The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLII-4/W4
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-4/W4, 23–26, 2017
https://doi.org/10.5194/isprs-archives-XLII-4-W4-23-2017
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-4/W4, 23–26, 2017
https://doi.org/10.5194/isprs-archives-XLII-4-W4-23-2017

  26 Sep 2017

26 Sep 2017

STATISTICAL METHOD TO OVERCOME OVERFITTING ISSUE IN RATIONAL FUNCTION MODELS

S. H. Alizadeh Moghaddam1, M. Mokhtarzade1, A. Alizadeh Naeini2, and S. A. Alizadeh Moghaddam1 S. H. Alizadeh Moghaddam et al.
  • 1Faculty of Geodesy and Geomatics Engineering, Khaje Nasir Toosi University of Technology, Tehran, Iran
  • 2Dept. Geomatics, Faculty of Civil Engineering and Transportation, University of Isfahan, Isfahan, Iran

Keywords: Rational Function Models (RFMs), Overfitting, Statistical test, Regularization

Abstract. Rational function models (RFMs) are known as one of the most appealing models which are extensively applied in geometric correction of satellite images and map production. Overfitting is a common issue, in the case of terrain dependent RFMs, that degrades the accuracy of RFMs-derived geospatial products. This issue, resulting from the high number of RFMs’ parameters, leads to ill-posedness of the RFMs. To tackle this problem, in this study, a fast and robust statistical approach is proposed and compared to Tikhonov regularization (TR) method, as a frequently-used solution to RFMs’ overfitting. In the proposed method, a statistical test, namely, significance test is applied to search for the RFMs’ parameters that are resistant against overfitting issue. The performance of the proposed method was evaluated for two real data sets of Cartosat-1 satellite images. The obtained results demonstrate the efficiency of the proposed method in term of the achievable level of accuracy. This technique, indeed, shows an improvement of 50–80% over the TR.