You are on page 1of 15

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/295636499

An Improved Spatial Downscaling Procedure


for TRMM 3B43 Precipitation Product Using
Geographically Weighted Regression

DATASET · FEBRUARY 2016

READS

4 AUTHORS, INCLUDING:

Cheng Chen Zheng Duan


Nanjing University Delft University of Technology
4 PUBLICATIONS 0 CITATIONS 30 PUBLICATIONS 551 CITATIONS

SEE PROFILE SEE PROFILE

Zhihao Qin
Chinese Academy of Agricultural Sciences
76 PUBLICATIONS 1,061 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Cheng Chen
letting you access and read them immediately. Retrieved on: 25 February 2016
4592 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 9, SEPTEMBER 2015

An Improved Spatial Downscaling Procedure


for TRMM 3B43 Precipitation Product Using
Geographically Weighted Regression
Cheng Chen, Shuhe Zhao, Member, IEEE, Zheng Duan, and Zhihao Qin

Abstract—Precipitation data at high spatio-temporal resolu- Index Terms—Disaggregation, geographically weighted


tion are essential for hydrological, meteorological, and ecological regression (GWR), multivariate regression (MR), satellite
research in local basins and regions. The coarse spatial reso- precipitation, validation.
lution (0.25◦ ) of Tropical Rainfall Measuring Mission (TRMM)
3B43 product is insufficient for practical requirements. In this
paper, a multivariable geographically weighted regression (GWR) I. I NTRODUCTION
downscaling method was developed to obtain 1 km precipitation.
The GWR method was compared with two other downscaling
methods [univariate regression (UR) and multivariate regression
(MR)] in terms of the performance of downscaled annual precipi-
P RECIPITATION, one of the crucial meteorological ele-
ments of climate change, plays an important role in
atmospheric processes at different temporal and spatial scales.
01 tation. Variables selection procedures were proposed for selecting Precipitation is a major driving force in global climate change,
appropriate auxiliary factors in all three downscaling methods. To hydrological cycle, and ecological environment, which is of
celo Portuguez obtain the monthly 1 km precipitation, two monthly downscal-
ing strategies (annual-based fraction disaggregation method and significant importance to agricultural irrigation, disaster pre-
monthly based GWR method) were evaluated. All analysis was vention, and other production activities [1], [2]. The traditional
tested in Gansu province, China with a semiarid to arid climate method to obtain precipitation data is completely dependent
for three typical years. Validation with measurements from 24 rain on the observations from point-based rain gauge stations.
gauge stations showed that the proposed GWR method performed However, the method faces challenges in some areas with rare
consistently better than the UR and MR methods. Two monthly
downscaling methods were efficient in deriving the monthly pre- rain gauge sites, especially in the mountainous and remote
cipitation at 1 km. The former method faces the challenge of areas. Remote sensing has the advantages of completely scan-
precipitation spatial heterogeneity and the derived monthly pre- ning the entire region under study, shortly revisiting to the
cipitation heavily depends on the annual downscaled results, which same region, and conveniently accessing to the data. With the
could lead to the accumulation of errors. The monthly based development of modern remote sensing technology, precipita-
GWR method is suitable for downscaling monthly precipitation,
but the accuracy of original TRMM 3B43 data would have large tion could be detected at a regional or global scale. Tropical
influence on downscaling results. It was demonstrated that the Rainfall Measuring Mission (TRMM) precipitation data with a
proposed method was effective for obtaining both annual and relatively high-spatial resolution and a relatively large scope of
monthly TRMM 1 km precipitation with high accuracy. space had been widely used in recent years [3], [4]. However,
the finest spatial resolution of gridded product of TRMM is
Manuscript received February 06, 2015; revised May 12, 2015; accepted May 0.25◦ , and this resolution is too coarse for local basin and
27, 2015. Date of publication June 24, 2015; date of current version December regions [5]. Precipitation with high spatio–temporal resolution
21, 2015. This work was supported in part by the National Basic Research and high accuracy is a key input parameter in hydrological
Program of China under Grant 2010CB951503, in part by the “Strategic Priority
Research Program—Climate Change: Carbon Budget and Related Issues” of model, which could provide a detailed characterization of the
the Chinese Academy of Sciences under Grant XDA0505 0106, and in part by spatial distribution of flood and drought disasters, debris flow,
the Project funded by the Priority Academic Program Development of Jiangsu and so on, hence the research on spatial downscaling of TRMM
Higher Education Institutions (PAPD). (Corresponding authors: Shuhe Zhao
precipitation products from relatively coarse to finer spatial
and Zheng Duan.)
C. Chen and S. Zhao are with the School of Geographic and Oceanographic resolution has attracted more attentions in recent years [5]–[7].
Sciences, Nanjing University, Nanjing 210023, China, and with the The general spatial downscaling procedure is to construct
Jiangsu Provincial Key Laboratory of Geographic Information Science and the regression relationship between the target variable and its
Technology, Nanjing University, Nanjing 210023, China, and also with the
Jiangsu Center for Collaborative Innovation in Geographical Information
controlling/influencing variables at the relatively coarse spatial
Resource Development and Application, Nanjing 210023, China (e-mail: resolution of the target variable, and then the constructed rela-
chcheng@whu.edu.cn; zhaosh@nju.edu.cn). tionship is used with the controlling variables at finer spatial
Z. Duan is with Department of Civil Engineering and Geographic Sciences, resolution to retrieve the target variable at the same finer res-
Delft University of Technology, 2628 CN Delft, The Netherlands (e-mail:
duanzheng2008@gmail.com).
olution. The downscaling procedures have been widely used
Z. Qin is with the Institute of Agricultural Resources and Regional Planning, in many fields to obtain the variables at fine/high resolution,
Chinese Academy of Agricultural Sciences, Beijing 100081, China (e-mail: such as land surface temperature (LST) [8]–[11] and soil mois-
zhihaoqin@163.com). ture [12], [13]. The research on the downscaling of the TRMM
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. precipitation data is relatively fewer. Considering the close rela-
Digital Object Identifier 10.1109/JSTARS.2015.2441734 tionship between the Normalized Difference Vegetation Index
1939-1404 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
CHEN et al.: IMPROVED SPATIAL DOWNSCALING PROCEDURE FOR TRMM 4593

(NDVI) and precipitation, Immerzeel et al. [5] first constructed addressing spatial nonstationary and scale-dependent problems
the exponential regression to downscale the TRMM 3B43 data, in landscape ecology. Song et al. [22] developed a satellite-
and the 1 km TRMM annual precipitation data were obtained on based GWR model to estimate the daily concentration of
the Iberian Peninsula. A quadratic polynomial regression rela- ground-level PM2.5 coincident to satellite overpass at regional
tionship between NDVI and TRMM 3B43 data was established scale. The GWR method was also used to estimate the LST
by Duan and Bastiaanssen [6] to construct the downscaling [23] and urban heat island [24]. To our knowledge, Chen et al.
model. The calibration procedures using the rain gauges data [25] was the first study, which introduced the GWR method
were also introduced to improve the accuracy of the downscaled to downscale the TRMM annual precipitation in north China.
TRMM annual precipitation, and the annual TRMM precipita- They found that the GWR method could fit the relationship
tion data were also disaggregated to obtain the monthly TRMM between NDVI, DEM, and TRMM better than multiple linear
precipitation. The proposed integrated downscaling-calibration regression method, and the accuracy of the downscaled precip-
procedure performed well in humid region (Lake Tana Basin itation was significantly improved (e.g., R2 from 0.80 to 0.82,
in Ethiopia) and the semiarid region (Caspian Sea Region in RMSE from 125.4 to 91.0 mm/year in 2001) compared to the
Iran). Jia et al. [2] used a combination of NDVI and Digital measurements from 49 rain gauge stations.
Elevation Model (DEM) in the Qaidam Basin of China to down- As summarized above, there are two main research gaps
scale TRMM 3B43 precipitation for obtaining the 1 km annual to be filled about the GWR-based spatial downscaling proce-
precipitation. Hunink et al. [7] conducted a research on the esti- dure of precipitation. The first gap is the limited number of
mation of weekly spatial precipitation based on NDVI, DEM, factors (NDVI and DEM) that are considered; the correlation
TRMM-based monthly average rainfall and field observations between DEM and TRMM precipitation could be insignificant
in a tropical mountainous region in Ecuador. It was shown that and not useful in certain regions. The introduction of factors
40% of the variance in this highly variable area was explained such as LST, topography (e.g. DEM, Slope, and Aspect), and
by these proxies with vegetation being the strongest proxy. geographical location (e.g. Latitude and Longitude) may be
Considering the impact of precipitation by multiple factors, good indications of precipitation in local area of the study area.
Fang et al. [14] used the slope, aspect, and terrain rough- The second gap is that almost all previous studies focused on
ness extracted from DEM, and humidity and temperature from only annual scale and fewer studies [6] on the downscaling of
meteorological conditions, into the multiple linear regression the TRMM precipitation at monthly or finer time scales. In the
equation to fit the regression function. All the above stud- paper, we aimed at filling the above two gaps by developing a
ies were based on the fundamental global regression between new multivariable GWR downscaling method to downscale the
TRMM precipitation data and a variety of auxiliary variables. TRMM 3B43 precipitation at both annual and monthly scales.
The global regression such as ordinary least squares (OLSs) The proposed GWR method was also compared with two other
regression means that the functional relationship is established commonly used downscaling methods, i.e., univariate regres-
using the same model for all the pixels. Although global regres- sion (UR) and multivariate regression (MR) to demonstrate the
sion relationships are relatively well-established, such use of superiority of the proposed method.
regression models is based upon the assumption of spatial sta-
tionarity in the relationship between the variables under study
[15]. However, the relationships are often not stable in space. II. S TUDY A REA AND DATASETS
Tobler’s first law (TFL) [16] shows that everything is related
A. Study Area
to everything else, but near things are more related than dis-
tant things. It indicates that geographical entities often have The Gansu province was selected as a case study area in the
local properties, and numerous examples in remote sensing- paper. Gansu province is located in the northwest of China and
related literature show the spatial nonstationarity of observed upstream of the Yellow River with the latitude of 32◦ 11 N–
geographical and ecological patterns and processes in nature 42◦ 57 N and longitude of 92◦ 13 E–108◦ 46 E (Fig. 1). Owing
[15], [17]–[19]. Li et al. [20] conducted an investigation of to the impact of topography, geographical location, and atmo-
urban surface temperature and found that the global regression spheric circulation, precipitation in Gansu is not only uneven
results in a failure to capture the spatial dependence when it in spatial distribution, but also great on the annual and inter-
was applied to geo-referenced data analyses. The meteorologi- annual variability. Territory of Gansu is vast and accounts for
cal variables such as precipitation and temperature are typically 4.73% of total area of China. It lies in the intersection zone of
spatial variable. Therefore, the global regression method is of Loess Plateaus, the Qinghai Tibet Plateaus, and Inner Mongolia
great limitation in establishing the relationship between tar- plateau. Gansu has a complex landform and is very moun-
get variable and its controlling variables of the precipitation tainous in the south and flat in the north. The elevations have
downscaling. greater disparity and the topography is slanted downward in
Recently, the geographically weighted regression (GWR) the northeast direction from the southwest. The area of arable
proposed by Brunsdon et al. [21] has been increasingly used land is 81.15 million acres, and the rain-fed agriculture and
in various fields to detect the spatial nonstationarity. For exam- dryland farming occupy the absolute superiority (80%). Gansu
ple, Zhao et al. [19] explored spatially variable relationships is landlocked and generally has a semiarid to arid, continental
between NDVI and climatic factors (annual precipitation and climate, with warm to hot summers, and cold to very cold win-
annual mean temperature) in a transition zone using GWR, and ters. The annual average temperature is between 0 and 14 ◦ C
it was shown that this local regression method is practical for and the annual average precipitation is about 300 mm/year.
4594 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 9, SEPTEMBER 2015

intervals using the TRMM multisatellite precipitation anal-


ysis (TMPA). The TMPA combines multiple independent
precipitation estimates from the TRMM Microwave Image
(TMI), Advanced Microwave Scanning Radiometer for Earth
Observing Systems (AMSR-E), Special Sensor Microwave
Imager (SSMI), Special Sensor Microwave Imager/Sounder
(SSMIS), Advanced Microwave Sounding Unit (AMSU),
Microwave Humidity Sounder (MHS), microwave-adjusted
merged geo-infrared (IR), and monthly accumulated Global
Precipitation Climatology Centre (GPCC) rain gauge analysis.
The TMPA data take advantage of the fine scales to create aver-
ages appropriate to the user’s application, which has a good
performance at monthly and daily scales. More details on the
TMPA are available through Huffman et al. [26]. The TRMM
3B43 monthly precipitation data in a complete record from
Fig. 1. Geographical location and land cover types of Gansu province in China. January 1998 to December 2013 used in this paper were freely
Hill-shade and the location of the rain gauges are superimposed. available from the NASA database (http://www.nasa.gov/). The
monthly precipitation data were accumulated to obtain the
TRMM annual precipitation.

C. SPOT-NDVI
The used SPOT VGT-S10 products at 1 km resolution were
obtained from http://www.spot-vegetation.com/. The SPOT-
NDVI products are a result of the merging of the NDVI data
strips from 10 consecutive days, and the cloud cover influ-
ences are favorably reduced. Data preprocessing procedures
include geometric, radiometric, and atmospheric corrections.
More details about SPOT VGT-S10 products can be found in
documents available in the link mentioned above. The monthly
NDVI data were acquired by averaging the average values of
the 10-day SPOT-NDVI product. The monthly NDVI data were
averaged to get the annual NDVI.
Fig. 2. Average monthly precipitation and temperature from 24 rain gauge
stations during 2000–2013 in Gansu, China.
D. Shuttle Radar Topography Mission Digital Elevation Model
However, as far from the sea, the water vapor transport is The Shuttle Radar Topography Mission (SRTM) is
impeded. Consequently, the precipitation in Gansu shows great an international project spearheaded by the National
spatial variation. The maximum value of annual precipitation Geospatial-Intelligence Agency (NGA) and the NASA
occurs in the southeast of Gansu (700–800 mm/year), and the (http://www.nasa.gov/). The SRTM DEM data at the spatial
minimum value occurs in the northwest (Hexi Corridor, 40– resolution of 90 m were used in this study. The 90 m DEM
200 mm/year). Most of the precipitation falls in the summer was resampled to the targeted resolution (e.g., 0.25◦ and 1 km)
months (May–September, Fig. 2). The overall trends of both using the pixel averaging method during the downscaling
temperature and precipitation decrease from the southeast to the procedures. The variables (slope and aspect data) were further
northwest. extracted from the DEM data. In addition, the Longitude
and Latitude data were also extracted from the projection
information of the DEM data.
B. TRMM 3B43
E. Land Surface Temperature
The TRMM is a joint space mission between National
Aeronautical and Space Administration (NASA) of the USA The Terra-MODIS daytime LST products (MOD11A2)
and the National Space Development (NASDA) of Japan with 1 km resolution from January 2000 to December 2013
designed to monitor and study tropical rainfall. The satel- in the study area were obtained from the NASA database
lite was launched on November 27, 1997. TRMM 3B43 (http://www.nasa.gov/). The MOD11A2 products offer LST
data cover areas from 50◦ S to 50◦ N latitude, 180◦ W to data stored on a 1 km sinusoidal grid as the average values of
180◦ E longitude. The TRMM 3B43 data are on a calendar clear-sky LSTs during an 8-day period [27]. To eliminate the
month temporal resolution and a 0.25◦ × 0.25◦ spatial res- influence of cloud, the monthly values of LST were calculated
olution. The TRMM 3B43 product is processed at monthly weighting with the number of days belonging to each month
CHEN et al.: IMPROVED SPATIAL DOWNSCALING PROCEDURE FOR TRMM 4595

based on 8-day LST provided metadata after masking the filled which easily lead to large rainfall. In Gansu, the high longi-
and missing values [28]. The monthly LST data were averaged tude will also lead to large rainfall because of the close distance
to obtain the annual LST. of the land–sea location. Generally, precipitation increases with
elevation and slope and decreases with aspect due to the topo-
graphical lifting effect on airflow. Jin et al. [29] found that in
F. Rain Gauges Data
east part of northwest China (east of 93◦ E), the precipitation
A dataset with monthly rain gauges data in Gansu was used declined with global warming. Recent studies have confirmed
for the validation of the downscaling results. Measurements that LST is highly correlated with air temperature [30], thus
from 24 rain gauge stations during 2000–2013 were obtained the LST may be also a good indication of precipitation in local
from the China Meteorological Data Sharing Service System areas of the study area. Moreover, the growth of vegetation is
(http://data.cma.gov.cn/). The locations of 24 rain gauges are often influenced by LST and precipitation such as the effect of
shown in Fig. 1. The analysis of rain gauges data showed drought on vegetation. Therefore, the consideration and addi-
that average annual precipitation was 280 mm/year during tion of LST in the downscaling procedures could improve in
2000–2013. The year 2005 represents the average year with describing the complex interactions between NDVI and precip-
290 mm/year, and the year 2007 with 325 mm/year represents itation especially at monthly and daily scale, which is expected
the wet year. The year 2009 is an extremely dry year with to make the spatial downscaling results more realistic. Finally,
only 230 mm/year of precipitation in Gansu. These three typ- a total of seven auxiliary factors (i.e., NDVI, LST, DEM, Slope,
ical years were selected to evaluate the performance of spatial Aspect, Latitude, and Longitude) were considered in the paper.
downscaling procedures in different climatic conditions. Three different downscaling methods were considered and eval-
uated further to determine whether each individual variable is
useful or not in the spatial downscaling procedures.
III. M ETHODOLOGY
1) UR Method: In general, there is a significant relation-
The procedures of the methodology in this paper are pre- ship between NDVI and TRMM data at the annual scale [31],
sented in this section. The section is divided into five parts. In [32]. Accordingly, NDVI is the most common used factor in
Section III-A, the variables selection procedure was proposed the UR method to downscale the TRMM annual precipitation.
to select the appropriate auxiliary factors to construct three dif- However, this relationship between NDVI and TRMM could
ferent downscaling methods. The detailed principle explanation be not invariably significant in all regions. Therefore, a simple
of the GWR method is shown in Section III-B. The spatial linear regression analysis between precipitation and auxiliary
downscaling procedures of TRMM 3B43 annual precipitation topographical and meteorological variables would be first con-
are described in Section III-C. Two downscaling strategies were ducted in different years. The factor with the best correlation
also tested to obtain the monthly precipitation at 1 km resolu- with TRMM precipitation in all years should be considered
tion, which is detailed in Section III-D. Finally, Section III-E as the auxiliary variable. Second, various regression methods
describes four indices that were computed for validating the such as polynomial regression, logarithm regression, exponen-
downscaling results. tial regression, and linear regression are compared to choose the
one that results in best fitting.
2) MR Method: The spatial distribution of precipitation in
A. Variables Selection for Spatial Downscaling northwest China is affected by geographical location, topogra-
As a complex meteorological element, precipitation interacts phy, and other factors as mentioned above, so it is expected that
with various factors, thus the selection of auxiliary variables the MR method would generate better performance. Fang et al.
for spatial downscaling would have significant effects on the [14] used several factors such as DEM, slope, aspect, interpo-
results. One critical requirement for the inclusion/selection of lated temperature, and humidity in the MR method to down-
an auxiliary variable in the spatial downscaling procedure is scale the TRMM precipitation data at daily scale. However, an
that this variable should be easily obtainable at fine spatial res- in-depth analysis often lacks, which would easily lead to the
olutions. In this study, all considered auxiliary variables are the phenomenon of multicollinearity, which means two or more
ones that can be obtained from satellite observations at 1 km individual variables in a multiple regression method are highly
resolution. correlated and the coefficient estimates of the multiple regres-
The selection of appropriate variables for downscaling sion may change erratically in response to small changes in the
should take the characteristics of a given study area into method. Therefore, the autocorrelation analysis of independent
account. Gansu is a typical arid and semiarid region in the variables and multicollinear test of linear regression equation
northwest of China. The growth of green vegetation mainly are of great importance in the MR method. Those unsatisfied
depends on precipitation in Gansu, so it is generally expected factors in the MR method should be eliminated.
that the more rainfall is and the larger values of NDVI will 3) GWR Method: The above two methods are the global
be, which forms a close positive relationship between precip- regression model, which establishes the function relationship
itation and NDVI. In addition, the topography (e.g., DEM, using the same model for all the pixels. In global regres-
Slope, and Aspect) and geographical location (e.g., Latitude sion model, the precipitation responds to changes in inde-
and Longitude) also have significant impact on the precipitation pendent variables that may cause errors in local basins and
in the study area. Areas located in low latitude usually have rel- regions where precipitation is not governed by the indepen-
atively high temperature, evaporation, and water vapor content, dent variables. Considering the precipitation is typically spatial
4596 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 9, SEPTEMBER 2015

nonstationarity and has great spatial variation, it is essential The Gauss function describes the relationship between Wij
to develop the GWR method. Because the geographical loca- and dij using a continuous monotonically decreasing func-
tions parameters were already the input in the GWR method, tion. Because of its universal application, the Gauss function
the Latitude and Longitude data were used as the auxiliary is widely used. The formula is as follows:
variables. The regression parameters (1) in the GWR method  
should be interpolated into 1 km resolution. To reduce the errors Wij = exp(− dij /b)2 (4)
of the interpolation, the auxiliary variables should be suitable
and reasonable. Therefore, only the auxiliary variables with rel- where b represents the bandwidth and dij is the distance func-
atively high correlations with TRMM precipitation were used in tion. Qin [36] found that the GWR method is less affected
the GWR method. by the changes of the variable bin, but the parameter b is
sensible to the regression results and can be obtained by
the cross validation. Consequently, the GWR method is suit-
B. Geographically Weighted Regression able for detecting the spatial nonstationarity of precipitation.
Traditional regression model such as OLS method is estab- The econometrics toolbox in MATLAB version 7 (http://
lished on the assumption that the regression relationship www.spatial-econometrics.com/), which includes the GWR
between the variables and the dependent variable remains stable model was used in this study.
in the entire region. Precipitation is the interaction result of var-
ious factors. The local precipitation variation typically occurs
C. Spatial Downscaling Procedures of TRMM 3B43 Annual
at scales of 2 km and larger [5], [33]. Precipitation is typically
Precipitation
spatial nonstationarity and has great spatial differences even in
the same underlying surface. The OLS as a global model cannot To validate the reliability of the relationship between
reflect the real spatial characteristics of the regression parame- TRMM precipitation and NDVI at different spatial resolutions,
ters. On the basis of the study of variable parameters [34] and Immerzeel et al. [5] analyzed the fitting relationships at mul-
local regression, Brunsdon et al. [21] proposed a weighted least tiple scales (0.25◦ −1.25◦ ) and found that the resolution at
squares regression method. The GWR method is the expansion 0.75◦ was optimal. A multiresolution wavelet analysis, which
of the general linear regression, and it can detect the spatial is based on the wavelet decomposition and reconstruction, was
heterogeneity by introducing the geographical location into the used to determine the best regression function in the paper, and
regression equation [21], [35]. The general form of GWR is it was found that the regression function became better with
described in (1) increasing resolutions (0.25◦ −1.0◦ ) (results are not shown for
p
conciseness). However, with the increase of the resolution, the
 number of pixels used in the fitting function is decreased, which
yi = β0 (ui , vi ) + βk (ui , vi )xik + εi (1)
in turn reduces the statistical significances of the regression
k=1
analysis [6]. Therefore, the original fitting function at 0.25◦ res-
where yi is the i th observation of the dependent variable, xik is olution was used for establishing the regression relationship.
the i th observation of the kth independent variable, the (ui , vi ) In addition, we also used the local Moran’s index [2] to detect
represents the geographical coordinates of ith sample point, and and eliminate the outliers of NDVI pixels, which are not gov-
βk (ui , vi ) is the kth regression parameter in the ith regression erned by precipitation, but only few pixels were eliminated,
point, which is a function of geographical location. εi repre- which makes little contributions to the establishment of regres-
sents independent normally distributed error term with zero sion relationship. The specific steps of the multivariable GWR
mean. Equation (1) can be simplified as downscaling method are described as follows (the superscript
LR represents low spatial resolution at 0.25◦ and the superscript
n
 HR represents high/fine spatial resolution at 1 km).
yi = βi0 + βik xik + εi (2)
1) Sum the original 0.25◦ resolution TRMM 3B43 monthly
k=1
precipitation to annual precipitation (TRMMLR original ).
when β1 k = β2 k = · · · = βnk , the GWR model degenerates {NDVIHR , LSTHR , . . . , LongituteHR } represent the
to a general linear regression model. The parameters were annual values of the auxiliary factors (e.g. NDVI,
estimated by the weighted least squares estimation, and the LST, DEM, Slope, Aspect, Latitude, and Longitude) at
parameter estimation in the ith sample point can be written as 1 km resolution, respectively. {NDVILR , LSTLR , . . . ,
 −1  T  LongituteLR } are the resampled auxiliary factors at
βi = X T Wi X X Wi Y (3) 0.25◦ resolution.
2) Establish a local regression relationship between
where Wi is n × n matrix, and each element of the diagonal TRMMLR and {NDVILR , LSTLR , . . . ,
original
matrix is a function of distance between the observation in the LongituteLR } by (2), where X represents the auxiliary
jth position and regression point in the ith position. The βi in factor
(3) is the result of the parameter estimation. The common dis-
tance functions include the distance threshold method, inverse N

distance method, Gauss function, and the bisquare function TRMMLR = β0LR + βkLR XLR
k +ε
LR
. (5)
[21]. More details can be found in [21] and [36]. k=1
CHEN et al.: IMPROVED SPATIAL DOWNSCALING PROCEDURE FOR TRMM 4597

3) Use the weighted least squares estimation to obtain the precipitation by establishing the relationship between precipi-
model parameters β0LR , β1LR ,. . ., βN
LR
, and the simple tation and auxiliary factors for each month similar to the annual
spline interpolation method, which is typically used for downscaling procedure mentioned above.
regularly spaced data [5], was implemented to interpolate
the model parameters into 1 km resolution (β0HR , β1HR ,. . .,
βNHR
). The model residual (εLR ), which represents the E. Validation
part cannot be explained by the auxiliary factors, were The measured rainfall data from 24 rain gauge stations
also interpolated into 1 km spatial resolution (εHR ). were used to validate the downscaling results of three differ-
4) Following (5), the 1 km TRMM annual precipitation was ent methods (UR, MR, and GWR) in three typical years (2005,
then calculated as described in precipitation average year; 2007, wet year; 2009, dry year). The
N
 following four indices in (8)–(11) were computed for validation
TRMMHR = β0HR + βkHR XHR
k +ε
HR
. (6) purpose.
k=1 The coefficient of determination (R2 ): The coefficient of
determination reflects the degree of linear correlation between
For the UR and MR methods, the downscaling procedures
satellite precipitation and gauge observations. Formula is as
are basically similar to the above steps. The main difference is
follows:
in the second step, and the different regression relationships are
⎛  ⎞2
established in different methods. Besides, in the third step, we N  
M i − M P i − P
only need to interpolate the residual errors into 1 km resolution R 2 = ⎝   i
2  2
⎠ (8)
in the UR and MR methods. N
i Mi − M P i − P

where Mi represents the ith rain gauge observation, Pi repre-


D. Spatial Downscaling at Monthly Scale
sents the ith TRMM 3B43 precipitation, M represents the mean
Monthly precipitation data with high resolution are intensely precipitation of the rain gauges data, P represents the mean pre-
essential for hydrological, meteorological, and ecological cipitation of downscaling results, and N represents the number
research in local basins and regions. The monthly precipitation of the rain gauges.
data at 1 km resolution can provide more information for flood The relative bias (Bias): The Bias describes the systematic
and drought monitoring. So far, there are few articles about bias of satellite precipitation estimates. The smaller absolute
the monthly downscaling of the TRMM precipitation. Since it value of Bias indicates the smaller deviation. The value of Bias
was reported that the vegetation response reflected by NDVI greater than zero indicates that satellite precipitation is greater
had a time lag with the cumulative monthly precipitation [37], than the gauge observations, or vice versa. Bias is calculated in
[38], using NDVI as the auxiliary variable is significantly lim- N
ited in the UR and MR downscaling methods. In the paper, Pi
two monthly downscaling strategies were tested to obtain the Bias = Ni − 1. (9)
i Mi
monthly 1 km precipitation.
The first was the annual-based fraction disaggregation The root-mean-square error (RMSE): RMSE is frequently
method. Duan and Bastiaanssen [6] used the fraction derived used to measure the differences between values predicted by a
from original TRMM 3B43 monthly data to disaggregate the model or an estimator and the values actually observed, which
downscaled annual precipitation maps into maps at monthly describes the average error magnitude in the paper. The small
time steps. The annual precipitation data were obtained on the RMSE means the small error between the downscaling results
basis of the NDVI–TRMM relationship established by a global and rain gauge observations
regression. In this paper, we also introduced the fraction disag-
gregation procedures and compared the performances of three N 2
i (Pi − Mi )
annual-based downscaling results by the UR, MR, and GWR RMSE = . (10)
N
methods in three typical years. The fractions are defined in
OrgT RM Mi The mean absolute error (MAE): MAE also represents the
Fi = 12 (7) average magnitude of the error. Because the positive and nega-
i=1 OrgT RM Mi tive offsets will not cancel each other out, the MAE can describe
where the OrgT RM Mi represents the precipitation that occurs the error of downscaling results better
during the ith month in a year as estimated from the original N 2
TRMM 3B43 product. Considering the fractions at 0.25◦ scale i (Pi − Mi )
MAE = . (11)
are regularly spaced, the same simple spline method was used to N
interpolate the fractions into 1 km resolution. Then, the annual
downscaling results of the precipitation data at 1 km scale were IV. R ESULTS AND D ISCUSSION
multiplied by the fractions to get the monthly precipitation in
A. Annual Downscaling Results
2005, 2007, and 2009.
The second was the monthly based GWR method in which 1) Regression Analysis of Three Downscaling Methods:
the GWR method was directly used to downscale the monthly Table I presents the correlation between TRMM 3B43 annual
4598 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 9, SEPTEMBER 2015

TABLE I
R ESULTS OF R EGRESSION A NALYSIS B ETWEEN TRMM 3B43 A NNUAL P RECIPITATION AND AUXILIARY VARIABLES FOR T HREE T YPICAL Y EARS

Meanwhile the precipitation is increased as the rise of eleva-


tion, thus a negative relationship is generated between LST and
precipitation. In Fig. 4(d), the values are uniform and approxi-
mately equal to zero, suggesting that the slope for the variable
Slope is less important than the other factors. Fig. 4(e) is the
GWR residual map at 1 km resolution. The residual map rep-
resents the part of precipitation that cannot be explained by the
auxiliary factors. Positive values of the residual indicate that the
precipitation is underestimated by the GWR method, while neg-
ative values indicate the overestimate of precipitation. Fig. 4(e)
shows that most of the residual values are within the range of
−50 to 50 mm/year, suggesting that the predicted precipitation
by the GWR method is quite close to the original observations
from TRMM data. Fig. 4(f) shows the spatial distributions of
local coefficient of determination (R2 ) for the GWR method,
with the average value of 0.76 and the range of 0.24–0.98. High
values are distributed in the middle part of Gansu province,
Fig. 3. Fitting results between NDVI and TRMM annual precipitation in the where the values of intercept β0 are large [Fig. 4(a)] and the
UR method. spatial variations in slope for NDVI, LST, and Slope are large
[Fig. 4(b)–(d)]. Low values are distributed in the northwest-
precipitation and various auxiliary factors in three typical years. ern where desert is the dominant landscape; the southeastern
The “multivariate” in Table I means that all the seven factors region where the spatial variations are relatively small also
were used in the regression analysis. has low values in spite of large precipitation [Fig. 4(b)–(d)].
It can be seen that NDVI has positive and high linear correla- This may indicate that the GWR method is more suitable for
tions with TRMM annual precipitation in all three typical years, the regions where the spatial variations of precipitation and
therefore NDVI can be used as the factor in the UR method to auxiliary factors are large.
downscale the TRMM data. Various regression functions were The predictive performance of all three methods (UR, MR,
tested and the quadratic polynomials regression (Fig. 3) was and GWR) can be evaluated by comparing the predicted annual
the optimum with the largest R2 (P < 0.001) in all years. precipitation by each method with the original TRMM annual
Eventually the quadratic polynomials regression of NDVI was precipitation at 0.25◦ . Such evaluation was conducted, and the
selected in the UR method. results are shown in Table III for three different typical years.
After autocorrelation analysis of independent variable factors Table III shows that the UR method has relatively high degrees
and multicollinearity test of linear regression equation, six fac- of correlation in three typical years (R2 = 0.80−0.86, RMSE =
tors (NDVI, LST, DEM, Slope, Aspect, and Longitude) were 91.6−100.0 mm/year, MAE = 62.1−69.9 mm/year). The UR
finally selected in the MR method. The fitting functions in the method using only the NDVI data is very simple, and the
MR method for all three years are shown in Table II. data is easily available. As expected, the MR method, which
For the GWR method, three variables (NDVI, LST, and introduces more auxiliary factors into the regression equation,
Slope) with higher degrees of correlation were selected as the can describe the spatial distribution of the precipitation better,
regression/independent variables in the GWR method. The spa- with higher R2 and lower RMSE and MAE than those in the
tial distributions of the model parameters, residual and local UR method. The GWR method has the best performance in
R2 for the GWR method in the average 2005 are shown in all three typical years, and the RMSE and MAE are reduced
Fig. 4 for illustrative purpose. It is shown in Fig. 4(b) that the greatly when compared to the UR and MR methods. This is
slope for NDVI is positively related to precipitation in most mainly because the global regression model, which ignores
parts of the study area and large values of the slope occur in the spatial heterogeneity of precipitation, has the obvious
the southeast part of the study area. The slope for LST has phenomenon of over fitting, so it is easy to generate deviations
large negative values in the southwest of Gansu province with and extreme values. It can be concluded that the GWR method
a relatively high elevation [Fig. 4(c)]. Southwestern Gansu is is a superior description of spatial distribution of the annual
mountainous and LST is decreased with the rise of elevation. precipitation in Gansu province.
CHEN et al.: IMPROVED SPATIAL DOWNSCALING PROCEDURE FOR TRMM 4599

TABLE II
R EGRESSION F UNCTIONS IN THE MR M ETHOD FOR T HREE T YPICAL Y EARS

Fig. 4. Spline interpolation results of the estimation parameters at 1 km resolution in GWR model in the average year 2005. (a) Intercept β0 . (b) Slope β1 for
NDVI. (c) Slope β2 for LST. (d) Slope β3 for Slope. (e) Residuals. (f) Local R2 .

TABLE III because that the large rainfall could be caused by more complex
E VALUATION R ESULTS OF THE UR, MR, AND GWR M ETHODS U SING processes, such as the climate anomaly and atmospheric circu-
THE O RIGINAL TRMM 3B43 DATA FOR T HREE T YPICAL Y EARS
lation, which cannot be well explained by the topographical and
meteorological factors used in this paper. The extremum formu-
lations of exponential regression, which performs better even
for very wet years, maybe a good choice in elsewhere. However,
this method could lead to over-fitting and is unsuitable for
multivariable auxiliary factors.
2) Downscaling Results: The downscaling results in the
average year 2005 are shown, e.g., in Fig. 5. Fig. 5(a) shows
the original TRMM annual precipitation at 0.25◦ . Fig. 5(b)–(d)
shows the downscaled annual precipitation at 1 km resolution
using the UR, MR, and GWR methods, respectively. The white
regions (no data) shown in Fig. 5(b)–(d) represent the outliers
with extreme high or low values.
It can be found in Fig. 5(b) and (c) that there are many
As far as the performances in different years are concerned, outliers in the downscaling results. The reason for that is
the best fitting was in the average precipitation year 2005. the regression functions of the UR and MR methods estab-
The wet year 2007 was the least optimal among three years, lished in the entire area, thus a lot of independent variables
which indicates the explanatory ability of the auxiliary variables are difficult to fully fit, which easily leads to the over-fitting
becoming weak as the increasing of precipitation. This may be phenomenon. Therefore, these global regression methods have
4600 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 9, SEPTEMBER 2015

Fig. 5. Overview of the annual downscaling results in 2005. (a) Original TRMM precipitation data at 0.25◦ resolution. (b)–(d) Downscaling results of the UR,
MR, and GWR methods, respectively.

great limitations when used for the downscaling of precipita- weighted to the sampling point to establish a robust regression
tion. The GWR method, which allows different relationships to equation. The multivariable GWR method could avoid over-
exist at different locations, can visually detect the spatial het- fitting [black rectangle frame in Fig. 5(d)] and is more suitable
erogeneity of precipitation. The GWR method can construct for the downscaling of precipitation.
the robust local regression equation, and thus it can reduce the Fig. 6(a)–(c) presents the comparisons between measure-
outliers significantly. ments from 24 rain gauge stations and the downscaled precip-
From the downscaling results of the TRMM annual precip- itation from three downscaling methods: UR, MR, and GWR
itation, we can find that rainfall in Gansu province is slanted methods, respectively. The horizontal axis represents the rain
downward in the northwest direction from the southeast. The gauges measured data, and the vertical axis represents the
annual precipitation in the Hexi Corridor northwest of Gansu downscaled results of the TRMM precipitation data at 1 km
is intensely scarce, and the annual precipitation of the Loess resolution. For comparison, Fig. 6(d) represents the fitting rela-
Plateau area in the middle of Gansu is relatively small, whereas tionship between the rain gauges data and the original TRMM
the maximum annual precipitation is located in the east of precipitation data at 0.25◦ .
Gansu. The precipitation is significantly related with geographi- It can be seen from Fig. 6(a)–(c) that different downscaling
cal location. The main land cover type in Jiuquan city northwest methods have different results, and the coefficients of the deter-
of Gansu is bare land (see Fig. 1). The small amounts of farm- mination (R2 ) in the average year 2005 are 0.660, 0.685, and
land and grassland in the central area of Jiuquan city make 0.833, respectively. The UR method had the worst result, which
the NDVI values larger than the surrounding area. The variable indicates the single factor of NDVI is insufficient to describe
NDVI presents larger contributions in the UR and MR methods, the spatial distribution of the precipitation. Besides, the values
and there are positive correlations between NDVI and down- of RMSE, Bias and MAE of the MR and UR methods are basi-
scaling results of precipitation, which means the larger values cally the same (Table IV). The MR method with relatively small
of NDVI reflect higher precipitation. So the downscaling results values of MAE and RMSE has a better performance than the
of precipitation in the area are larger than the surrounding area UR method. However, in the meteorological sites of Zhangye
[black rectangle frame in Fig. 5(b) and (c)]. In fact, the north- and Wuwei [black rectangular frame in Fig. 6(a) and (b)], the
west of Gansu is a typical arid area with little rainfall, and the precipitation estimated by the UR and MR methods is greatly
growth of green vegetation mainly depends on the irrigation and overestimated. The NDVI, which is one of the most crucial
groundwater (protected farmland). The precipitation is obvi- parameters in the regression model, has a large change from
ously overestimated by global regression methods of the UR 0.25◦ to 1 km in the two sites. The values of NDVI in Zhangye
and MR in the area. The local fitting results of parameters are and Wuwei at 0.25◦ resolution are 0.22 and 0.26 and at 1 km
CHEN et al.: IMPROVED SPATIAL DOWNSCALING PROCEDURE FOR TRMM 4601

Fig. 7. Time series of the coefficients of determination (R2 ) between the


original 0.25◦ TRMM 3B43 monthly data, annual-based GWR fraction disag-
gregation downscaling results, monthly based GWR downscaling results (only
in May, June, and September), and the monthly rain gauges data in 2005, 2007,
and 2009.
Fig. 6. Validation results in 2005 using 24 rain gauges precipitation for three
downscaled TRMM precipitation at 1 km resolution. (a)–(c) UR, MR, and
GWR methods. (d) Original 0.25◦ resolution TRMM precipitation.

TABLE IV
VALIDATION R ESULTS U SING 24 R AIN G AUGES P RECIPITATION FOR
T HREE T YPICAL Y EARS

Fig. 8. Time series of the MAE of the UR, MR, and GWR methods and the
average monthly precipitation in 2005, 2007, and 2009 of the Gansu province.

(large values of Bias, RMSE, and MAE) between the origi-


nal TRMM data and the rain gauges data despite of high R2
shown in Table IV. As shown in Fig. 6(d), the TRMM 3B43
resolution are 0.31 and 0.37, respectively. The mismatch errors
data overestimated the precipitation in the study area. Since
at different scales make the downscaled precipitation overes-
all downscaling procedures are based on the original TRMM
timated. The GWR method, which constructs the robust local
3B43 data, the errors in TRMM 3B43 data would be inevitably
regression function, can reduce the scale mismatch errors to
introduced in the downscaled precipitation. Therefore, the orig-
some extent [Fig. 6(c)]. The GWR method with largest R2 and
inal TRMM 3B43 data need more improvements using more
smallest Bias, RMSE, and MAE is better than the UR and MR
advanced estimation algorithms and integrating more rainfall
methods in three typical years (Table IV). However, consider-
datasets from microwave and geostationary satellite sensors as
ing the annual average precipitation is about 300 mm/year, the
well as ground-based radar and gauge stations.
GWR downscaled precipitation still has quite large errors.
Previous studies [39]–[41] had shown that both temporal
errors (±8% to ±12% per month) and sampling errors (approx-
imately 30%) could be expected in TRMM rainfall estimates. B. Monthly Downscaling Results
Such errors can result in erroneous applications if without 1) Annual-Based Fraction Disaggregation Results: On the
calibration. Therefore, many researchers had made the corre- basis of annual downscaling results, we used the frac-
sponding research [40], [42]. There are significant deviations tion disaggregation method to obtain the monthly TRMM
4602 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 9, SEPTEMBER 2015

TABLE V
R ESULTS OF R EGRESSION A NALYSIS B ETWEEN P RECIPITATION AND AUXILIARY VARIABLES

precipitation in three typical years. Three different annual TABLE VI


downscaled precipitation data, which were obtained by the VALIDATION R ESULTS OF THE M ONTHLY P RECIPITATION DATA
O BTAINED BY THE M ONTHLY-BASED GWR M ETHOD
UR, MR, and GWR methods, were disaggregated to get the
monthly precipitation. Fig. 7 shows the R2 between original
0.25◦ TRMM 3B43 monthly data, annual-based GWR frac-
tion disaggregation downscaling results, monthly based GWR
downscaling results, and the monthly rain gauges data in three
typical years. Considering that most of the precipitation occurs
in the summer months (May–September), we only tested the
monthly based GWR downscaling method in May, July, and
September for each typical year. It was shown that the R2 are
mainly from 0.4 to 0.8, and the low R2 are mostly observed
for months with little rainfall (from November to February of
the following year). The R2 of annual- and monthly based are
basically the same with the results of original 0.25◦ TRMM
3B43 data. The annual-based fraction disaggregation method
has relatively large R2 . It can be seen from Fig. 8 that the
MAE of the annual-based fraction disaggregation downscal- results. The relatively low R2 values indicate the significant
ing results, which based on three different regression methods limitation in TRMM precipitation downscaling by the UR and
(UR, MR, and GWR), point toward the same trend with the MR methods when they are used at monthly scale. The GWR
monthly rainfall. The monthly MAE values of annual-based method based on the local regression may be a good selection to
GWR method are lower than the values of annual-based UR downscale the monthly TRMM 3B43 data in the study area. The
and MR methods, especially in summer months. In other words, NDVI, LST, and Slope were also selected to construct the GWR
annual-based GWR fraction disaggregation downscaling results model. The R2 between the original TRMM 3B43 monthly data
have the smallest errors. It indicated that annual-based GWR and the predictive results range from 0.93 to 0.98 for three
fraction disaggregation method is effective for the downscaling typical years; it indicates an excellent performance in spatial
of the monthly precipitation in the study area. distribution of the monthly based GWR downscaling method in
2) Monthly Based GWR Downscaling Results: The annual- Gansu province of China.
based GWR fraction disaggregation method could effectively Similarly, the monthly based GWR downscaling results were
downscale the monthly TRMM precipitation and produce small also validated using the rain gauges data. The validation results
errors. However, this downscaling procedure, which acquires are shown in Table VI and the results show that the GWR
the monthly fractions at 1 km scale by a simple spline inter- method can downscale the monthly precipitation with relatively
polator, ignores the spatial heterogeneity of precipitation. In higher R2 and lower MAE. However, the original 0.25◦ TRMM
addition, the fraction disaggregation results are based on the 3B43 monthly data also have considerable discrepancies with
downscaling results of annual precipitation, which could cause the rain gauge measurements, which have a significant impact
the accumulation of errors. Therefore, in this paper, we also on the downscaling results. As a whole, the annual-based
tried the monthly based GWR method directly to downscale GWR fraction disaggregation downscaling results have larger
the monthly precipitation data. This monthly based GWR R2 value than the monthly based GWR method. However,
downscaling procedure is similar to the downscaling of annual this method ignores the precipitation spatial heterogeneity,
precipitation. Similarly, we only tested the downscaling results whereas the monthly based GWR method can compensate
in May, July, and September of 2005, 2007, and 2009. The for the deficits. Subjectivity aside, the monthly based GWR
results of regression analysis between precipitation and auxil- method also has some disadvantages. For example, this method
iary variables are shown in Table V. It can be found that the R2 needs suitable auxiliary factors, and these factors would be
of the NDVI, Slope, Longitude, and Latitude are lower than the diverse in different regions and may fail in certain regions.
annual results, but the R2 of the LST are larger than the annual Although the monthly based method has a good performance
CHEN et al.: IMPROVED SPATIAL DOWNSCALING PROCEDURE FOR TRMM 4603

in the study area, it still needs to be verified for different time The original TRMM 3B43 data at 0.25◦ spatial resolution
scales in different regions. Consequently, two monthly based have considerable discrepancies from the rain gauge mea-
downscaling methods mentioned in the paper have their own surements. Since all downscaling procedures are based on the
merits and could be used to downscale the TRMM monthly original TRMM 3B43 data, the accuracy of TRMM 3B43 data
precipitation. The monthly based GWR downscaling method directly determines the accuracy of downscaled precipitation
may be more suitable when the original TRMM data can be no matter what methods are used. Future work should focus on
obtained at high accuracy. the downscaling and calibration simultaneously for obtaining
better precipitation for a given study area. Testing the proposed
GWR downscaling method for higher temporal scales (e.g.,
V. S UMMARY AND C ONCLUSION
weekly and daily) in different climatic regions would be
The objective of this study was to obtain the annual and interesting topics in future studies. It is worth noting that future
monthly precipitation at 1 km resolution with high accuracy in availability of TRMM data is not assured, thus integrating
Gansu, China. To achieve this objective, a new multivariable multisources satellite precipitation data to obtain high accuracy
GWR method was developed to downscale the TRMM 3B43 rainfall at high spatio-temporal resolution would also be a
annual precipitation at 0.25◦ to precipitation at 1 km resolu- popular tendency in future.
tion. The GWR downscaling method was evaluated in Gansu
for three typical precipitation years (average, wet, and dry).
Two other downscaling methods, i.e., UR and MR were also
implemented and compared with the developed GWR method. R EFERENCES
The variables selection procedure was proposed, and appro- [1] G. Langella, A. Basile, A. Bonfante, and F. Terribile, “High-resolution
priate variables were finally determined for three different space–time rainfall analysis using integrated ANN inference systems,” J.
Hydrol., vol. 387, pp. 328–342, 2010.
downscaling methods. Appropriate variable(s) for the UR [2] S. Jia, W. Zhu, A. Lü, and T. Yan, “A statistical spatial downscaling algo-
method was NDVI, for the MR method were six factors, rithm of TRMM precipitation based on NDVI and DEM in the Qaidam
i.e., NDVI, LST, DEM, Slope, Aspect, and Latitude, for the Basin of China,” Remote Sens. Environ., vol. 115, pp. 3069–3079, 2011.
[3] X. Li, Q. Zhang, and C. Xu, “Suitability of the TRMM satellite rainfalls
GWR were multivariable NDVI, LST, and Slope. The measured in driving a distributed hydrological model for water balance computa-
precipitation data from 24 rain gauge stations were used to vali- tions in Xinjiang catchment, Poyang lake basin,” J. Hydrol., vol. 426–427,
date the downscaled precipitation by three methods. Validation pp. 28–38, 2012.
[4] G. Naumann, P. Barbosa, H. Carrao, A. Singleton, and J. Vogt,
results showed that the simplest UR method was feasible for “Monitoring drought conditions and their uncertainties in Africa using
obtaining the 1 km TRMM precipitation to a certain extent. TRMM data,” J. Appl. Meteorol. Climatol., vol. 51, pp. 1867–1874, 2012.
With inclusion of more auxiliary factors, the MR method gener- [5] W. W. Immerzeel, M. M. Rutten, and P. Droogers, “Spatial downscal-
ing of TRMM precipitation using vegetative response on the Iberian
ated better results as expected. Both UR and MR methods were Peninsula,” Remote Sens. Environ., vol. 113, pp. 362–370, 2009.
based on the assumption that global regression relationship is [6] Z. Duan and W. G. M. Bastiaanssen, “First results from Version 7 TRMM
stable in the entire region, and they ignore the spatial hetero- 3B43 precipitation product in combination with a new downscaling–
calibration procedure,” Remote Sens. Environ., vol. 131, pp. 1–13, 2013.
geneity of precipitation, which could easily lead to over-fitting [7] J. E. Hunink, W. W. Immerzeel, and P. Droogers, “A high-resolution
and is more sensitive to scale mismatch errors. In contrast, the precipitation 2-step mapping procedure (HiP2P): Development and appli-
multivariable GWR method can capture the variation by cali- cation to a tropical mountainous area,” Remote Sens. Environ., vol. 140,
pp. 179–188, 2014.
brating spatial-varying multiple regression models at different [8] W. P. Kustas, J. M. Norman, M. C. Anderson, and A. N. French,
locations, thereby being able to detect the space nonstationarity “Estimating subpixel surface temperatures and energy fluxes from the
of precipitation. Validation results clearly demonstrated that the vegetation index–radiometric temperature relationship,” Remote Sens.
Environ., vol. 85, pp. 429–440, 2003.
GWR method performed significantly and consistently better [9] N. Agam, W. P. Kustas, M. C. Anderson, F. Li, and C. M. U. Neale,
than both UR and MR methods for all three typical years. “A vegetation index based technique for spatial sharpening of thermal
Two monthly downscaling strategies were evaluated to imagery,” Remote Sens. Environ., vol. 107, pp. 545–558, 2007.
[10] W. Zhan et al., “Downscaling land surface temperatures with multi-
obtain the monthly precipitation at 1 km resolution; they are the spectral and multi-resolution images,” Int. J. Appl. Earth Observ. Geoinf.,
annual-based fraction disaggregation method and the monthly vol. 18, pp. 23–36, 2012.
based GWR method. It was shown that the annual-based GWR [11] W. Zhan et al., “Disaggregation of remotely sensed land surface tem-
perature: Literature survey, taxonomy, issues, and caveats,” Remote Sens.
fraction disaggregation method was performing better than the Environ., vol. 131, pp. 119–139, 2013.
other method, especially in summer months. However, it faces [12] O. Merlin, A. Al Bitar, J. P. Walker, and Y. Kerr, “A sequential model
two major shortcomings. The first is that the monthly fractions for disaggregating near-surface soil moisture observations using multi-
resolution thermal sensors,” Remote Sens. Environ., vol. 113, pp. 2275–
at 1 km resolution were obtained by a simple spline inter- 2284, 2009.
polation method, which ignores the spatial heterogeneity of [13] M. Piles et al., “A downscaling approach for SMOS land observa-
precipitation; the other one is the monthly downscaling results tions: Evaluation of high-resolution soil moisture maps over the Iberian
Peninsula,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 7,
are on the basis of the annual downscaling results, which could no. 9, pp. 3845–3857, Sep. 2014.
easily lead to the accumulation of errors. The monthly based [14] J. Fang, J. Du, W. Xu, P. Shi, M. Li, and X. Ming, “Spatial downscaling
GWR method producing high R2 and low MAE has an excel- of TRMM precipitation data based on the orographical effect and meteo-
rological conditions in a mountainous area,” Adv. Water Resour., vol. 61,
lent performance in spatial distribution of precipitation in the pp. 42–50, 2013.
study area. Two methods could be used for the downscaling of [15] G. M. Foody, “Geographical weighting as a further refinement to regres-
the TRMM monthly data, and the monthly based GWR method sion modelling: An example focused on the NDVI–rainfall relationship,”
Remote Sens. Environ., vol. 88, pp. 283–293, 2003.
is more suitable for the monthly downscaling when the original [16] W. R. Tobler, “A computer movie simulating urban growth in the Detroit
TRMM 3B43 data can be obtained with high accuracy. region,” Econ. Geogr., vol. 46, pp. 234–240, 1970.
4604 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 9, SEPTEMBER 2015

[17] P. Dutilleul and P. Legendre, “Spatial heterogeneity against heteroscedas- [40] T. Condom, P. Rau, and J. C. Espinoza, “Correction of TRMM 3B43
ticity: An ecological paradigm versus a statistical concept,” Oikos, monthly precipitation data over the mountainous areas of Peru during the
vol. 66, pp. 152–171, 1993. period 1998–2007,” Hydrol. Processes, vol. 25, pp. 1924–1933, Jun. 15,
[18] B. Li, S. Tao, and R. W. Dawson, “Relations between AVHRR NDVI and 2011.
ecoclimatic parameters in China,” Int. J. Remote Sens., vol. 23, pp. 989– [41] M. Almazroui, “Calibration of TRMM rainfall climatology over Saudi
999, 2002. Arabia during 1998–2009,” Atmos. Res., vol. 99, pp. 400–414, 2011.
[19] Z. Zhao, J. Gao, Y. Wang, J. Liu, and S. Li, “Exploring spatially vari- [42] H. Heidinger, C. Yarlequé, A. Posadas, and R. Quiroz, “TRMM rain-
able relationships between NDVI and climatic factors in a transition fall correction over the Andean Plateau using wavelet multi-resolution
zone using geographically weighted regression,” Theor. Appl. Climatol., analysis,” Int. J. Remote Sens., vol. 33, pp. 4583–4602, Jul. 20, 2012.
vol. 120, pp. 507–519, Jun. 12, 2014.
[20] S. Li, Z. Zhao, X. Miaomiao, and Y. Wang, “Investigating spatial
non-stationary and scale-dependent relationships between urban surface Cheng Chen was born in Suqian, Jiangsu, China,
temperature and environmental factors using geographically weighted in 1990. He received the B.Eng. degree in survey-
regression,” Environ. Model. Softw., vol. 25, pp. 1789–1800, 2010. ing and mapping engineering from Wuhan University,
[21] C. Brunsdon, A. S. Fotheringham, and M. E. Charlton, “Geographically Wuhan, China, in 2013. He is currently pursuing the
weighted regression: A method for exploring spatial nonstationarity,” M.S. degree in photogrammetry and remote sensing
Geogr. Anal., vol. 28, pp. 281–298, Oct. 01, 1996. at Nanjing University, Nanjing, China.
[22] W. Song, H. Jia, J. Huang, and Y. Zhang, “A satellite-based geographi- His research interests include land use and land
cally weighted regression model for regional PM2.5 estimation over the cover change, satellite-based precipitation, drought
Pearl River Delta region in China,” Remote Sens. Environ., vol. 154, monitoring, and remote sensing image processing.
pp. 1–7, 2014.
[23] F. Tian, G. Y. Qiu, Y. H. Yang, Y. J. Xiong, and P. Wang, “Studies on the
relationships between land surface temperature and environmental factors
in an inland river catchment based on geographically weighted regression Shuhe Zhao (M’03) received the Ph.D. degree from
and MODIS data,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., Nanjing University, Nanjing, China, in 2003.
vol. 5, no. 3, pp. 687–698, Jun. 2012. From 2003 to 2005, he was a Research Fellow
[24] D. Ivajnšič, M. Kaligarič, and I. Žiberna, “Geographically weighted with the Institute of Remote Sensing & GIS,
regression of the urban heat island of a small city,” Appl. Geogr., vol. 53, Peking University, Beijing, China. He is currently
pp. 341–353, 2014. a Vice Director and Associate Professor with the
[25] F. Chen, Y. Liu, Q. Liu, and X. Li, “Spatial downscaling of TRMM 3B43 Department of Geographic Information Science,
precipitation considering spatial heterogeneity,” Int. J. Remote Sens., School of Geographic and Oceanographic Sciences,
vol. 35, pp. 3074–3093, May 03, 2014. Nanjing University. From 2009 to 2009, he is
[26] G. J. Huffman et al., “The TRMM multisatellite precipitation anal- an Academic Visitor at the School of Geography,
ysis (TMPA): Quasi-global, multiyear, combined-sensor precipitation University of Southampton, Southampton, U.K. From
estimates at fine scales,” J. Hydrometeorol., vol. 8, pp. 38–55, 2007. May 2013 to June 2013, he visited the Department of Environment and
[27] W. Wang, S. Liang, and T. Meyers, “Validating MODIS land surface Geographic Science, University of Cape Town, Cape Town, South Africa.
temperature products using long-term nighttime ground measurements,” From 2013 to 2014, he worked as a Visiting Professor at the School of
Remote Sens. Environ., vol. 112, pp. 623–635, 2008. Civil and Environmental Engineering, Yonsei University, Yonsei, Korea. His
[28] J. Rhee, J. Im, and G. J. Carbone, “Monitoring agricultural drought for research interests include hyperspectral remote sensing, ecological remote sens-
arid and humid regions using multi-sensor remote sensing data,” Remote ing and global change, and application of pattern recognition methods in remote
Sens. Environ., vol. 114, pp. 2875–2887, Dec. 15, 2010. sensing.
[29] L. Y. Jin, J. L. Fu, and F. H. Chen, “Spatial differences of precipitation
over northwest China during the last 44 years and its response to global
warming,” Sci. Geogr. Sin., vol. 25, pp. 565–572, 2005. Zheng Duan received the M.S. degree in cartogra-
[30] S. Kawashima, T. Ishida, M. Minomura, and T. Miwa, “Relations between phy and GIS from the Graduate University of Chinese
surface temperature and air temperature on a local scale during winter Academy of Sciences, Beijing, China, in 2010 and
nights,” J. Appl. Meteorol., vol. 39, pp. 1570–1579, 2000. the Ph.D. degree in remote sensing and hydrologi-
[31] S. E. Nicholson, M. L. Davenport, and A. R. Malo, “A comparison of the cal model from Delft University of Technology, Delft,
vegetation response to rainfall in the Sahel and East Africa, using normal- The Netherlands, in 2014.
ized difference vegetation index from NOAA AVHRR,” Clim. Change, He was a Scientific Consultant with several
vol. 17, pp. 209–241, 1990. institutes including UNESCO-IHE, Delft, The
[32] X. Zhao, K. Tan, S. Zhao, and J. Fang, “Changing climate affects vegeta- Netherlands, in the field of satellite-based estimation
tion growth in the arid region of the northwestern China,” J. Arid Environ., of actual evapotranspiration, soil moisture and
vol. 75, pp. 946–952, 2011. biomass production, and land cover classification.
[33] I. Orlanski, “A rational subdivision of scales for atmospheric processes,” From July 2015, he is an Assistant Professor with the Technische Universität
Bull. Amer. Meteorol. Soc., vol. 56, pp. 527–530, 1975. München, München, Germany. His research interests include integration of
[34] E. Casetti, “Bayesian regression and the expansion method,” Geogr. remote sensing in hydrological model and water balance studies of lakes and
Anal., vol. 24, pp. 58–74, Jan. 01, 1992. catchments, satellite-based precipitation, evapotranspiration, and open water
[35] A. S. Fotheringham, M. Charlton, and C. Brunsdon, “The geography evaporation estimation.
of parameter space: An investigation of spatial non-stationarity,” Int. J.
Geogr. Inf. Syst., vol. 10, pp. 605–627, 1996.
[36] W. Z. Qin, “The basic theoretics and application research on geo- Zhihao Qin received the B.Sc. degree in economic
graphically weighted regression,” Ph.D. dissertation, School of Civil geography from the Department of Geography, Sun
Engineering Dept. Surveying and Geo-informatics, Tongji Univ., Yetsen University, Guangzhou, China, in 1983, and
Shanghai, China, 2007. the Ph.D. degree in remote sensing of arid envi-
[37] P. Rahimzadeh-Bajgiran, K. Omasa, and Y. Shimizu, “Comparative eval- ronment from the Remote Sensing Laboratory, J.
uation of the vegetation dryness index (VDI), the temperature vegetation Blaustein Institute for Desert Researches, Ben Gurion
dryness index (TVDI) and the improved TVDI (iTVDI) for water stress University of the Negev, Israel, in 2001.
detection in semi-arid regions of Iran,” ISPRS J. Photogramm. Remote Currently, he is a Professor of Remote Sensing
Sens., vol. 68, pp. 1–12, 2012. Applications to Agriculture with the Institute of
[38] S. Tanaka, T. Sugimura, and S. Mishima, “Monitoring of vegetation Agro-Resources and Regional Planning, Chinese
extent around Kitui pilot forest (afforestation test site) in Kenya with Academy of Agricultural Sciences, Beijing, China.
rainfall by satellite data,” Adv. Space Res., vol. 26, pp. 1039–1042, 2000. He has authored 150 papers, including the extensively cited one on mono-
[39] S. H. Franchito, V. B. Rao, A. C. Vasques, C. M. E. Santo, and window algorithm for land surface temperature retrieval from Landsat
J. C. Conforte, “Validation of TRMM precipitation radar monthly rain- TM/ETM thermal band data. His research interests include thermal remote
fall estimates over Brazil,” J. Geophys. Res., vol. 114, no. D02, pp. 1–9, sensing, remote sensing and GIS applications to agriculture, agro-disasters
Jan. 23, 2009. monitoring, and regional planning for agricultural development.
Annotations

An Improved Spatial Downscaling Procedure for TRMM 3B43


Precipitation Product Using Geographically Weighted
Regression
Chen, Cheng; Zhao, Shuhe; Duan, Zheng; Qin, Zhihao

01 Marcelo Portuguez Page 2


18/5/2018 23:57

You might also like