Evaluation of AVHRR-to-MODIS cross-sensor translation using Landsat Tm for long-term monitoring of terrestrial environment

Abstract

Improving our understanding of Earth's terrestrial environmental changes requires longterm data records constructed from multiple satellite sensors. Recently, a multi-sensor normalized difference vegetation index (NDVI) translation algorithm has been developed for the Land Long-Term Data Record (LTDR) NDVI data, and used to converted a historic LTDR Advanced Very High Resolution Radiometer (AVHRR/2) NDVI (1981--1999) dataset into a LTDR Moderate Resolution Imaging Spectroradiometer (MODIS) (2000--present) compatible NDVI. The objectives of this study were to develop a methodology to evaluate the continuity of nonoverlapping NDVI products and to assess the performance of the LTDR translation equation. This study utilized a 23 year time-series consisting of 20,776 LTDR/TM paired scenes from 14 test sites across the North American continent representing 10 distinct land cover classes. While the Chow test showed significant differences between AVHRR and MODIS data both before and after translation, the translation of AVHRR data resulted in an overall increase in compatibility between the AVHRR and MODIS NDVI data. Mean and median differences between AVHRR vs. TM difference and MODIS vs. TM difference decreased from 0.053 to 0.0064 NDVI units and from 0.04 to 0.0056 NDVI units, respectively. Differences between translated AVHRR NDVI and TM NDVI exhibited a seasonal dependency also observed in those between MODIS and TM NDVI. Paired LTDR/TM observations showed that after translation AVHRR and MODIS both had a slope of 1.05 and 1.06 respectively with scene pairs consistently above the 1:1 line while original AVHRR/TM observations had a slope of 0.8449, crossing the below the 1:1 line at NDVI values greater than 0.3. Finally, while a multiple linear regression model of the combined LTDR dataset showed that sensor type still had a significant influence on slope, the effect of sensor type on the model slope was estimated to be 0.8%. This study has shown that while the translated AVHRR data and the MODIS data are still statistically different, the translation resulted in a great reduction in the difference (an increased compatibility) between the two data. This study has also demonstrated the utility of Landsat TM in evaluating intersensor compatibility of NDVI data sets which lack a common observation period.