Sensitivity of the BFAST algorithm to MODIS satellite and vegetation index

Abstract

The Breaks for Additive Seasonal and Trend (BFAST) algorithm combines the additive decomposition of time series with abrupt change detection. It potentially allows for the effective use of the temporal detail available in satellite image time series for examining vegetation response patterns across regional extents, accounting for variation at the seasonal scale while detecting changes in the long term trends. While BFAST has been validated using NDVI time series (Verbesselt et al. 2010a), its sensitivity to different parameters and input data has not yet been assessed. Understanding the effects of the data source and type and the variation of the algorithm's user-defined parameters on the timing and number of abrupt changes it detects will allow it to be used more effectively. This study aims to assess the effects of the satellite used for data collection, the vegetation index (EVI or NDVI) and varying the value of a BFAST argument called the h parameter, which controls the potential number of trend breaks detected, on abrupt change detection for a study area in the Paroo region of north-western New South Wales, Australia. Moderate Resolution Imaging Spectroradiometer (MODIS) EVI and NDVI time series were decomposed for 165 sample points chosen to include a range of land cover types in the study region. The effect of the MODIS satellite (Aqua or Terra) used to collect data was assessed by comparing the number of breaks detected and their timing between time series derived from each satellite. The effect of changing the h parameter was assessed by comparing the similarity in the length of time periods with a significant trend component slope between different values of h (1/3, 1/5 and 1/7). The timing of detected breaks was affected by the satellite used to collect data, despite the visual similarity of Aqua and Terra time series. Greater certainty in the timing of breaks was achieved when using smaller values for the h parameter. Of the three factors tested, the vegetation index had the greatest impact on the timing and number of breaks in the long term trend detected by BFAST. The effect of the vegetation index was dependent upon the h parameter used, and the effects of both the h parameter and the satellite varied between EVI and NDVI. These results suggest a moderate sensitivity of the BFAST algorithm to all three of these factors, and also an interaction between them. This should be taken into consideration when using BFAST for long term vegetation change detection.