Inter-comparison of four models for smoothing satellite sensor time-series data to estimate vegetation phenology

Abstract

Several models have been fitted in the past to smooth time-series vegetation index data from different satellite sensors to estimate vegetation phenological parameters. However, differences between the models and fine tuning of model parameters lead to potential differences, uncertainty and bias between the results amongst users. The current research assessed four techniques: Fourier analysis, asymmetric Gaussian model, double logistic model and the Whittaker filter for smoothing multi-temporal satellite sensor observations with the ultimate purpose of deriving an appropriate annual vegetation growth cycle and estimating phenological parameters reliably. The research used Level 3 Medium Resolution Imaging Spectrometer (MERIS, spatial resolution ~4.6km) Terrestrial Chlorophyll Index (MTCI) data over the years 2004 to 2006 composited at eight day intervals covering the Indian sub-continent. First, the four models were fitted to representative sample time-series of the major vegetation types in India, and the quality of the fit was analysed. Second, the effect of noise on model fitting was analysed by adding Gaussian noise to a standard profile. Finally, the four models were fitted to the whole study area to characterise variation in the quality of model fitting as a function of single and double vegetation seasons. These smoothed data were used to estimate the onset of greenness (OG), a major phenological parameter. The models were evaluated using the root mean square error (RMSE), Akaike Information Criteria (AIC), and Bayesian Information Criteria (BIC). The first test (fitting to representative sample time series) revealed the consistently superior performance of the Whittaker and Fourier approaches in most cases. The second test (fitting after the addition of Gaussian noise) revealed the superior performance of the double logistic and Fourier approaches. Finally, when the approaches were applied to the whole study, thus, including vegetation with different phenological profiles and multiple growing seasons (third test), it was found that it was necessary to tune each of the models according to the number of annual growing seasons to produce reliable fits. The double logistic and asymmetric Gaussian models did not perform well for areas with more than one growing season per year. The mean absolute deviation in OG derived from these models was a maximum (3 to 4weeks) within the dry deciduous vegetation type and minimum (1week) in evergreen vegetation. All techniques yielded consistent results over the south-western and north-eastern regions of India characterised by tropical climate.