PE&RS February 2018 Full - page 65

“Blend-then-Index” or “Index-then-Blend”:
A Theoretical Analysis for Generating High-
resolution NDVI Time Series by STARFM
Xuehong Chen, Meng Liu, Xiaolin Zhu, Jin Chen, Yanfei Zhong, and Xin Cao
Abstract
There are two strategies for generating high-resolution vegeta-
tion index time series using spatiotemporal data blending
methods, named as “Blend-then-Index” (
BI
) and “Index-
then-Blend” (
IB
), according to the order of vegetation index
calculation and data blending. This study aims to determine
which strategy can obtain better results for generating a high-
resolution normalized difference vegetation index (
NDVI
) time
series using the spatial and temporal adaptive reflectance
fusion model (
STARFM
). The theoretical error analysis sug-
gests that the more accurate strategy depends on the vegeta-
tion growth stages:
BI
has a smaller error than
IB
when the
NDVI
values at the prediction date are higher than the input
NDVI
values and vice versa. Simulated experiments using
Landsat images were conducted to verify the theoretical
analysis. This study provides guidelines for producing better
high-resolution vegetation index time series using
STARFM
.
Introduction
Time-series vegetation index (
VI
) data derived from satellite
images are indispensable for monitoring land surface dy-
namics, such as phenology, ecosystem production, and land
cover changes (Kim
et al
., 2014; Shen
et al
., 2015; Wardlow
et al
., 2007; Zhang
et al
., 2003; Zhang
et al
., 2013; Zhong
et al
., 2012). However, owing to technological and budget
limitations, there is trade-off between spatial resolution and
acquisition frequency of satellite image data. Consequently,
the available satellite
VI
time-series datasets are acquired with
frequent observations (e.g., daily) but coarse spatial resolution
(250 to 8,000 m), such as Moderate Resolution Imaging Spec-
troradiometer (
MODIS
) and Advanced Very High Resolution
Radiometer (
AVHRR
)
VI
products. The low spatial resolution
limits their application to heterogeneous areas.
Recently, blending techniques that combine the frequent
coverage data with coarse spatial resolution (e.g.,
MODIS
)
and the long revisit cycle data with fine spatial resolution
(e.g., Landsat) have been proposed for generating synthetic
data with both high spatial resolution and temporal resolu-
tion (Chen
et al
., 2015a), such as the spatial and temporal
adaptive reflectance fusion model (
STARFM
) (Gao
et al
., 2006),
the enhanced
STARFM
(
ESTARFM
) (Zhu
et al
., 2010), the spatial
and temporal adaptive algorithm for mapping reflectance
change (STAARCH) (Hilker
et al
., 2009), the spatial temporal
data fusion approach (
STDFA
) (Wu
et al
., 2012), the sparse-
representation-based spatiotemporal reflectance fusion model
(
SPSTFM
) (Huang
et al
., 2012), the linear mixing growth model
(
LMGM
) for normalized different vegetation index (
NDVI
) (Rao
et al
., 2015), and the flexible spatiotemporal data fusion
(
FSDAF
) method (Zhu
et al
., 2016). These techniques provide
promising solutions to generate
VI
time-series data with high
spatial resolution. Based on the order of the
VI
calculations
and data blending, there are two blending strategies for
VI
data: “Blend-then-Index” (
BI
) and “Index-then-Blend” (
IB
)
(Jarihani et al., 2014; Tian et al., 2013). The
BI
method first
blends the reflectance of individual
MODIS
and Landsat bands
and then calculates
VI
from the blended data. In contrast, the
IB
method first calculates the
VI
from the original
MODIS
and
Landsat reflectance data and then blends the
VI
using blend-
ing methods. Therefore, the computation cost of
IB
is less than
that of
BI
because
IB
blends only one band. On the other hand,
the
BI
method seems more reasonable because the blending
algorithm is originally designed for reflectance instead of the
VI
. Both these strategies have been adopted in practice (Bhan-
dari
et al
., 2012; Walker
et al
., 2012; Chen
et al
., 2015b; Dong
et al
., 2016). However, it is unclear which strategy produces
better results, which is critical to choose between
BI
and
IB
to blend the
VI
data. Several investigations have dealt with
this problem. Tian
et al
. (2013) evaluated
STARFM
by simu-
lating a time series of 12
NDVI
images on the Loess Plateau
and concluded that
IB
performs better than
BI
. Jarihani
et al
.
(2014) used three sites with nine indices and both
STARFM
and
ESTARFM
and demonstrated that
IB
is more suitable for index
blending. However, the abovementioned conclusions were
based on experiments from a small number of sites; therefore,
their applicability is questionable. A comprehensive study
with solid theoretical analysis on this topic is necessary.
Herein, we compare the
BI
and
IB
strategies for vegetation
pixels by considering the error propagations in both the strat-
egies.
NDVI
, the most widely used
VI
, and
STARFM
, the first and
most commonly used spatiotemporal data blending method
are considered in our theoretical analysis. Simulated experi-
ments were also performed to confirm the conclusions of the
conducted theoretical analysis.
Xuehong Chen, Jin Chen, and Xin Cao are with the State Key
Laboratory of Earth Surface Processes and Resource Ecology,
Faculty of Geographical Science, Beijing Normal University,
Beijing 100875, China (
).
Meng Liu is with the College of Global Change and Earth System
Science, Beijing Normal University, Beijing 100875, China.
Xiaolin Zhu is with the Department of Land Surveying and
Geo-Informatics, Hong Kong Polytechnic University, Hong
Kong, China.
Yanfei Zhong is with the State Key Laboratory of Information
Engineering in Surveying Mapping and Remote Sensing,
Wuhan University, Wuhan, China.
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 2, February 2018, pp. 65–73.
0099-1112/17/65–73
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.2.65
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2018
65
51...,55,56,57,58,59,60,61,62,63,64 66,67,68,69,70,71,72,73,74,75,...114
Powered by FlippingBook