Calibrating long-term AVHRR-derived NDVI imagery

Abstract

Several studies have confirmed that the radiometric calibration of AVHRR instruments changes following launch, and drifts over time. These changes cause spurious trends to appear in NDVI data that confound the reliable detection of vegetation change. It has previously been established that approximate additive corrections can be applied to an entire NDVI image, if the postlaunch calibration coefficients are known. However, the results of postlaunch calibration studies published in the literature vary, leading to different NDVI corrections depending on which coefficients are used. We acquired ten years (1981–1991) of monthly NDVI images covering Western Australia (Latitude 13°S to 36°S) in which the NDVI had been derived using prelaunch calibration coefficients. To address the calibration issue, we have computed theoretical corrections to the NDVI based on the prelaunch coefficients used to compute the NDVI in our data set, and postlaunch calibration coefficients reported by Che and Price (1992) and Kaufman and Holben (1993). Large differences were found in the NDVI corrections, depending on which set of postlaunch calibrations was adopted. To resolve the differences, we assumed that the entire image was a seasonally invariant target, and that the trend in the mean NDVI time series was solely due to changes in sensor calibration. The results obtained using this “time series” approach agreed with those computed in theoretical simulations using the calibration coefficients of Kaufman and Holben (1993). It is concluded that the time series technique described in this article can be used to remove sensor-induced trends from a time series of NDVI data without any knowledge of the postlaunch calibration coefficients. To apply the time series calibration technique, the dates of any changes in the sensor prelaunch calibration coefficients (including changes in the satellite platform) must be known. No other information is necessary, offering a significant advantage over existing methods for users of NDVI time series image data.