method, a Cartosat-1 image pair and an SRTM model were provided from a mountainous region. This method compared with the original slope-based approach and provided a better approximation of the three-dimensional offset values between the relative and absolute models.
The generation of the relative DEM (MATDEM) has been implemented in the MATLAB environment. In this research, it was considered a dense image matching method for generating the relative DEM. Therefore, an area-based solution has been used to reach the desired density of the results.
The least squares image matching (LSM) method, as an area-based method, has the potential to achieve high precision and is mainly used as an alternative to increasing the accuracy of other matching methods. However, the LSM method has a low convergence radius and it is possible to fall into the local minima of the correlation function, which in turn reduces the reliability of the results. Therefore, it is important to produce the proper seed points which are located within the small convergence radius of this method. Here, these seed points were obtained using raw RPC data and the SRTM model. In this regard, a regular grid was assumed on the first image, and after the extraction of the seed points, the precise matching was performed using this method. Also, for comparison, a relative model (PCIDEM) was produced using the PCI-Geomatica software.
The most important achievement was to discover the actual values of the bias of the raw RPCs in the MATDEM case. It was assumed that the systematic errors were propagated from the RPC data in the generated relative DEMs. The estimated values for offset parameters, particularly the offset in the longitude direction, were different for PCIDEM and MATDEM. According to the evaluations, the values obtained from the MATDEM have been more accurate. The reference data for this assessment was the offset calculated using ground control points. In the evaluations using ground control points, the offset values estimated by the proposed method along latitude and longitude directions were 0.77 and 1.23 m, respectively. With regard to the pixel size of Cartosat-1 images, the planimetric offset value was estimated as 0.58 pixels.