Comparison of data and methods to best estimate starting of season dates across RENECOFOR forest plots (France) based on MODIS imagery

Researches based on ground phenological dataset emphasizes a new challenge to the remote sensing (RS) scientific community. In fact, despite the role of RS in monitoring large area's vegetation dynamics has been widely recognized in the last decades (Reed et al. 2003), generate precise estimates of phenological parameters from satellite data is still an important issue to RS researchers. In fact, according to BrA¼gger et al. (2003), phenological events occurrence within stands spanned by a few days, approximately between 7 and 14. Moreover, the forest's average blooming anticipation trend due to the climate change was measured by Nordli et al. (2008) and it was ranging between 0.2 and 5.1 days per decade. Similar trends were found in several works reviewed in Richardson et al. (2013) The aim of this work is to find the best combination of: a) vegetation index (VI), b) satellite data, c) fitting method and d) phenological parameters' estimation methods to accurately assess starting of season (SOS) and end of season (EOS) dates over not-disturbed temperate forests. The sample areas we considered are 52 plots of the REseau National de suivi A  long terme des ECOsystA¨mes FORestiers (RENECOFOR) French phenological network covered by coetaneous, homogeneous, broad-leaf stands. Its 1992 onwards weekly phenological measurements are a precious dataset to validate RS estimations of phenological metrics' occurrence. To fulfil our objective, we are implementing two vegetation indices, the Normalized Difference Vegetation Index (NDVI) and the Wide Dynamic Range Vegetation Index (WDRVI) at different levels of correction. Time-series (TS) of the above-cited VIs are generated based on both MODIS daily (MYD09GQ/MYD09GA) and 16-day composite (MYD13Q1) reflectances. We are basing our research on data from MODIS Aqua because of the problem related with MODIS Terra sensor's degradation (Wang et al. 2012). The scientific community developed a number of methods to reduce VIs TS' noise. The methods we are implementing and testing are the following: double-logistic (DL, Beck et al. 2006; Hmimina et al. 2013; Soudani et al. 2008; Zhang et al. 2003) and gaussian (GA, Jonsson and Eklundh 2002, 2004) fitting function, Savitzky-Golay filtering (SG, Chen et al. 2004; Jonsson and Eklundh 2002, 2004), fast Fourier transform (FFT, Jakubauskas et al. 2001) and principal component analysis (PCA). We will estimate phenological metrics' timing based on two methods: i) fixed threshold of VI increase above the annual winter minimum and ii) dynamic estimation based on the first derivative and the rate of curvature of the smoothed TS. Finally, we will obtain a date of SOS and EOS for each year in the VIs TS for each combination of VI, data, noise-removing algorithm and estimation method and we will compare such result with RENECOR's ground measurements. Our comparison will allow to find the combination of data, VI and TS treatment leading to the phenological metrics' estimation best matching with ground data. In addition, we will quantify the loss of accuracy involved by the use of composite data respective to daily data. We retain this point very important since it would eventually justify the use of the composite data for operations requiring fast responses and easiness of use.