PE&RS July 2019 - page 510

mixture of woody and herbaceous vegetation, and the phase of
the harmonic suggested the timing of leaf-on (Moody and John-
son 2001). Although harmonics have been used to extract and
classify phenology information, to our knowledge, they have
not been used to separate phenological dynamics of woody
and herbaceous vegetation components in African savannas.
This paper aims to (1) develop a new approach based
on harmonic analysis to separate the deciduous woody and
herbaceous phenological dynamics in savanna area, and (2)
apply the method to map the spatial and temporal changes of
vegetation components in southern Africa. Deciduous woody
and herbaceous covers have different timings of green-up and
senescence in southern Africa (Archibald and Scholes 2007).
For example, woody vegetation flushes its leaves right before
the rainy season, while herbaceous vegetation does not grow
leaves until the rainy season starts (Higgins et al. 2011). The
phenological difference leads to distinct seasonal NDVI charac-
teristics for woody and herbaceous covers, which can be cap-
tured by frequency components in the harmonic analysis. Our
science questions are: (1) Can the new approach developed in
this study be used to characterize the compositional phenol-
ogy changes of savanna (using southern African savanna as a
case study)? (2) What are the spatial and temporal dynamics
of the phenology patterns of the woody and herbaceous com-
ponents in southern African savanna? (3) How does climatic
variation drive the phenological patterns?
Study Area
The study area covers portions of Angola, Zambia, Zimbabwe,
Namibia, and Botswana of southern Africa, lying between
10°52′23′′S and 21°44′44′′ S and between 15°59′24′′ E and
26°58′52′′E. The area covers approximately 1 450 000 km2
with a mean minimum and maximum temperature range
from 9°C to 27°C, and annual precipitation from 1500 to 291
mm (Hijmans et al. 2005). From north to south, the vegeta-
tion shifts from woodlands and open deciduous forests to
arid shrubland (Cowling et al. 2004). Three types of woody
canopy are found in the study area: 1) Miombois covered by
a dense canopy that is mostly broad leave evergreen but with
a short partially deciduous period (Fuller 1999; Chidumayo
2002; Richer 2008; Vancutsem et al. 2009); 2) Mopane is
mostly covered by broad leave deciduous, but more mixed in
density (Fuller 1999; Veenendaal et al. 2008); and 3) Kalahari
is mostly deciduous forest, varying from a heterogeneous mix-
ture of open woodland savannas (Burkeo-Pterocarpetea) to
shrublands (Acacietea, Strohbach and Petersen 2007; Gessner
et al. 2009). The study area consists of eight ecoregions (Table
1) defined by the Terrestrial Ecoregions of the World (Figure
1) (Olson et al. 2001). Nonsavanna ecoregions (Etosha Pan
halophytics, Western Zambezian grasslands, Zambezian Cryp-
tosepalum dry forests, Zambezian flooded grasslands, and
Zambezian halophytics) are excluded in this study. The land
use and grazing intensity across the study region is shown in
Figure 2a. The map shows that the the drier grassy savanna
types are grazed whilst the wetter denser woody savanas of
the Miombo woodlands are laregly ungrazed. The precipita-
tion that drives growth across the study region exhibits a very
strong latitudinal gradient declining markedly from north to
south (Figure 2b).
Material and Methods
The MODIS eight-day 500 m Bidirectional Reflectance Distribu-
tion Function (BRDF)-adjusted reflectance product (MCD43A4)
(Schaaf et al. 2002, 2010) from January 2002 to December
2011 was acquired for the study area. The NDVI time series
was derived from the red band (620–670 nm) and near-infra-
red band (841–876 nm). The MODIS observations often contain
noise due to clouds, cloud shadows, and other artificial sourc-
es, which creates false high or low values in the NDVI time
series (Lunetta et al. 2006; Shao et al. 2016). In this study, we
detected the noise using the quality assurance layer within
the MODIS reflectance product, which labeled it as “not good
quality”. Additionally, we detected the noises using a moving
window differencing filter, which calculated the difference
between the current observation and the observation at the
previous time step (Lu et al. 2003; Zhou et al. 2016; Hill et al.
2016). The detected noise was removed from the time series
and replaced by applying the spline interpolation (Zhou et al.
2016). We chose the 500 m MODIS dataset because our research
project included a separate study on separation of photosyn-
thetic vegetation, nonphotosynthetic vegetation, and bare soil,
which required short-wave infrared (SWIR) bands that were
not available in the finer resolution 250 m dataset.
Frequency Decomposition
We propose a frequency decomposition method to separate
the seasonal time series variation of a mixed pixel into woody
and herbaceous variation (Figure 3). Time series can be rep-
resented by a sequence of regular frequencies according to a
standard Fourier transform (Bracewell 1965):
X x e k N
k
n
N
n
i
N
kn
= ⋅ = … −
=
− − ∑
0
1 2
0 1
π
, , (1)
where N is the total sample size (e.g., N = 460 for the eight-
day composites from 2002 to 2011), k is the number of cycles.
k/N is the frequency (e.g., n = 10, 20, 30 represents once a
year, twice a year, and three times a year in this study), xn is
the nth value of time series, and Xk is a complex number that
represents the amplitude and phase at the given frequency. In
this study, we use a subset range of frequencies to represent
the seasonal vegetation variation.
Table 1. Characteristics of African savanna ecoregions.
Ecoregion Annual Rainfall (mm) Tree Genus and Species Grass Genus and Species
Angolan Miombo (ER1) 641–1472
Brachystegia; Isoberlinia;
Julbernardia spp.
Loudetia; Hyparrhenia;
Tristachya spp. Central Zambezian Miombo (ER2) 667–1503
Southern Miombo (ER3) 605–832
Zambezian & Mopane (ER4) 847–922 Colophospermum mopane;
Combretum; Acacia; Kirkia spp.
Aristida; Digitaria; Erarostis;
Echniochloa spp. Angolan Mopane (ER5) 310–730
Southern Africa Bushveld (ER6) 336–844 C. mopane; Acacias;
Terminalia spp. Burkea spp. Hyparrhenia spp.
Zambezian Baikiaea (ER7) 459–1027 Baikaiea plurijuga Sparse grasses
Kalahari Acacia-Baikiaea (ER8) 298–594
Lonchocarpus; Terminalia; Burkea;
Combretum; Grewia; Acacia;
Commiphora spp.
Aristida; Eragrostis;
Heteropogon;
Digitaria spp.
*Data source: Vegetation species: Burgess et al. (2004); Hill et al. (2011). Rainfall: Hijmans et al. (2005).
510 July 2019 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING