The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B5-2021
https://doi.org/10.5194/isprs-archives-XLIII-B5-2021-51-2021
https://doi.org/10.5194/isprs-archives-XLIII-B5-2021-51-2021
30 Jun 2021
 | 30 Jun 2021

REGRESSION ANALYSIS OF ERRORS OF SAR-BASED DEMS AND CONTROLLING FACTORS

Y. Y. Wu and H. Ren

Keywords: InSAR, digital elevation model, water vapor variation, surface deformation, linear regression analysis

Abstract. Interferometric Synthetic Aperture Radar (InSAR) has been well developed for several decades and is known for its powerful capability of retrieving three-dimensional ground information from SAR imagery. One of the most important application of InSAR technique is topographic mapping. The technique is limited when confronting certain poor conditions which lead to low coherence. In this research, we aim at investigating the relationship between SAR-based digital elevation models (DEMs) and related factors that contribute to the error budget by conducting a linear regression analysis. The surface deformation in line of sight (LOS) direction and the amount of integral refractivity change over two acquisition events are considered as two related factors. Eight pairs of Sentinel-1 images were selected to conduct InSAR processing over Chaiyi City of Taiwan, and SNAP software was used to generate SAR-based DEMs. The coherence mask was applied during the InSAR workflow in order to alleviate unwrapping error. The result has shown that the coherence thresholds help to improve the accuracy by up to 52.61%. Since some large errors were observed from the resulting InSAR-DEMs, these points were removed based on standard error. In regression analysis, there were 15 set of data, categorized by different coherence threshold and data removal standard, to test the model. As the result has shown, when the coherence threshold is 0.3 and the points were filtered with half standard error, the R2 can achieve 0.85. However, the rest of the dataset did not produce desirable results. In our discussion, we have provided several reasons which might have contributed to this outcome.