Evaluation of long-term NDVI time series derived from Landsat data through blending with MODIS data

The aim of this study is to capitalize on the spatial detail of Landsat and the temporal regularity of MODIS acquisitions using a fusion approach (Spatial and Temporal Adaptive Reflectance Fusion Model, STARFM). Specifically, the 30 m Landsat-7 ETM+ (Enhanced Thematic Mapper plus) surface reflectance was predicted for a period of eight years (2002-2009) as the product of observed ETM+ and MODIS surface reflectance (MOD09Q1) on the predicted and observed ETM+ dates. A pixel based analysis for observed ETM+ dates covering winter and summer crop seasons showed that the prediction method was more accurate for NI R (mean r(2)= 0.87, p <= 0.01) compared to red band (mean r(2)= 0.65; p <= 0.01). The NDVI was computed from observed Landsat and predicted surface reflectance. The difference between NDVI from predicted and observed ETM+ data (prediction residual) was compared with the temporal residuals of NDVI from observed Landsat and MODIS data at two different dates. The prediction residuals for NDVI (spatial mean value of 0.0085) were found to be significantly lower than the temporal residuals (spatial mean value of 0.056 for MODIS and 0.051 for observed ETM+) implying that the prediction method was better than temporal pixel substitution. Investigating the trend in synthetic ETM+ NDVI values over a growing season revealed that phenological patterns were well captured. A direct comparison between the NDVI values obtained from MODIS and synthetic ETM+ images has shown a good consistency of the temporal dynamics but a systematic error that can be read as bias (MODIS NDVI over estimation). The relationship between synthetic ETM+ NDVI with observed precipitation and evaporation data was also studied and it was observed that monthly total precipitation and monthly evaporation of the preceding month have higher correlation coefficients (r(2)= 0.56 and r(2)= 0.59) with mean monthly synthetic ETM+ NDVI.