A physical explanation of the variation in threshold for delineating terrestrial water surfaces from multi-temporal images: effects of radiometric correction

Abstract

Multi-temporal satellite images are widely used to delineate objects of interest for monitoring surface changes. Threshold value(s) are often determined from a histogram of a delineation index. However, the threshold determined may vary and be case-dependent, with images taken at different times. Although the variation is well known, its cause remains unclear, and this raises doubts about the reliability of the classification results. This study selects three widely used indices, the near-infrared (NIR) band, the normalized difference vegetation index (NDVI) and the normalized difference water index (NDWI), all of which can be used to delineate water surfaces. Our theoretical analysis reveals that sensor calibration, the Sun–target–satellite geometry and the atmospheric optical properties create synthetic effects on the satellite's digital number (DN) and, subsequently, on the thresholds for delineation. The DN-based threshold has a significant dependence on the reflectance-based counterpart, which has been proved with multi-temporal Moderate Resolution Imaging Spectroradiometer (MODIS) data for the Poyang Lake region of China. Our results show that a DN-based threshold is generally higher than a reflectance-based one, and ∼90% of the difference is accounted for by temporal influences. A quantification of the temporal influences provides a physical explanation to the variation in thresholds, and the findings should be valuable for improving the reliability of long-term studies using multi-temporal images.

Acknowledgements

This work was supported by a 973 Programme of National Basic Research Programme of China (2012CB417003), a Programme of National Science Foundation of China (No. 41171272) and the 100-talent Project of The Chinese Academy of Sciences (CAS). MODIS data were obtained from the Warehouse Inventory Search Tool, USA. Landsat TM/ETM + data were obtained from Landsat.org, Global Observatory for Ecosystem Services, Michigan State University, USA. ASTER was obtained from the Earth Remote Sensing Data Analysis Centre (ERSDAC), Japan. The anonymous reviewers are highly appreciated for their constructive comments.