You are on page 1of 9

International Journal of Applied Earth Observation and Geoinformation 39 (2015) 40–48

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and


Geoinformation
journal homepage: www.elsevier.com/locate/jag

RNDSI: A ratio normalized difference soil index for remote sensing of


urban/suburban environments
Yingbin Deng a , Changshan Wu b,a,∗ , Miao Li c , Renrong Chen d
a
Department of Geography, University of Wisconsin-Milwaukee, P.O. Box 413, Milwaukee WI 53201-0413, USA
b
Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing 100094, People’s Republic of China
c
Key Laboratory of Remote Sensing Monitoring of Geographic Environment, Harbin Normal University, Heilongjiang 150025, People’s Republic of China
d
School of Geography Science, South China Normal University, Guangzhou 510631, People’s Republic of China

a r t i c l e i n f o a b s t r a c t

Article history: Understanding land use land cover change (LULCC) is a prerequisite for urban planning and environ-
Received 13 March 2014 ment management. For LULCC studies in urban/suburban environments, the abundance and spatial
Accepted 20 February 2015 distributions of bare soil are essential due to its biophysically different properties when compared to
Available online 6 March 2015
anthropologic materials. Soil, however, is very difficult to be identified using remote sensing technologies
majorly due to its complex physical and chemical compositions, as well as the lack of a direct relation-
Keywords:
ship between soil abundance and its spectral signatures. This paper presents an empirical approach to
Soil index
enhance soil information through developing the ratio normalized difference soil index (RNDSI). The first
Ratio normalized difference soil index
(RNDSI)
step involves the generation of random samples of three major land cover types, namely soil, impervious
Biophysical composition index (BCI) surface areas (ISAs), and vegetation. With spectral signatures of these samples, a normalized difference
Enhanced built-up and bareness index soil index (NDSI) was proposed using the combination of bands 7 and 2 of Landsat Thematic Mapper
(EBBI) Image. Finally, a ratio index was developed to further highlight soil covers through dividing the NDSI by
Land use land cover change (LUCC) the first component of tasseled cap transformation (TC1). Qualitative (e.g., frequency histogram and box
charts) and quantitative analyses (e.g., spectral discrimination index and classification accuracy) were
adopted to examine the performance of the developed RNDSI. Analyses of results and comparative anal-
yses with two other relevant indices, biophysical composition index (BCI) and enhanced built-up and
bareness Index (EBBI), indicate that RNDSI is promising in separating soil from ISAs and vegetation, and
can serve as an input to LULCC models.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction erties (e.g., vegetation, impervious surfaces). Specified land cover


information can be highlighted and thereafter separated from other
Land use land cover change (LULCC) plays an important role types by computing the ratio (Major et al., 1990), normalize differ-
in shaping our living environments. While LULCC is associated ence (Gao, 1996; Townshend and Justice, 1986), and difference of
with economic benefits and improved quality of life, it also brings spectral signatures of two bands. Taking vegetation as an exam-
numerous environmental problems to human societies, such as ple, numerous vegetation indices, including normalized difference
urban heat island, biodiversity loss, water, air, and soil pollutions, vegetation index (NDVI) (Becker and Choudhury, 1988), ratio vege-
etc. For examining the intensity and patterns of LULCC, multi- tation index (RVI) (Major et al., 1990), soil adjusted vegetation index
temporal remotely sensed variables (e.g., reflectance, spectral (Huete, 1988) have been developed to examine the condition of
indices) have been typically applied at different geographical scales. vegetation growth. Similarly, for impervious surfaces, normalized
Within these variables, spectral indices, computed using the spec- different built-up index (NDBI) (Zha et al., 2003), normalized differ-
tral signatures of two or more bands of remotely sensed imagery, ence impervious surface index (NDISI) (Xu, 2010), and biophysical
are one of the most convenient means for extracting land prop- composition index (BCI) (Deng and Wu, 2012), have been devel-
oped to examine the quantity and distribution of urban impervious
surfaces. In addition to vegetation and impervious surfaces, soil is
also an important biophysical component in urban/suburban envi-
∗ Corresponding author. Tel.: +1 414 229 4860; fax: +1 414 229 4860. ronments (Raison, 1979). For urban/suburban environments, Ridd
E-mail addresses: yingbin@uwm.edu (Y. Deng), cswu@uwm.edu (C. Wu), (1995) developed a conceptual vegetation-impervious surface-soil
dkylimiao@gmail.com (M. Li), renrong.chen@gmail.com (R. Chen). (V-I-S) model to characterize dynamic urban land use land covers.

http://dx.doi.org/10.1016/j.jag.2015.02.010
0303-2434/© 2015 Elsevier B.V. All rights reserved.
Y. Deng et al. / International Journal of Applied Earth Observation and Geoinformation 39 (2015) 40–48 41

Besides, soil is a key parameter in examining agriculture produc- software. Atmospheric correction was not applied to the image
tion (Magdoff and Weil, 2004), hydrological process (Zhang et al., due to generally good weather condition. In order to examine the
2002), and sandstorm statues (Yang et al., 2010). performance of RNDSI, we employed digital orthophoto quarter
The developments of soil indices, however, are difficult due quadrangle (DOQQ) images of Waukesha and Milwaukee (acquired
to several reasons. First, soil is a complex material with various on April 13, 2000 and March 30, 2002 with a spatial resolution of
chemical and physical compositions, and the spectra of soil are 0.61 m) as the reference data. With the Landsat TM5 reflectance
very complex that prevents direct connections between soil prop- image, water pixels were identified and removed following three
erties and its spectral responses (Ben-Dor, 2002). The formation steps, including (1) identifying training samples of water, urban,
of an efficient spectral index always relies on the identification forest, grass, and soil land cover types, (2) performing a support
of spectral bands with distinct spectral properties (Mahlein et al., vector machine classification using these training samples, and
2013). Such characteristically spectral signatures, however, cannot (3) removing identified water pixels from the image. Finally, all
be found in the spectra of soil (Ge et al., 2011). Second, soil’s spec- the images were re-projected to the universal transverse mercator
tral signatures are varied with its construction, texture, moisture, (UTM) with zone 16 and datum WGS84.
color, and surface roughness (Lõhmus et al., 1989), and significantly
different spectral signatures may exist among different types of
3. Methods
soil. For example, moist soil has a similar spectral signatures when
compared to those of water, shadow, and dark urban impervious
3.1. Construction of Ratio normalized difference soil index
surfaces, majorly due to the large amount of water content in the
(RNDSI)
soil (Wu, 2004). Comparatively, dry bare soil’s spectral signatures
are comparable to those of bright urban impervious surfaces and
3.1.1. Generation of stratified random samples
bare rocks (Weng and Lu, 2008). As a result, the confusions between
Due to the complexity of soil compositions and their cor-
moist soil and water, shadow, and dark impervious surfaces, as well
responding spectral signatures, direct connections between soil
as between dry bare soil and bright urban impervious surfaces and
properties and their spectral responses could not be constructed.
bare rock are always considered as essential problems, especially
To address this problem, we developed an empirical approach to
for analyzing urban and suburban environments.
formulate the soil index with samples chosen through applying a
To address the aforementioned problems, this paper proposed a
stratified random sampling strategy. The selection of random sam-
three-step empirical approach to generate a soil index, named ratio
ples is essential for constructing an effective index. Theoretically, it
normalized difference soil index (RNDSI). In particular, the first step
would be ideal to select a large number of samples representing all
involves the generations of random samples of three major land
land use land cover types in the study area. It is, however, impracti-
cover types of urban/suburban areas, e.g., vegetation, impervious
cal due to its prohibitive cost (Van Genderen and Lock, 1977). On the
surfaces, and soil. With the spectral signatures of these samples,
other hand, a very small number of samples may lead to a reduced
we empirically examined a number of spectral variables for devel-
reliability for representing the most typical land cover classes (Gao,
oping normalized difference indices, and identified the one that
2008). As a compromise of efficiency and reliability, we generated
can highlight soil covers. The last step involves the construction of
two sample sets (training and testing) using the stratified random
a ratio index to further highlight soil information. This empirically
sampling strategy. The training set includes 390 samples (soil: 130,
created index can be considered as an integration of the normalized
vegetation: 130, impervious surface: 130), and the testing set con-
difference index and the ratio index. The remainder of this paper is
sists 1397 samples (soil: 400, vegetation: 554, impervious surface:
organized as follows. The next section introduces the study area and
422). We intentionally chose a larger number of testing samples to
data sources. Section 3 describes the methods applied for develop-
comprehensively examine the effectiveness of the developed soil
ing the RNDSI, as well as comparative analyses with other relevant
index. The validity of all samples was carefully checked through
indices. Results and comparative analyses are reported in Section
visualizing both TM5 and DOQQ images to avoid mis-labeling and
4. Finally, discussion and conclusions are provided in Section 5 and
mixing-pixel problems. The mean and standard deviation of the
Section 6.
spectral reflectance of impervious surface area (ISA), vegetation,
and soil, are illustrated for all training samples (see Fig. 2).
2. Study area and data
3.1.2. Development of normalized difference soil index (NDSI)
Two counties of Wisconsin, Milwaukee and Waukesha, in the With the spectra of training samples of ISA, vegetation, and soil,
United States were chosen as the study area. These two counties we attempted to develop a normalized difference index to enhance
are located in the Great Lake Region (Fig. 1) with a humid conti- soil information. A normalized difference index is typically applied
nental climate. Milwaukee and Waukesha have geographic areas to highlight the difference between the strongest and weakest spec-
of 607 km2 and 2058 km2 and are with total population of 947,735 tral responses of a target land cover. NDVI, as an example, was
and 389,891 respectively. Within these two counties, Milwaukee developed to highlight the difference of the spectral responses of
is majorly covered by urban and suburban land uses (e.g., com- vegetation at the red and near infrared bands. Unfortunately, such
mercial, industrial, residential etc.), and Waukesha is dominated characteristically strong or weak spectral responses do not exist for
by suburban and rural lands (e.g., farmland and forest). This study soil due to its complex physical and chemical compositions, as well
area contains a large amount of bare soil (e.g., silt loam), impervious as regional differences. Therefore, the first task is to find out the
surfaces, and vegetation, and is an ideal site for examining the effec- band combinations to construct a normalized difference index for
tiveness of the proposed RNDSI in separating soil from impervious soil covers. Through analyzing the spectral reflectance of soil sam-
surfaces and vegetated covers. ples (see Fig. 2), we found that the mean reflectance values of bands
A scene of Landsat TM5 image acquired on September 11, 2001 4, 5, and 7 are likely to be higher than those of bands 1, 2 and 3,
was employed for developing the RNDSI. The image has six spec- and it might be feasible to construct a normalized difference index
tral bands (except the thermal band) with a spatial resolution of through selecting one band with higher reflectance and the other
30 m. Digital numbers (DNs) of the image were converted into with lower reflectance. Further visual examinations (Fig. 3) indicate
at-satellite reflectance using the Landsat calibration model pro- that, among all possible combinations, the normalized differences
vided by ENVI, a commercial remote sensing image processing of bands 5 and 4, bands 7 and 1, and bands 7 and 2 are generally
42 Y. Deng et al. / International Journal of Applied Earth Observation and Geoinformation 39 (2015) 40–48

Fig. 1. Study area, Waukesha, and Milwaukee. R: b7, G: b5, B: b2.

higher for soil samples, and significantly lower for ISA and vegeta- Where Band7 and Band2 represent the reflectance of TM band 7
tion samples. Therefore, their ability of enhancing soil information (middle infrared) and band 2 (green) respectively.
merits further analyses. Fig. 3 also indicates that the normalized dif-
ference indices calculated using the combinations of bands 7 and
3.1.3. Construction of ratio normalized difference soil index
1, and bands 7 and 2, can better separate soil samples from ISA and
(RNDSI)
vegetation samples. Finally, through a careful examination of these
To further alleviate the confusion between soil and other
two band combinations, we found out that the index derived using
land covers, especially ISAs, we constructed a ratio normalized
bands 7 and 2 combination serves as a better alternative as it can
difference soil index (RNDSI), with which the calculated NDSI
adequately enhance the soil information and greatly suppress the
value was employed as the numerator, and another variable as
ISA and vegetation values (see Fig. 3). Hence, bands 7 and 2 were
the denominator. To find the right variable as the denominator,
chosen as the best-fit model to compute the NDSI. The formulation
we carefully examined the spectral characteristics of ISA and soil
of NDSI is expressed in Eq. (1).
samples (see Fig. 2), and attempted a few spectral transformation
techniques, including maximum noise fraction (MNF) and tasseled
cap (TC) transformations.TC is an orthogonal transformation which
(Band7 − Band2) converts the original data into a new dimensional space. The first
NDSI = (1)
(Band7 + Band2) three components of TC transformation are considered to represent
Y. Deng et al. / International Journal of Applied Earth Observation and Geoinformation 39 (2015) 40–48 43

brightness, greenness, and wetness respectively. Through a com-


parative analysis (see Fig. 4), we found that two ratio indices,
NDSI/TC1 (first component of tasseled cap transformation) and
NDSI/TM4(fourth band of TM image) are likely to have a better
performance for highlighting soil information. Further, when these
two indices were examined, we found that the performance of
NDSI/TC1 is better with lower standard deviation (SD) values for
ISA and soil (e.g., ISA: 0.36, Soil: 0.88) when compared to those of
NDSI/TM4 (ISA: 0.62, soil: 1.44), indicating soil can be better sepa-
rated with NDSI/TC1. Therefore, TC1 could be a viable option as the
denominator to further highlight soil information. Noticing that
bright ISAs are with much higher values in TC1, it is reasonable to
apply TC1 as the denominator to highlight soil information, as well
as suppress the values associated with ISAs. Note that we employed
Fig. 2. Spectral reflectance (mean and standard deviation) of ISA, vegetation, and
soil.

Potential band combinations for calculating NDSI


0.8
ISA Vegetation Soil

0.6

0.4

0.2

0
(b5-b1)/(b5+b1)

(b5-b2)/(b5+b2)

(b5-b3)/(b5+b3)

(b5-b4)/(b5+b4)

(b5-b7)/(b5+b7)

(b4-b1)/(b4+b1)

(b4-b2)/(b4+b2)

(b4-b3)/(b4+b3)

(b4-b7)/(b4+b7)

(b7-b1)/(b7+b1)

(b7-b2)/(b7+b2)

(b7-b3)/(b7+b3)

(b1-b2)/(b1+b2)

(b1-b3)/(b1+b3)

(b3-b2)/(b3+b2)
-0.2

-0.4

Fig. 3. Potential band combinations for calculating NDSI.

Potential band combinations for calculating RNDSI


ISA Vegetation Soil
6

0
NDSI/MNF1 NDSI/BCI NDSI/TM1 NDSI/TM2 NDSI/TM3 NDSI/TM4 NDSI/TC1 NDSI/TC2 NDSI/TC3

Fig. 4. Potential band combinations for calculating RNDSI.


44 Y. Deng et al. / International Journal of Applied Earth Observation and Geoinformation 39 (2015) 40–48

the normalized NDSI (NNDSI) and normalized TC1 (NTC1) for the Table 1
Results of SDI.
calculation of the RNDSI. The derived RNDSI can be expressed as:
SDI RNDSI BCI EBBI
NNDSI
RNDSI = (2)
NTCI ISA-S 0.85 0.77 0.24
S-V 0.97 1.53 3.06
where
ISA-S means the separability between impervious surface and soil; S-V means the
NDSI − NDSImin separability between soil and vegetation. Because we mainly focused on relation-
NNDSI =
NDSImax − NDSImin ship between soil and other land use types, separability between vegetation and
impervious surface was not discussed in this paper.
TCl − TClmin
NTCI =
TClmax − TClmin
RNDSI through visual analysis and quantitative assessments. In par-
ticular, samples’ histogram plots and resultant index images were
3.2. Performance assessment employed for visual analysis, and SDI and classification accuracy
measures were utilized as quantitative assessments to confirm the
To examine the performance of RNDSI in separating soil from results obtained from the visual analyses.
other land covers in urban/suburban environments, two mea-
surements, spectral discrimination index (SDI) and classification
4. Results
accuracy, were calculated and applied to the study area. SDI is an
index that measures the degree of separation between the spec-
4.1. Development of RNDSI
tra of two land cover types (Radeloff et al., 1999). Specifically, it
is calculated through dividing the differences of the mean index
Through applying Eq. (1), the normalized difference soil index
values of two land cover types by the summation of their standard
(NDSI) was calculated using the combination of bands 7 and 2 of
deviations (see Eq. (3))
the TM imagery. Results (see Fig. 5A) indicate that NDSI can, to a
|1 − 2 | certain degree, separate soil with other land cover types. In partic-
SDI = (3)
1 + 2 ular, the values of NDSI for soil are clustered in the range of 0.1 to
where 1 and 2 are the mean values of indices representing land 0.4, while ISA and vegetation are with lower NDSI values. There-
cover types 1 and 2; and 1 and 2 are the standard deviations of fore, with NDSI, soil information has been enhanced and could
the indices associated with land cover types 1 and 2 respectively. be, to some degree, separated from vegetation and ISA. Further,
A higher value of SDI indicates these two land cover types can be RNDSI was derived through dividing the normalized NDSI by the
effectively separated, while a lower value indicates the existence normalized TC1 component (see Eq. (2)). Visual assessments of the
of spectral confusion. As a rule of thumb, an SDI of greater than 1 resulting RNDSI image for the study area (Fig. 6B) and the boxplot
indicates satisfactory separability. In addition to SDI, we also imple- of RNDSI values of soil, vegetation, and ISA testing samples (see
mented a decision tree (DT) classification approach using RNDSI as Fig. 8A) indicate that RNDSI is with reasonably satisfactory perfor-
the single explanatory variable. The DT classification was imple- mances. Specifically, the RNDSI image (see Fig. 6B) indicates that
mented using See5 program developed by Quinlan (1997), and two higher values can be found in the western portion of the study area,
classes, soil and non-soil, were identified. We consider the resul- corresponding to bare soil patches in farmland. In urban areas (e.g.,
tant classification accuracy (e.g., overall accuracy and kappa index) Milwaukee city in the eastern portion of the study area), the RNDSI
and decision rules may shed lights on the performance of RNDSI in values are significantly lower than those in farmlands. For a bet-
separating soil and other land covers. ter illustration, three image subsets, two farmland areas with bare
soil and one commercial area with pure ISAs, are shown in Fig. 7.
3.3. Comparative analysis Through visualizing these three subsets, we can find that the RNDSI
values of bare soil are likely to be much higher than those of urban
In order to better examine the performance of RNDSI, we ISA (see Fig. 7B). This observation is confirmed by the analysis of
also conducted a comparative analysis. Two soil related spectral boxplots of RNDSI (see Fig. 8A), which indicates that the RNDSI val-
indices: biophysical composition index (BCI) (Deng and Wu, 2012), ues of soil samples are generally higher than approximately 2.3,
enhanced built-up and bareness Index (EBBI) (As-Syakur et al., with a mean of 2.9 but a relatively large standard deviation (about
2012), were implemented for the comparative analysis. These two 1.4). Comparatively, ISAs are associated with much lower RNDSI
indices, in some degree, can also enhance soil information. These values, with a mean of 1.5–1.6 and a standard deviation of 0.8. For
indices are expressed as follows. vegetation, it has a similar mean value of RNDSI when compared
to ISA, although with a much smaller standard deviation (approx-
(H + L)/2 − V imately 0.1). As a summary, visual analyses of the resultant RNDSI
BCI = (4)
(H + L)/2 + V image and the associated boxplots of samples indicate that RNDSI
with is promising in separating soil from the other two land cover types.
In addition to visual examination, we also performed quanti-
TC1 − TC1min
H= tative analyses using SDI and classification accuracy as indicators.
TC1max − TC1min
Results show that the SDI values (see Table 1) are 0.85 and 0.97
respectively for the pairs of ISAs and soil, and soil and vegetation.
TC2 − TC2min
V= With a value of 0.85 (less than 1), we acknowledged that the sep-
TC2max − TC2min
arability between soil and ISA is still less than desired. However,
as it is always very difficult to differentiate soil and ISAs from
Band5 − Band4
EBBI = √ (5) remote sensing images, we consider an SDI value approaching one
10 Band5 + Band6 as acceptable. Further, the results of DT classification confirmed
where Band i represents the reflectance of the ith band of the TM this observation (see Table 2 and Fig. 9). With RNDSI as the single
image, and TCi(i = 1,2,3) indicates the ith component of TC transfor- variable, soil class has been separated with other land covers
mation. The performances of these indices were compared to that of with an RNDSI value greater than 2.3, and the resultant overall
Y. Deng et al. / International Journal of Applied Earth Observation and Geoinformation 39 (2015) 40–48 45

Fig. 5. Frequency Histogram of (A) NDSI, (B) RNDSI, (C) BCI, and (D) EBBI.

Fig. 6. Images of (A) original Landsat TM, (B) RNDSI, (C) BCI, and (D) EBBI.
46 Y. Deng et al. / International Journal of Applied Earth Observation and Geoinformation 39 (2015) 40–48

Fig. 7. Comparisons of (A) reflectance TM image, (B) RNDSI, (C) BCI, and (D) EBBI in farmlands with bare soil (P1 and P2) and commercial areas (P3).

Fig. 8. Box chart of testing samples.


Y. Deng et al. / International Journal of Applied Earth Observation and Geoinformation 39 (2015) 40–48 47

Table 2 pared to RNDSI. The purpose of BCI, however, is to enhance urban


Decision rules for soil using the DT classification.
ISAs, instead of soil.
Indices RNDSI BCI EBBI

Soil thresholds Value > 2.2653 0.0693 > Value > −0.0712 Value > 0.0002 4.2.2. EBBI

As EBBI was developed to highlight both built-up (e.g., ISAs) and


bareness (e.g., bare soil) (see Eq. (5)), it appears that both bare soil
and ISAs are with very high values (see Fig. 5D, Fig. 6D, Fig. 7D,
and Fig. 8C), while vegetated areas are with very low values. This
observation was further proved by their low SDI values for ISA and
soil (e.g., 0.24 for EBBI), high SDI value for soil and vegetation (e.g.,
3.06 for EBBI). Finally, the classification accuracy of soil with EBBI
(see Fig. 9) is 79.65%. Such a relative high accuracy may be due to
its strong ability in separating vegetation from ISAs and bare soil.

5. Discussion

Due to convenience of spectral indices in highlighting a partic-


ular land cover type, a large number of indices, such as Normalized
Fig. 9. Comparisons of classification accuracy and kappa coefficients using RNDSI, Difference Vegetation Index (NDVI), Normalized Difference Water
BCI, and EBBI as variables.
Index (NDWI), Normalized Difference Impervious surface Index
(NDISI), Ratio Vegetation Index (RVI), have been developed in
classification accuracy is 95.49% with a kappa index of 0.8917. This the past decades. Few indices, however, have been developed to
result indicates that RNDSI has successfully separated soil from enhance soil information directly. A major reason is associated with
other land cover types with a relatively high precision. the complexity of soil properties and their spectra, as the moisture,
texture, and physical and chemical compositions vary from region
4.2. Comparative analyses with other indices to region, and a direct association between soil and reflectance
spectra is difficult. Fortunately, the proposed empirical method
In order to further examine the effectiveness of RNDSI, two may serve as a better alternative to address this difficulty. With the
relevant indices, BCI, and EBBI, were also implemented, and the developed approach, a different ratio normalized difference index
aforementioned visual and quantitative indices were applied for may be chosen to effectively avoid the spectral variability prob-
an objective comparison. For visual comparisons, the resultant BCI lem in various geographic regions. In the study area with a humid
and EBBI images and their boxplots are shown in Figs. 6–8, respec- continental climate, the RNDSI was constructed based on the com-
tively. Due to the differences of value ranges of these indices, it binations of bands 7 and 2, as well as the TC1. As aforementioned,
is less meaningful to simply compare their values. Instead, visual the selections of NDSI and TC1 are based on comparative analy-
comparisons of the index images of the entire study area (Fig. 6), ses with numerous potential band combinations. Further, when
typical agricultural and urban areas (Fig. 7), and boxplots of the applied to an arid or semi-arid region, as an example, the resultant
indices for testing samples (Fig. 8) would be helpful in judging their RNDSI with the band combinations may not be satisfactory. Never-
abilities in separating soil covers from other types. theless, with such an empirical method, it is very likely to generate
a site specific and well-performed RNDSI index as this method
4.2.1. BCI searches for optimal band combinations. We do acknowledge that
the developed RNDSI lacks physical meanings to differentiate the
Through applying Eq. (4), BCI images for the entire study area spectra of soil, ISAs, and vegetation. This is, nevertheless, a con-
and typical farmlands and urban areas are shown in Fig. 6C and venient and straightforward approach to create a soil index for a
Fig. 7C, and boxplots of training and testing samples of soil, ISAs specific geographic area. Obviously, when compared to other two
and vegetation are illustrated in Fig. 5C and Fig. 8B respectively. soil-related indexes, RNDSI was the only one that could highlight
Analysis of BCI images (Fig. 6C and Fig. 7C) shows that urban ISAs soil information and be separated from ISAs to a certain degree.
are with the highest values; bare soil patches in the farmlands area Although its performance was acceptable when applied to the
with the medium values; and vegetated areas are with the lowest study area, RNDSI still has limitations. First, RNDSI is able to effec-
BCI values. Further, visual examinations of the boxplots (Fig. 8C) tively identify bare soil located in farmland or in forest, but its
show that urban ISAs are with the highest BCI values, with a mean performance is poor when applied to the sandy areas (e.g., con-
approximately 0.12, and a standard deviation of 0.10. In addition, struction sites, mining industry, beaches, etc.), majorly due to the
BCI values of soil covers have a mean of −0.02, and a standard devia- high-degree spectral similarity between sand and bright ISAs. In
tion of 0.09. Finally, vegetation covers are with the lowest BCI values most cases, reflectance is rather high in sandy areas when com-
(e.g., mean of −0.2 and standard deviation of 0.02). In summary, as pared to that of bright ISAs, such as concrete, glass, etc. Therefore,
BCI is designed to highlight urban ISAs, instead of soil, it is reason- RNDSI values of sandy soil may be relatively low and comparable
able to have relative lower values for bare soil. Further, quantitative to those of ISAs. This has led to the relatively low value of SDI in
analyses of the performance of BCI confirmed the above observa- separating soil and ISAs in the study area. Second, ISAs and vegeta-
tions. Especially, the SDI value of BCI for separating bare soil and tion could not be separated with RNDSI. We do not consider this is a
ISA is 0.77, slightly lower than that of the RNDSI. BCI, however, per- major concern as vegetation indices, such as NDVI, could solve this
forms much better in separating soil and vegetation, with a value problem without any difficulty. Finally, the effectiveness of RNDSI
of 1.53. Further, the DT classification analysis (Table 2) indicated is still needed to be tested using a variety of remote sensing imagery
that bare soil can be identified with a BCI value between −0.0712 at different study sites. In this paper, only Landsat TM5 data were
and 0.0693, and the overall classification accuracy is 94.69% with used to examine the performance of RNDSI, and it may be more
a kappa index of 0.8665. In summary, analysis of results indicates conclusive to examine its applicability with other remote sensing
that BCI has a comparable (slightly worse) performance when com- imagery such as SPOT, Hyperion, MODIS, etc.
48 Y. Deng et al. / International Journal of Applied Earth Observation and Geoinformation 39 (2015) 40–48

6. Conclusions Research Committee Award of University of Wisconsin-Milwaukee.


The authors would like to thank anonymous reviewers for their
As a major biophysical component of urban/suburban environ- constructive comments.
ments, soil abundance and spatial distribution are essential for
LULC studies. Soil indices may serve as a convenient means of References
extracting soil information from remotely sensed imagery. The
development of soil indices, however, is extremely difficult due As-Syakur, A.R., Adnyana, I., Arthana, I.W., Nuarsa, I.W., 2012. Enhanced built-up
and bareness index (ebbi) for mapping built-up and bare land in an urban area.
to the complexity of soil spectra, and their dependency on soil’s Remote Sens. 4, 2957–2970.
chemical and physical compositions, construction, texture, mois- Becker, F., Choudhury, B.J., 1988. Relative sensitivity of normalized difference
ture, color, and surface roughness. In order to address this issue, vegetation index (NDVI) and microwave polarization difference index (MPDI)
for vegetation and desertification monitoring. Remote Sens, Environ. 24,
this paper developed an empirical approach to formulate the RNDSI 297–311.
index, and examined its ability in separating soil covers from ISAs Ben-Dor, E., 2002. Quantitative remote sensing of soil properties. Adv. Agron. 75,
and vegetation covers obtained from the Landsat TM imagery. Anal- 173–243.
Deng, C., Wu, C., 2012. BCI: a biophysical composition index for remote sensing of
yses of results and comparative analyses with two other relevant
urban environments. Remote Sens.Environ. 127, 247–259.
indices, BCI and EBBI, suggest several conclusions. Gao, B.-C., 1996. NDWI–a normalized difference water index for remote sensing of
First, although without a direct association between soil prop- vegetation liquid water from space. Remote Sens. Environ. 58, 257–266.
erties and their spectral signatures, an empirical approach with Gao, J., 2008. Digital analysis of remotely sensed imagery. McGraw-Hill
Professional.
randomly selected samples could serve as an alternative for devel- Ge, Y., Thomasson, J.A., Sui, R., 2011. Remote sensing of soil properties in precision
oping the soil index. Using an integrated normalized difference and agriculture: a review. Fron. Earth Sci. 5, 229–238.
ratio index formulation, we found that the combination of Landsat Huete, A., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25,
295–309.
TM bands 2 and 7 can highlight soil information, and the ratio index Lõhmus, K., Oja, T., Lasn, R., 1989. Specific root area: a soil characteristic. Plant Soil
using TC1 as the denominator could further highlight soil informa- 119, 245–249.
tion and suppress noises from other land covers. With this empirical Magdoff, F., Weil, R.R., 2004. Soil organic matter in sustainable agriculture. CRC
Press.
approach, it is feasible to develop the ratio normalized difference Mahlein, A.-K., Rumpf, T., Welke, P., Dehne, H.-W., Plümer, L., Steiner, U., Oerke,
soil index (RNDSI) to highlight soil information. E.-C., 2013. Development of spectral indices for detecting and identifying plant
Second, qualitative and quantitative analyses indicate that diseases. Remote Sens.Environ. 128, 21–30.
Major, D., Baret, F., Guyot, G., 1990. A ratio vegetation index adjusted for soil
RNDSI is promising in enhancing soil information, as well as sup- brightness. Int. J. Remote Sens. 11, 727–740.
pressing noises from other land covers. With RNDSI, soil patches in Radeloff, V.C., Mladenoff, D.J., Boyce, M.S., 1999. Detecting jack pine budworm
the study area were clearly highlighted, with a mean value of 2.9, defoliation using spectral mixture analysis: separating effects from
determinants. Remote Sens. Environ. 69, 156–169.
and a standard deviation of 1.4. Comparatively, pure ISAs are with
Raison, R.J., 1979. Modification of the soil environment by vegetation fires, with
much lower mean values of RNDSI (e.g., 1.5–1.6), and a standard particular reference to nitrogen transformations: a review. Plant Soil 51,
deviation of 0.8. Furthermore, quantitative analyses indicate that 73–108.
RNDSI performs relatively well in separating from ISAs, with an Townshend, J.R., Justice, C., 1986. Analysis of the dynamics of African vegetation
using the normalized difference vegetation index. Int. J. Remote Sens. 7,
SDI of 0.85 and classification accuracy of 95.49%. Finally, compar- 1435–1445.
ative analyses with two relevant indices, BCI and EBBI, confirmed Van Genderen, J.L., Lock, B.F., 1977. Testing land-use map accuracy. Photogramm.
the better performance of RNDSI in separating soil from ISAs and Eng. Remote Sens 43, 1135–1137.
Weng, Q., Lu, D., 2008. A sub-pixel analysis of urbanization effect on land surface
vegetation. temperature and its interplay with impervious surface and vegetation coverage
Admittedly, although RNDSI could separate soil from ISAs and in Indianapolis, United States. Int. J. Appl. Earth Obs. Geoinf. 10, 68–83.
vegetation, some problems exist. First, the ability of RNDSI in sep- Wu, C., 2004. Normalized spectral mixture analysis for monitoring urban
composition using ETM+ imagery. Remote Sens. Environ. 93, 480–492.
arating sandy soil and ISAs is questionable, and we did not test its Xu, H., 2010. Analysis of impervious surface and its impact on urban heat
performance in separating moist soil and dark ISAs as we could environment using the Normalized Difference Impervious Surface Index
not locate moist soil in the study area. Second, the performance of (NDISI). Photogramm. Eng. Remote Sens. 76, 557–565.
Yang, L., Zhang, F., Gao, Q., Mao, R., Liu, X., 2010. Impact of land-use types on soil
RNDSI has not been examined in other study areas and/or with nitrogen net mineralization in the sandstorm and water source area of Beijing,
other remote sensing imagery. As an empirically derived index, China. Catena 82, 15–22.
we do expect minor performance variations of RNDSI in different Zha, Y., Gao, J., Ni, S., 2003. Use of normalized difference built-up index in
automatically mapping urban areas from TM imagery. Int. J. Remote Sens. 24,
spatial locations. Therefore, a comprehensive examination of the
583–594.
RNDSI is necessary and will be one of our future research directions. Zhang, Y., Li, C., Trettin, C.C., Li, H., Sun, G., 2002. An integrated model of soil,
hydrology, and vegetation for carbon dynamics in wetland ecosystems. Global
Biogeochem. Cycles 16, 9-1–9-17.
Acknowledgements

This research was partially supported by the One-hundred Tal-


ent Award of Chinese Academy of Sciences and the Graduate School