There is increasing interest among the user communities for using satellite data products from multiple sensors for improved environmental monitoring. Spectral vegetation indices (VIs) are one of the more important products in observing spatial and temporal variations of vegetation biophysical properties and photosynthetic activities as well as in biogeochemical cycle modeling. To accomplish this goal, VIs from multiple sensors need to be normalized for differences in sensor characteristics and algorithms. In this study, we evaluated several empirical strategies in cross-calibrating VIs from different sensors for the spectral band pass filter differences. A satellite-borne hyperspectral image was obtained with the Earth Observing-1 (EO-1) Hyperion sensor over a tropical forest-savanna transitional area in South America. The image was first spectrally convolved to simulate Advanced Very High Resolution Radiometer (AVHRR), Moderate Resolution Imaging Spectroradiometer (MODIS), and Landsat-7 Enhanced Thematic Mapper Plus (ETM+) band passes and corrected for atmospheric effects. Data were then extracted from six land cover types with a wide range of biogeophysical conditions and used to empirically derive cross-calibration (translation) equations for the Normalized Difference Vegetation Index (NDVI). The empirical strategies examined included: cross-calibration at the VI level using the NDVI as a predictor variable, cross-calibration at the reflectance level using the reflectance as a predictor variable, and cross-calibration at the reflectance level using the NDVI as a predictor variable. We also examined a two-steps approach in which the cross-calibrations were performed first at the reflectance level and then at the NDVI level. Overall, all of the cross-calibration methods performed well, resulting in root mean square errors less than .05 NDVI units. In nearly all the cases, however, the translations resulted in large residual bias errors with their values reaching .16 units for dark, little or non-vegetated land targets. Depending on cross-calibration methods used, both the magnitudes and directions of bias errors varied significantly. Although the NDVI-based cross-calibration of the NDVI produced the best results with small RMSE values (< .01 unit), there still existed small bias errors. These results indicate that data continuity studies require a theoretical basis in developing a mechanistic understanding of discontinuity and that cross-calibration results need to be evaluated from a real application point of view in order to assess the impact of persistent bias errors and to establish acceptable difference, or error levels in multi-sensor data sets.
|