8-day interval) time series of Landsat type imagery (Gao et al, 
2006, Roy et al, 2008). On a regional level these data contribute 
to a better ecosystem understanding and an improved estimation 
of carbon fluxes for vegetation communities (Linderholm et al, 
2006, Schmidt et al, 2012). 
In this contribution we link a regular spaced (8-day interval) 
time series of MODIS-Landsat fused imagery via the STARFM 
algorithm (Gao et al, 2006) for a period of 7.5 years using 322 
observations of NDVI imagery and rainfall surfaces. The data 
are first analysed classically with a lagged correlation and 
subsequently with a more complex distributed lag model 
including trend and noise removal. The objective of this 
contribution is a) to investigate if the synthetic high 
spatiotemporal NDVI time series and rainfall time series exhibit 
a significant correlation in a test region in an Australian 
Savanna and; b) investigate the influence of the seasonality of 
the correlation. 
2. DATA AND METHODS 
2.1 Data 
Regional setting 
A 12 km x 10 km sample region in a typical Australian northern 
savanna region was chosen. The area includes homogeneous 
woody forests vegetation, grasslands and heterogeneous areas 
with a mixture of surface covers, such as a palustrine wetland 
and riparian vegetation. Regional ecosystem (RE) data of 
Queensland are generally mapped at 1:100,000 scale 
(http://www.derm.qld.gov.au/REDATA). The major forested 
communities in the test region are shown in Figure 1 and are 
mapped by RE data as a) low open-woodland to occasionally 
low open-forest of Eucalyptus shirleyi (silver-leaved ironbark), 
b) semi-evergreen vine thicket with many codominant species 
on young igneous rock, Woodland to open-woodland of 
Eucalyptus platyphylla (poplar gum), Corymbia clarksoniana 
(Clarkson's bloodwood), Corymbia tessellaris (Moreton Bay 
ash) and Eucalyptus tereticornis (bluegum). The northern part 
of the subset is part of the Great Basalt Wall national park and 
has undergone very little change in the recent history (e.g. no 
fire history). 
— 
     
i 
à 
Figure 1. Study area in a Landsat true color composite, 
superimposed is a subset of the regional ecosystem 
classification. 
Satellite imagery 
Satellite imagery of the Landsat and MODIS sensors were 
utilised to create a 8-day regularly interval time series, see 
Schmidt et al (2012) for details. 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
  
Start date Stop date Num. of obs. Interval 
MODIS 02/2000 09/2007 322 8-day 
Landsat 12/1999 09/2007 90 irregular 
STARFM 02/2000 09/2007 322 8-day 
Rainfall 02/2000 09/2007 322 8-day 
  
Table 1. Time span and resolution of the raster data used. 
NDVI data were extracted from this time series of STARFM 
generated Landsat type imagery as a representation for 
vegetation greenness. 
Rainfall data 
Rainfall data were extracted form the SILO interpolated 
surfaces (Jeffery, et al., 2001) as 8-day rainfall totals. 
Foliage Projective Cover 
Foliage Projective Cover (FPC) data were extracted for the 
study region. FPC is defined as the horizontal percentage cover 
of photosynthetic foliage of all strata and provides a 
biophysically meaningful description of vegetation cover, 
particularly for Australian vegetation communities (Armston, et 
al. 2009). 
  
Figure 2. FPC as a measure of overstorey foliage cover in the 
study area. 
2.2 Methods 
A lagged correlation analysis was performed on a per pixel 
basis of z-transfomed NDVI and rainfall data for the common 
time series from 02/2000-09/2007. This type of analysis is 
commonly applied, but leads in problems with significance 
assessment if two autocorrelated time-series are regressed are 
regressed against each other. In this case the ordinary least 
square (OLS) estimator results in a autocorrelated residual 
structure, which constitutes a severe violation of the assumption 
of OLS-regression, since the risk of a type I error in the 
significance assessment of (lagged) regression coefficients is 
increased. 
Thus, in a second step of the analysis a distributed lag model 
has been applied, in which OLS-estimator was substituted by a 
generalized least square (GLS) parameter estimation. To this 
end were first long term trends and seasonal components were 
eliminated from both datasets and only seasonal anomalies were 
retained. Than a regression analysis in form of a distributed lag 
model (Udelhoven et al, 2009) is applied to the anomalies as 
follows using simple OLS-regression for parameter estimation: