- Email: [email protected]
Estimation of leaf water content in cotton by means of hyperspectral indices
Computers and Electronics in Agriculture 90 (2013) 144–151
Contents lists available at SciVerse ScienceDirect
Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag
Estimation of leaf water content in cotton by means of hyperspectral indices Qiu-xiang Yi ⇑, An-ming Bao, Qiang Wang, Jin Zhao Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, 830011 Urumqi, China
a r t i c l e
i n f o
Article history: Received 8 May 2012 Received in revised form 18 September 2012 Accepted 28 September 2012
Keywords: Equivalent water content Fuel moisture content Ratio type of vegetation index Normalized difference type of vegetation index Two-band combination
a b s t r a c t The knowledge of vegetation water conditions can contribute to drought assessment. Remote sensing has a proven ability to assess vegetation properties. In this study, all two-band combinations (350–2500 nm) in the ratio type of vegetation index (RVI) and the normalized difference type of vegetation index (NDVI) were performed on cotton leaf raw spectral reflectance (R) and the first derivative reflectance (DR). The correlation coefficient (r) between all two-band combinations and two leaf water parameters (EWT: equivalent water thickness, and FMC: fuel moisture content) were determined, and the results of this comprehensive analysis were presented by matrix plots. Band centers (k1 and k2) and band widths (Dk1 and Dk2) that combine to form the best indices were identified for EWT and FMC through matrix plots. Then the evaluation of the predictive power of three predictors, i.e. single narrow band reflectance, the widely used published water indices and the best band combination indices, were performed. The results shown that the new indices DR1647/DR1133 and DR1653/DR1687, proposed by two-band combinations, were considered as the optimal indices for EWT and FMC estimation, respectively. The models based on these two best combination indices could explain 58% and 67% variability in EWT and FMC, respectively. Besides, bands with center wavelengths in region from 950 nm to 1100 nm, and 1650 nm to 1750 nm were represented almost all selected bands. The study should further our understanding of the relationships between leaf water content and hyperspectral reflectance. Ó 2012 Elsevier B.V. All rights reserved.
1. Introduction The knowledge of vegetation water conditions can in fact contribute to detect vegetation physiological status (Carter, 1993; Peñuelas et al., 1994; Stimson et al., 2005), to provide useful information in agriculture for irrigation decisions and drought assessment (Peñuelas et al., 1993, 1994) and it is important in forestry in determining fire susceptibility (Carlson and Burgan, 2003; Chuvieco et al., 2004; Ustin et al., 1998). Remote sensing offers an important opportunity for quantitative assessment of vegetation properties at different scales. Together with other parameters, vegetation water content is an important property that can be investigated by using remotely sensed data. With the development of hyperspectral remote sensing technique, direct detection approaches of vegetation water content have been proposed and further used to evaluate crop drought condition. The main parameters describing the amount of water in vegetation that are usually investigated by remote sensing are fuel moisture content (FMC, %) and leaf equivalent water thickness (EWT, cm). FMC defined as the proportion of water over the vegetation dry mass (Burgan, 1996), and EWT is the ratio between the quantity of water and leaf area (Danson et al., 1992). EWT and FMC pro⇑ Corresponding author. Tel.: +86 (0)9917823131. E-mail address: [email protected] (Q.-x. Yi). 0168-1699/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.compag.2012.09.011
vide information on the amount of water present in vegetation. The possibility of estimating FMC and EWT by means of remotely sensed data derives from the fact that water absorbs radiant energy throughout the near-infrared (750–1300 nm) and short-infrared (1300–2500 nm) spectral regions. Depending on tissue water content, reflectance is thus reduced to a varying extent within the water absorption features centered on 970, 1200, 1450, 1940, and 2500 nm (Knipling, 1970; Thomas et al., 1971; Tucker, 1980) and these changes can be recognized and quantified as water content variations. Our study investigates the potential of retrieving both of these measurements via leaf reflectance. Several hyperspectral indices such as the Normalized Difference Infrared Index (NDII), the Water Index (WI) have been established based on hyperspectral remote sensing data. The most widely used ratio type of vegetation index and normalized difference type of vegetation index at leaf level were summarized in Table 1. The results obtained in these studies indicate that knowledge of connection between the investigated variable and the spectral data can improve the performance of the vegetation indices (Thenkabail et al., 2000). Further improvement in indices is generally possible through use of spectral data from distinct narrow bands or through corrections for soil background effects, and further, through different band combinations. Selection of new wavebands in hyperspectral data has been performed in some cases, mainly focusing on how to increase the sensitivity of the vegetation indices to
145
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151 Table 1 Correlation analysis between published water indices and water content (n = 113). Water indices NDII NDII1 NDWI2130 NDWI1240 NDWI1260 NDVI WI WBI FWBI SRWI1 SRWI2 SRWI3 SRWI4 MSI MSI1 MSI1 SIWSI * **
Name Normalized different infrared index Normalized different infrared index Normalized different water index Normalized different water index Normalized different water index Normalized different vegetation index Water index Water band index Floating position water band Simple ratio water index Simple ratio water index Simple ratio water index Simple ratio water index Moisture stress index Moisture stress index Moisture stress index Shortwave infrared water stress
Formula
Reference
(q850 q1650)/(q850 + q1650) (q835 q1650)/(q835 + q1650) (q858 q2130)/(q858 + q2130) (q860 q1240)/(q860 + q1240) (q870 q1260)/(q870 + q1260) (q858 q648)/(q858 + q648) q900/q970 q970/q900 q900/min(q930–980) q858/q1240 q1070/q1340 q678/q1070 q880/q1265 q1600/q820 q870/q1350 q1650/q835 (q858 q1640)/(q858 + q1640)
Hardisky et al. (1983), Kimes et al. (1981) Hardisky et al. (1983), Van Niel et al. (2003) Chen et al. (2005) Gao (1996) José et al. (2007) Rouse et al. (1974) Peñuelas et al. (1993), (1997) Peñuelas et al. (1993) Strachan et al. (2002) Zarco-Tejada et al. (2001), (2003) José et al. (2007) José et al. (2007) José et al. (2007) Rock et al. (1986), Hunt (1991) Rock et al. (1986), Hunt (1991) Rock et al. (1986), Hunt (1991) Fensholt and Sandholt (2003)
EWT
FMC **
0.341 0.331** 0.020 0.539** 0.600** 0.472** 0.647** 0.647** 0.659** 0.527** 0.258** 0.498** 0.618** 0.262** 0.374** 0.336** 0.332**
0.231* 0.247* 0.370* 0.163 0.099 0.313** 0.082 0.079 0.058 0.173 0.299** 0.160 0.078 0.294** 0.296** 0.248* 0.233*
Indicates significant differences at 95% confidence level (r = 0.195). Indicates significant difference at 99% confidence level (r = 0.254).
chlorophyll and other pigments (Thenkabail et al., 2000; Blackburn, 1998; Hansen and Schjoerring, 2003). These investigations have mainly been performed on other crop variables but not water parameters or on vegetation very distinct from cotton. As a consequence of the different measurement conditions and the different crop variables, some degree of disagreement exists in the selection of wavebands. Several studies successfully exploited empirical relationships to estimate leaf water content (Peñuelas et al., 1993; Ceccato et al., 2001; Datt, 1999), first derivative reflectance spectra (Danson et al., 1992), continuum removed spectra analysis (Curran et al., 2001; Pu et al., 2003; Tian et al., 2001) and artificial neural networks (Dawson et al., 1998). To date, however, relationships between leaf water content (EWT and FMC) and the narrow band ratio type of vegetation index (RVI) and normalized difference type of vegetation index (NDVI) involving all possible twoband combinations of 2150 bands (from 350 nm to 2500 nm) have not been well investigated. Based on the above background, the objective of the present investigation was to (i) exploit the relationships between leaf water content (EWT and FMC) and all possible two-band combinations indices, which based on raw reflectance spectra and first derivative reflectance spectra; (ii) identify the best combinations of narrow wavebands for ratio type of vegetation indices and normalized difference type of indices for EWT and FMC estimation and (iii) evaluate the performance of various types of hyperspectral vegetation indices in characterizing leaf water content. The final goal is to determine and recommend an optimal number of hyperspectral bands, their centers and widths, thus reducing the redundancy in hyperspectral data. 2. Materials and methods 2.1. Study area The field experiment was conducted in June–October 2010 at agricultural belts in Shihezi, Xinjiang, Northwest of China (85°590 E, 44°190 N), where cotton is the dominate crop. The continental arid climate of Xinjiang is characterized by aridity, rich sunlight and rare rainfall, with sharply defined seasons, high annual and diurnal fluctuations in air temperature, and low precipitation. The total annual precipitation for the whole study area is about 193 mm, and the total precipitation of whole cotton growth stage from April to October is about 108 mm. Sites of cotton were selected for the experiment. Cotton is generally planted in April–
May, and harvested in September–October. The whole growth period is about 180 days. The medium loam soil at the experiment area had the following properties: the field moisture capacity at depth of 10 cm is 0.33 g cm3, the volumetric water content at depth of 10 cm is 1.59 g cm3, and the saturation moisture content is 0.44 g cm3. Besides, field sampling was complemented by a water-controlled experiment in order to obtain very low vegetation water content that could not be obtained in the field, except in very extreme situations.
2.2. Plant sampling and water content measurements Leaf spectral readings and corresponding water status measurements were performed four times from seedling stage until boll stage (dates are 9–12 June, 14–18 July, 4–8 August, and 8–12 September, 2011, respectively). This procedure ensured that the normally occurring variation due to growth stage and measurement factors was included in the models giving a more realistic basis for model development. Three average-looking plants per plot were pulled out with their roots, placed and sealed in a plastic bag, and then placed in a cool dark container to avoid water loss as much as possible. Upon return to the laboratory, leaf sampling was conducted near the tops, middle and bottom of every sampling plant, for a total of 113 leaf samples. Fresh weight (FW) of leaves was recorded immediately using an analytical balance, after which optical properties were measured and leaf photos were taken. Fresh leaves were then put into oven to dry with 105 °C for half an hour and 70 °C till the constant weight were acquired. In order to make all measurement simultaneous, four groups worked like a line operation for leaf sampling, weighting, leaf spectra measurement and leaf photo taken. Leaf FWC and EWT were calculated for each leaf sample using Eqs. (1) and (2), respectively.
FMC ¼
FW DW 100% DW
ð1Þ
EWT ¼
FW DW dw A
ð2Þ
where FW is the leaf fresh weight and DW is the dry leaf weight of the same sample, A is the area of fresh leaf (cm2), dw is the density of water (1 g cm3).
146
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151
2.3. Leaf hyperspectral measurements Leaf reflectance was measured over the spectral region between 350 nm and 2500 nm by coupling a leaf clip (ASD, Inc., Boulder, Co., USA) with the ASD FieldSpec FR. The reflectance was measured in the ‘‘reflectance’’ mode against a black background. Fully expanded leaves near the tops, middle and bottom of sampling plants were respectively excised for leaf reflectance measurements, for a total of 113 leaf samples. The reflectance of a white Spectralon (BaSO4) panel was measured before every reflectance was taken, and reflectance was then calculated as the ratio between energy reflected by the leaf and energy incident on the leaf. Every reflectance was an average of ten repeated scans that were automatically acquired by the FieldSpec. 2.4. Specific waveband selection A rigorous and exhaustive approach is adopted in computing and evaluating hyperspectral indices. Because the ratio type of vegetation index (RVI, ration vegetation index) and the normalized difference type of vegetation index (NDVI, normalized difference vegetation index) are the classical vegetation index type, specific waveband selections for EWT and FMC were based on these two types of vegetation indices. The two-band ratio vegetation index (Eq. (3)) and two-band normalized difference vegetation index (Eq. (4)) have the following forms, respectively:
RVI ¼ k1 =k2
ð3Þ
NDVI ¼ ðk1 k2 Þ=ðk1 þ k2 Þ
ð4Þ
Two different hyperspectral reflectances were used for the different band combinations. In the first strategy, all raw spectral reflectances for wavelength from 350 to 2500 nm were directly used for the computation of RVI and NDVI band combinations. In the second strategy, the first order derivative spectral reflectances (DR, first derivative reflectance) based on raw reflectance were used for computation of RVI and NDVI band combinations. The first order derivative spetra was calculated by the following equation:
qðkiþ1 Þ ¼
qðkiþ1 Þ qðki1 Þ 2Dk
ð5Þ
All possible two-pair combinations of 2150 wavelengths (4, 622, 500 combinations) were used in Eqs. (3) and (4) and correlation analysis were performed in order to determine the correlation coefficient (r). All the r values were plotted in a matrix plot and the plot revealed a characteristic pattern with a number of ‘‘hot spot’’ with relatively high correlation coefficients. The optimal water indices were selected by choosing the wavelength combinations that showed the highest correlation coefficients. 3. Results and discussion 3.1. Single narrow band reflectance relationships with water content As a first step, raw reflectance and the first derivative reflectance from 350 to 2500 nm was correlated with EWT and FMC (Fig. 1). Maximum negative correlation coefficients (r) were mostly centered around 1550–1850 nm. This is a region of high water absorption. The maximum negative r values for relationship between raw reflectance and EWT (Fig. 1a) was at 1725 nm (r = 0.668), for FMC it was centered at 722 nm (r = 0.470). The maximum negative r values for relationship between the first derivative reflectance and EWT (Fig. 1b) was at 1482 nm (r = 0.693), for FMC it was at 1675 nm (r = 0.690). Recent investigations have revealed that reflectance is related to changes in leaf EWT rather than to changes in FWC (Ceccato et al., 2001; Datt,
1999; Davidson et al., 2006; Maki et al., 2004). FWC in fact, depends on two independent leaf variables; EWT and leaf dry mass area (LMA), both affecting leaf optical properties (Ceccato et al., 2001; Danson and Bowyer, 2004). 3.2. Published water indices relationships with water content In order to illustrate the efficiency of the new band combination indices, the published and most widely used ratio type of water indices and normalized differences type of water indices were adopted (Table 1). The performance of every water indices was evaluated by computing their correlations with water parameters (Table 1). As seen in Table 1, the correlationships between water indices and EWT were generally better than FMC. For EWT, the significant difference of almost all seventeen water indices reached 99% confidence level, but for FMC, the significant difference for almost all water indices just reached 95% confidence level. Compared with FMC, the better relationship between water indices and EWT was proved again. Besides, comparing the performance of published water indices with single narrow band reflectance, the maximum correlation coefficients values for EWT and FMC were both reduced. Among these published water indices, q900/min(q930–980 nm) and (q858 q2130)/(q858 + q2130) were best correlated with EWT and FMC, with r was 0.659 and 0.370, respectively. 3.3. RVI and NDVI band combinations relationship with water content Correlation coefficients r between all possible two-band ratio type of vegetation indices and water content, as well as two-band normalized type of vegetation indices and water content, were determined. The results of this comprehensive analysis were illustrated in matrix plots of r values, in Fig. 2. Matrix plots shown the correlation coefficients between water content and all RVI and NDVI band combinations for 2150 narrow bands spread across wavelength1 (350–2500 nm) and wavelength2 (350–2500 nm). A number of ‘‘hot spot’’ (Fig. 2) with high correlation coefficients were revealed. As seen in Fig. 1, the coefficients distributions of the matrix plots of normalized type vegetation index were completely symmetrical, but it was not true for ratio type vegetation index. Furthermore, the numbers of ‘‘hot spots’’ in plots for raw spectra were much more than in plots for the first derivative spectra. Analysis of the band centers and bandwidth in both directions were performed basing on these ‘‘hot spots’’. Band centers (k1 and k2) and band widths (Dk1 and Dk2) that combine to form the best indices were identified for EWT and FMC. This can be determined through matrix plots. For example, R-NDVI–EWT (NDVI band combinations based on raw reflectance relationship with EWT) plot, the index that has highest correlation coefficient (r = 0.693) has band centers (k1 and k2) and band widths (Dk1 and Dk2) extract from the plot range of 0.65–0.70, same to RRVI–EWT. Similar procedure was adopted for determining k1, k2, Dk1 and Dk2 for indices of other band combinations. It was noted that band centers (k1 and k2) and band widths (Dk1 and Dk2) for DR-NDVI–EWT, and DR-RVI–EWT extract from the plots range of 0.7–0.75, for R-NDVI–FMC, R-RVI–FMC, DR-NDVI–FMC, and DRRVI–FMC extract from the plots range of 0.75–1. Band centers and bandwidths for all optimal indices were tabulated in Table 2. As seen in Table 2, bands with center wavelengths in near infrared region from 950 nm to 1100 nm, and 1650 nm to 1700 nm were represented almost all selected bands for both EWT and FMC. The correlation coefficients for band combinations based on the first derivative reflectance were generally higher than these band combinations based on raw reflectance, for both EWT and FMC, the greatest correlation coefficients were acquired from the
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151
147
Fig. 1. Correlation coefficients (r) between water parameters and spectra reflectance. (a) Raw spectral reflectance and (b) the first derivative reflectance. r 0.05 and r 0.01 indicate significant differences at 95% and 99% confidence levels, respectively.
ratio type of band combinations based on the first derivative reflectance. Among all type of band combinations, DR-RVI index DR1647/DR1133 has the greatest coefficient (r = 0.748) with EWT, and DR1653/DR1687 has the greatest negative coefficient (r = 0.821) with FMC. Through above analysis, an interesting phenomenon was worth mentioning. Just as mentioned above, either for single narrow band reflectance or for the published water indices, their correlationships with EWT were generally better than with FMC, but for band combination indices, the coefficients for FMC were higher than EWT. Compared the correlationships between FMC and three types of reflectance, i.e. single narrow band reflectance, published water indices and new band combination indices, the coefficients between FMC and new band combination indices were greatly improve by increasing the maximum r from 0.690 for single band reflectance and 0.370 for published indices to 0.821. However, the improvement of correlation between EWT and the new band combination indices was not as obvious as FMC.
sions based on the best performance indices of all three types of independent variables were presented here. 3.5. Model validation Validation of the models was performed by comparing differences in R2 and the relative RMSE (REP). REP correlated with another widely used test criterion, i.e. root mean square error (RMSE). Because the differences among the values of RMSE were too small to effectively reflect the differences of the models’ performance, so REP was adopted. RMSE and REP values were calculated according to Eqs. (6) and (7), respectively:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn 2 ~ i¼1 ðyi yi Þ RMSE ¼ n
~i and yi were the predicted and measured water parameters, where y respectively, and n is the number of samples (for cross validation n = 75, for independent validation n = 38).
3.4. Model development In order to have the water parameter values equally distributed for model calibration and validation, the grouping of data for calibration and validation was accomplished. It was noted that here 75 samples out of 113 samples were used as calibration data set for model calibration and the remaining 38 samples were used as validation data set to examine the validity of the models. The statistics of EWT and FMC values for model calibration and validation were described in Table 3. Three different input strategies were used for the development of EWT and FMC estimation models. In the first strategy, single narrow bands of raw spectra and first derivative spectra that strongly correlated with EWT and FMC were used as independent variables. Basing on the above correlation analysis, raw reflectance at 1725 nm and 722 nm, and the first derivative reflectance at 1482 nm and 1675 nm were respectively selected as input variables for the EWT and FMC estimation model development. In the second strategy, among seventeen published water indices, those that significantly correlated with EWT and FMC were selected as independent variables. Then in the third strategy, the optimal indices based on band combinations analysis were used as independent variables. The model expressions for all water indices were presented in Table 4. For clarity, only the model expres-
ð6Þ
100 REP ¼ y
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn 2 ~ 100 i¼1 ðyi yi Þ RMSE ¼ y n
ð7Þ
was the mean value of the measured values of EWT and FMC. y In order to well demonstrate the performance of all indices, the REP and R2 were calculated with both calibration and validation data sets for all developed models, and the results were calculated, but only the evaluation results of the best performance models were presented here (Table 5). As can been seen, compared to the models based on the single narrow band reflectance, i.e. DR1482 and DR1675, or the published water indices, i.e. (q858 q2130)/(q858 + q2130) and q900/min(q930–980), the models based on the new band combination indices, i.e. DR1653/ DR1687and DR1647/DR1133, resulted in generally lower REP values and higher R2 values, which indicated good performance of the new band combination indices. Finally, in order to make the comparison results more convictive and more visual, the measured EWT and FMC values against the estimated values obtained by the best indices were plotted in Fig. 3. Ideally, when the regression line entirely overlaps the 1:1 line, it should be a perfect match. The slope and correlation coefficients between measured and estimated EWT and FMC values were also presented in the figure. It was obviously from the figure
148
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151
Fig. 2. Correlation coefficients for the relation of all band combinations (bands spread across 350–2500 nm) of RVI and NDVI based on raw reflectance (R) and the first derivative reflectance (DR) against EWT and FMC, respectively. (a) R-NDVI–EWT (NDVI band combinations based on raw reflectance relationship with EWT); (b) R-NDVI– FMC; (c) R-RVI–EWT; (d) R-RVI–FMC; (e) DR-NDVI–EWT; (f) DR-NDVI–FMC; (g) DR-RVI–EWT and (h) DR-RVI–FMC. Similar naming method for the rest plots.
149
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151 Table 2 Optimal band combinations for EWT and FMC. Band centers (k1 and k2) and bandwidths (Dk1and Dk2) in nm for all optimal band combinations. Type of index
EWT
FMC
R-NDVI
k1 Dk 1 954 882–959 Index (R954 R985)/(R954 + R985)
R-RVI
k1 954 Index R954/R985
k2 985
Dk 2 977–994 r 0.693
k1 Dk1 1693 1690–1694 Index (R1693 R1645)/(R1693 + R1645)
k2 985
Dk 2 977–994 r 0.693
k1 1693 Index R1693/R1645
DR-NDVI
k1 Dk 1 k2 1647 1646–1651 954 Index (DR1647 DR954)/(DR1647 + DR954)
DR-RVI
k1 Dk 1 1647 1645–1650 Index DR1647/DR1133
Dk 1 882–959
k2 1133
k2 1645
Dk 2 1642–1647 r 0.775
k2 1645
Dk 2 1642–1647 r 0.776
Dk 2 953–955 r 0.743
k1 Dk1 k2 1693 1691–1696 1660 Index (DR1693 DR1660)/(DR1693 + DR1660)
Dk 2 1660–1663 r 0.804
Dk 2 1127–1155 r 0.748
k1 Dk1 1653 1646–1661 Index DR1653/DR1687
Dk 2 1678–1696 r 0.821
EWT (cm) Total Calibration Validation FMC (%) Total Calibration Validation
Number
Std. deviation
Min.
Max.
Mean
113 75 38
0.0006 0.0008 0.0009
0.0131 0.0145 0.0131
0.0432 0.0432 0.0421
0.0287 0.0286 0.0288
113 75 38
9.595 13.22 8.108
240.31 240.31 261.76
762.16 762.16 518.70
369.13 390.71 326.53
Table 4 Expressions of estimation models for EWT and FMC. Independent variables
EWT models
DR1482 DR1675 (q858 q2130)/ (q858 + q2130) q900/min(q930–980) R954/R985 DR1647/DR1133 R1693/R1645 DR1653/DR1687
y = 15.274 +0.0483
k2 1687
to FMC model, the difference between the evaluation results for both calibration and validation set was small.
Table 3 Descriptive statistics of water content for model calibration and validation. Group
Dk1 1690–1695
FMC models y = 2E + 06 +585.68 y = 808.24 51.237
y = 0.4996 0.4915 y = 1.4195 1.4113 y = 0.1117 +0.0531 y = 17521 17065 y = 380.93 +141.37
that the FMC estimation model based on the new band combination index DR1647/DR1133 was the most efficient when evaluated by the calibration data set, but when evaluated by validation data set, the result was not so satisfied. Additionally, it was apparent that although the EWT model was a little less predictive compared
4. Conclusions In this study, in order to identified the optimal index for the estimation of cotton water content, i.e. EWT and FMC, the narrow band ratio type of vegetation index (RVI) and normalized difference type of vegetation index (NDVI) involving all possible twoband combinations of 2150 bands (from 350 nm to 2500 nm) were performed on cotton leaf reflectance, and the band combinations performed on both raw spectral reflectance and the first derivative reflectance. Then the predictive ability of the best combination of narrow wavebands for RVI and NDVI type of vegetation indices were compared with the single narrow band reflectance and some widely used published water indices. Correlation coefficients between all possible two-band combinations and water content were determined and the results of this comprehensive analysis were presented by matrix plots. Band centers (k1 and k2) and band widths (Dk1 and Dk2) that combine to form the best indices were identified for EWT and FMC through matrix plots. Bands with center wavelengths in near infrared region from 950 nm to 1100 nm, and 1650 nm to 1700 nm were represented almost all selected bands for both EWT and FMC. Furthermore, the performance of the best band combination indices based on the first derivative reflectance were always better than these based on the raw spectra. Among all type of band combinations, DR1647/DR1133 and DR1653/DR1687 were considered as the optimal indices for EWT and FMC estimation, respectively. Additionally, among the two water content, i.e. EWT and FMC, the correlation between EWT and spectral reflectance were
Table 5 Results of model performance analysis by calibration and validation data sets. Water content Water indices DR1482 DR1675 (q858 q2130)/(q858 + q2130) q900/min(q930–980) R954/R985 DR1647/DR1133 R1693/R1645 DR1653/DR1687
EWT REPcv 15.20
16.91 15.58 14.79
FMC 2
R cv 0.562
0.458 0.540 0.585
REP 17.72
16.49 17.16 14.57
2
R 0.343
REPcv
R2cv
REP
R2
19.78 26.12
0.551 0.205
22.53 28.78
0.191 0.116
17.53 16.74
0.642 0.674
17.74 12.91
0.363 0.472
0.424 0.402 0.523
REPcv and R2cv represent the cross-validation results based on the calibration data set; REPcv and R2cv represent the validation results based on the validation data set.
150
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151
Fig. 3. Scatter plots of measured versus estimated EWT and FMC for linear regression based on the optimal band combination indices. (a) Calibration data set and (b) validation data set. The dash line represents the 1:1 line, and the thick solid line represents the regression line.
generally better than that of FMC. However, the correlation between FMC and reflectance can be significantly improved by the best band combination indices, to some extent, which indicated that the selected published water indices used in the study were not suitable for FMC estimation. Finally, comparing the predictive power of the three types of hyperspectral predictors, all selected published water indices were found to be the worst performance, even worse than the single band reflectance. However, as expected, although the low REP and high R2cv were obtained from the best single narrow band, the REP and R2 were unsatisfied, which reflected the poor predictive ability of the single narrow band. The bad performance of the single narrow band reflectance and the published water indices once again demonstrate the necessity of the optimal vegetation index selection for different crop and different crop variables. However, this was the first attempt to estimate the cotton water content through two-band combination index, and the study should further our understanding of the relationships between cotton water content and vegetation indices. Although encouraging performance of the two new band indices, i.e. DR1647/DR1133 and DR1653/DR1687, for the estimation of EWT and FMC, was proven, more research is needed to test the band centers and band widths proposed in the study. Acknowledgements The work supported by West Light Foundation of the Chinese Academy of Sciences (XBBS200902), Major State Basic Research
Development Program (2009CB421301), National Natural Science Foundation of China (41101430), Knowledge Innovation Project (KZCX2-YW-BR-12), and China Postdoctoral Science Foundation (20100471681). The work assistances given by colleagues in Xinjiang Institute of Ecology and Geography Chinese Academy of Sciences are highly appreciated. References Blackburn, G.A., 1998. Spectral indices for estimating photosynthetic pigment concentrations: a test using senescent tree leaves. Int. J. Remote Sens. 19, 657– 675. Burgan, R.E., 1996. Use of remotely sensed data for fire danger estimation. EARSeL Adv. Remote Sens. 4 (4), 1–8. Carlson, J.D., Burgan, R.E., 2003. Review of user needs in operational fire danger estimation: the Oklahoma example. Int. J. Remote Sens. 24, 1601–1620. Carter, G.A., 1993. Responses of leaf spectral reflectance to plant stress. Am. J. Bot. 80, 239–243. Ceccato, P., Flasse, S., Tarantola, S., Jacquemoud, S., Gregoire, J.M., 2001. Detecting vegetation leaf water content using reflectance in the optical domain. Remote Sens. Environ. 77, 22–33. Chen, D.Y., Huang, J.F., Thomas, J.J., 2005. Vegetation water content estimation for corn and soybeans using spectral indices derived from MODIS near- and shortwave infrared bands. Remote Sens. Environ. 98, 225–236. Chuvieco, E., Cocero, D., Aguado, I., Palacios-Orueta, A., Prado, E., 2004. Improving burning efficiency estimates through satellite assessment of fuel moisture content. J. Geophys. Res. – Atmos. 109, 1–8. Curran, P.J., Dungan, J.L., Peterson, D.L., 2001. Estimating the foliar biochemical concentration of leaves with reflectance spectrometry: testing the Kokaly and Clark methodologies. Remote Sens. Environ. 76, 349–359. Danson, F.M., Bowyer, P., 2004. Estimating live fuel moisture content from remotely sensed reflectance. Remote Sens. Environ. 92, 309–321. Danson, F.M., Steven, M.D., Malthus, T.J., Clark, J.A., 1992. Highspectral resolution data for determining leaf water content. Int. J. Remote Sens. 13 (3), 461–470.
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151 Datt, B., 1999. Remote sensing of water content in eucalyptus leaves. Aust. J. Bot. 47, 909–923. Davidson, A., Wang, S., Wilmshurst, J., 2006. Remote sensing of grassland– shrubland vegetation water content in the shortwave domain. Int. J. Appl. Earth Obs. Geoinf. 8, 225–236. Dawson, T.P., Curran, P.J., Plummer, S.E., 1998. The biochemical decomposition of slash pine needles from reflectance spectra using neural networks. Int. J. Remote Sens. 19, 1433–1438. Fensholt, R., Sandholt, I., 2003. Derivation of a shortwave infrared water stress index from MODIS near- and shortwave infrared data in a semiarid environment. Remote Sens. Environ. 87, 111–121. Gao, B.C., 1996. NDWI-a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 58, 257–266. Hansen, P.M., Schjoerring, J.K., 2003. Reflectance measurement of canopy biomass and nitrogen status in wheat crops using normalized difference vegetation indices and partial least squares regression. Remote Sens. Environ. 86, 542–553. Hardisky, M.A., Klemas, V., Smart, R.M., 1983. The influences of soil salinity, growth form, and leaf moisture on the spectral reflectance of Spartina alterniflora canopies. Photogramm. Eng. Remote Sens. 49, 77–83. Hunt Jr., E.R., 1991. Airborne remote sensing of canopy water thickness scaled from leaf spectrometer data. Int. J. Remote Sens. 12, 643–649. José, R.R., David R., Eli C., Susan, U., David, R.S., 2007. Evaluation of hyperspectral reflectance indexes to detect grapevine water status in vineyards. Am. J. Enol. Vitic. 58 (3), 302–317. Kimes, D.S., Markham, B.L., Tucker, C.J., 1981. Temporal relationships between spectral response and agronomic variables of a corn canopy. Remote Sens. Environ. 11, 401–411. Knipling, E.B., 1970. Physical and physiological basis for the reflectance of visible and near-infrared radiation from vegetation. Remote Sens. Environ. 1, 155–159. Maki, M., Ishiahra, M., Tamura, M., 2004. Estimation of leaf water status to monitor the risk of forest fires by using remotely sensed data. Remote Sens. Environ. 90, 441–450. Peñuelas, J., Filella, I., Biel, C., Serrano, L., Save, R., 1993. The reflectance at the 950– 970 nm region as an indicator of plant water status. Int. J. Remote Sens. 14, 1887–1905. Peñuelas, J., Gamon, J.A., Fredeen, A.L., Merino, J., Field, C.B., 1994. Reflectance indices associated with physiological changes in nitrogen and water limited sunflower leaves. Remote Sens. Environ. 48, 135–146.
151
Peñuelas, J., Pinol, J., Ogaya, R., Filella, I., 1997. Estimation of plant water concentration by the reflectance water index WI (R900/R970). Int. J. Remote Sens. 14, 1887–1905. Pu, R., Ge, S., Kelly, N.M., Gong, P., 2003. Spectral absorption features as indicators of water status in coast live oak (Quercus agrifolia) leaves. Int. J. Remote Sens. 24, 1799–1810. Rock, B.N., Vogelmann, J.E., Williams, D.L., Vogelmann, A.F., Hoshizaki, T., 1986. Remote detection of forest damage. J. Biosci. 36, 439–445. Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W. 1974. Monitoring vegetation systems in the Great Plains with ERTS. In: Proceedings, Third Earth Resources Technology Satellite-1 Symposium. Greenbelt: NASA SP-351, pp. 301–317. Stimson, H.C., Breshears, D.D., Ustin, S.L., Kefauver, S.C., 2005. Spectral sensing of foliar water conditions in two co-occurring conifer species: pinus edulis and Juniperus monosperma. Remote Sens. Environ. 96, 108–118. Strachan, I.B., Pattey, E., Boisvert, J.B., 2002. Impact of nitrogen and environmental conditions on corn as detected by hyperspectral reflectance. Remote Sens. Environ. 80, 213–224. Thenkabail, P.S., Smith, R.B., Pauw, E.D., 2000. Hyperspectral vegetation indices and their relationships with agricultural crop characteristics. Remote Sens. Environ. 71, 158–182. Thomas, J.R., Namken, L.N., Oerther, G.F., Brown, R.G., 1971. Estimating leaf water content by reflectance measurements. Agron. J. 63, 845–847. Tian, Q., Tong, Q., Pu, R., Guo, X., Zhao, C., 2001. Spectroscopic determinations of wheat water status using 1650–1850 nm spectral absorption features. Int. J. Remote Sens. 22, 2329–2338. Tucker, C.J., 1980. Remote sensing of leaf water content in the near infrared. Remote Sens. Environ. 10, 23–32. Ustin, S.L., Roberts, D.A., Pinzón, J., Jacquemoud, S., Gardner, M., Scheer, G.C., 1998. Estimating canopy water content of chaparral shrubs using optical methods. Remote Sens. Environ. 65, 280–291. Van Niel, T.G., McVicar, T.R., Fang, H., Liang, S., 2003. Calculating environmental moisture for per-field discrimination of rice crops. Int. J. Remote Sens. 24, 885–890. Zarco-Tejada, P.J., Miller, J.R., Noland, T.L., Mohammed, G.H., Sampson, P.H., 2001. Scaling-up and model inversion methods with narrow-band optical indices for chlorophyll content estimation in closed forest canopies with hyperspectral data. IEEE Trans. Geosci. Remote Sens. 39 (7), 1491–1507. Zarco-Tejada, P.J., Rueda, C.A., Ustin, S.L., 2003. Water content estimation in vegetation with MODIS reflectance data and model inversion methods. Remote Sens. Environ. 85, 109–124.
Contents lists available at SciVerse ScienceDirect
Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag
Estimation of leaf water content in cotton by means of hyperspectral indices Qiu-xiang Yi ⇑, An-ming Bao, Qiang Wang, Jin Zhao Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, 830011 Urumqi, China
a r t i c l e
i n f o
Article history: Received 8 May 2012 Received in revised form 18 September 2012 Accepted 28 September 2012
Keywords: Equivalent water content Fuel moisture content Ratio type of vegetation index Normalized difference type of vegetation index Two-band combination
a b s t r a c t The knowledge of vegetation water conditions can contribute to drought assessment. Remote sensing has a proven ability to assess vegetation properties. In this study, all two-band combinations (350–2500 nm) in the ratio type of vegetation index (RVI) and the normalized difference type of vegetation index (NDVI) were performed on cotton leaf raw spectral reflectance (R) and the first derivative reflectance (DR). The correlation coefficient (r) between all two-band combinations and two leaf water parameters (EWT: equivalent water thickness, and FMC: fuel moisture content) were determined, and the results of this comprehensive analysis were presented by matrix plots. Band centers (k1 and k2) and band widths (Dk1 and Dk2) that combine to form the best indices were identified for EWT and FMC through matrix plots. Then the evaluation of the predictive power of three predictors, i.e. single narrow band reflectance, the widely used published water indices and the best band combination indices, were performed. The results shown that the new indices DR1647/DR1133 and DR1653/DR1687, proposed by two-band combinations, were considered as the optimal indices for EWT and FMC estimation, respectively. The models based on these two best combination indices could explain 58% and 67% variability in EWT and FMC, respectively. Besides, bands with center wavelengths in region from 950 nm to 1100 nm, and 1650 nm to 1750 nm were represented almost all selected bands. The study should further our understanding of the relationships between leaf water content and hyperspectral reflectance. Ó 2012 Elsevier B.V. All rights reserved.
1. Introduction The knowledge of vegetation water conditions can in fact contribute to detect vegetation physiological status (Carter, 1993; Peñuelas et al., 1994; Stimson et al., 2005), to provide useful information in agriculture for irrigation decisions and drought assessment (Peñuelas et al., 1993, 1994) and it is important in forestry in determining fire susceptibility (Carlson and Burgan, 2003; Chuvieco et al., 2004; Ustin et al., 1998). Remote sensing offers an important opportunity for quantitative assessment of vegetation properties at different scales. Together with other parameters, vegetation water content is an important property that can be investigated by using remotely sensed data. With the development of hyperspectral remote sensing technique, direct detection approaches of vegetation water content have been proposed and further used to evaluate crop drought condition. The main parameters describing the amount of water in vegetation that are usually investigated by remote sensing are fuel moisture content (FMC, %) and leaf equivalent water thickness (EWT, cm). FMC defined as the proportion of water over the vegetation dry mass (Burgan, 1996), and EWT is the ratio between the quantity of water and leaf area (Danson et al., 1992). EWT and FMC pro⇑ Corresponding author. Tel.: +86 (0)9917823131. E-mail address: [email protected] (Q.-x. Yi). 0168-1699/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.compag.2012.09.011
vide information on the amount of water present in vegetation. The possibility of estimating FMC and EWT by means of remotely sensed data derives from the fact that water absorbs radiant energy throughout the near-infrared (750–1300 nm) and short-infrared (1300–2500 nm) spectral regions. Depending on tissue water content, reflectance is thus reduced to a varying extent within the water absorption features centered on 970, 1200, 1450, 1940, and 2500 nm (Knipling, 1970; Thomas et al., 1971; Tucker, 1980) and these changes can be recognized and quantified as water content variations. Our study investigates the potential of retrieving both of these measurements via leaf reflectance. Several hyperspectral indices such as the Normalized Difference Infrared Index (NDII), the Water Index (WI) have been established based on hyperspectral remote sensing data. The most widely used ratio type of vegetation index and normalized difference type of vegetation index at leaf level were summarized in Table 1. The results obtained in these studies indicate that knowledge of connection between the investigated variable and the spectral data can improve the performance of the vegetation indices (Thenkabail et al., 2000). Further improvement in indices is generally possible through use of spectral data from distinct narrow bands or through corrections for soil background effects, and further, through different band combinations. Selection of new wavebands in hyperspectral data has been performed in some cases, mainly focusing on how to increase the sensitivity of the vegetation indices to
145
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151 Table 1 Correlation analysis between published water indices and water content (n = 113). Water indices NDII NDII1 NDWI2130 NDWI1240 NDWI1260 NDVI WI WBI FWBI SRWI1 SRWI2 SRWI3 SRWI4 MSI MSI1 MSI1 SIWSI * **
Name Normalized different infrared index Normalized different infrared index Normalized different water index Normalized different water index Normalized different water index Normalized different vegetation index Water index Water band index Floating position water band Simple ratio water index Simple ratio water index Simple ratio water index Simple ratio water index Moisture stress index Moisture stress index Moisture stress index Shortwave infrared water stress
Formula
Reference
(q850 q1650)/(q850 + q1650) (q835 q1650)/(q835 + q1650) (q858 q2130)/(q858 + q2130) (q860 q1240)/(q860 + q1240) (q870 q1260)/(q870 + q1260) (q858 q648)/(q858 + q648) q900/q970 q970/q900 q900/min(q930–980) q858/q1240 q1070/q1340 q678/q1070 q880/q1265 q1600/q820 q870/q1350 q1650/q835 (q858 q1640)/(q858 + q1640)
Hardisky et al. (1983), Kimes et al. (1981) Hardisky et al. (1983), Van Niel et al. (2003) Chen et al. (2005) Gao (1996) José et al. (2007) Rouse et al. (1974) Peñuelas et al. (1993), (1997) Peñuelas et al. (1993) Strachan et al. (2002) Zarco-Tejada et al. (2001), (2003) José et al. (2007) José et al. (2007) José et al. (2007) Rock et al. (1986), Hunt (1991) Rock et al. (1986), Hunt (1991) Rock et al. (1986), Hunt (1991) Fensholt and Sandholt (2003)
EWT
FMC **
0.341 0.331** 0.020 0.539** 0.600** 0.472** 0.647** 0.647** 0.659** 0.527** 0.258** 0.498** 0.618** 0.262** 0.374** 0.336** 0.332**
0.231* 0.247* 0.370* 0.163 0.099 0.313** 0.082 0.079 0.058 0.173 0.299** 0.160 0.078 0.294** 0.296** 0.248* 0.233*
Indicates significant differences at 95% confidence level (r = 0.195). Indicates significant difference at 99% confidence level (r = 0.254).
chlorophyll and other pigments (Thenkabail et al., 2000; Blackburn, 1998; Hansen and Schjoerring, 2003). These investigations have mainly been performed on other crop variables but not water parameters or on vegetation very distinct from cotton. As a consequence of the different measurement conditions and the different crop variables, some degree of disagreement exists in the selection of wavebands. Several studies successfully exploited empirical relationships to estimate leaf water content (Peñuelas et al., 1993; Ceccato et al., 2001; Datt, 1999), first derivative reflectance spectra (Danson et al., 1992), continuum removed spectra analysis (Curran et al., 2001; Pu et al., 2003; Tian et al., 2001) and artificial neural networks (Dawson et al., 1998). To date, however, relationships between leaf water content (EWT and FMC) and the narrow band ratio type of vegetation index (RVI) and normalized difference type of vegetation index (NDVI) involving all possible twoband combinations of 2150 bands (from 350 nm to 2500 nm) have not been well investigated. Based on the above background, the objective of the present investigation was to (i) exploit the relationships between leaf water content (EWT and FMC) and all possible two-band combinations indices, which based on raw reflectance spectra and first derivative reflectance spectra; (ii) identify the best combinations of narrow wavebands for ratio type of vegetation indices and normalized difference type of indices for EWT and FMC estimation and (iii) evaluate the performance of various types of hyperspectral vegetation indices in characterizing leaf water content. The final goal is to determine and recommend an optimal number of hyperspectral bands, their centers and widths, thus reducing the redundancy in hyperspectral data. 2. Materials and methods 2.1. Study area The field experiment was conducted in June–October 2010 at agricultural belts in Shihezi, Xinjiang, Northwest of China (85°590 E, 44°190 N), where cotton is the dominate crop. The continental arid climate of Xinjiang is characterized by aridity, rich sunlight and rare rainfall, with sharply defined seasons, high annual and diurnal fluctuations in air temperature, and low precipitation. The total annual precipitation for the whole study area is about 193 mm, and the total precipitation of whole cotton growth stage from April to October is about 108 mm. Sites of cotton were selected for the experiment. Cotton is generally planted in April–
May, and harvested in September–October. The whole growth period is about 180 days. The medium loam soil at the experiment area had the following properties: the field moisture capacity at depth of 10 cm is 0.33 g cm3, the volumetric water content at depth of 10 cm is 1.59 g cm3, and the saturation moisture content is 0.44 g cm3. Besides, field sampling was complemented by a water-controlled experiment in order to obtain very low vegetation water content that could not be obtained in the field, except in very extreme situations.
2.2. Plant sampling and water content measurements Leaf spectral readings and corresponding water status measurements were performed four times from seedling stage until boll stage (dates are 9–12 June, 14–18 July, 4–8 August, and 8–12 September, 2011, respectively). This procedure ensured that the normally occurring variation due to growth stage and measurement factors was included in the models giving a more realistic basis for model development. Three average-looking plants per plot were pulled out with their roots, placed and sealed in a plastic bag, and then placed in a cool dark container to avoid water loss as much as possible. Upon return to the laboratory, leaf sampling was conducted near the tops, middle and bottom of every sampling plant, for a total of 113 leaf samples. Fresh weight (FW) of leaves was recorded immediately using an analytical balance, after which optical properties were measured and leaf photos were taken. Fresh leaves were then put into oven to dry with 105 °C for half an hour and 70 °C till the constant weight were acquired. In order to make all measurement simultaneous, four groups worked like a line operation for leaf sampling, weighting, leaf spectra measurement and leaf photo taken. Leaf FWC and EWT were calculated for each leaf sample using Eqs. (1) and (2), respectively.
FMC ¼
FW DW 100% DW
ð1Þ
EWT ¼
FW DW dw A
ð2Þ
where FW is the leaf fresh weight and DW is the dry leaf weight of the same sample, A is the area of fresh leaf (cm2), dw is the density of water (1 g cm3).
146
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151
2.3. Leaf hyperspectral measurements Leaf reflectance was measured over the spectral region between 350 nm and 2500 nm by coupling a leaf clip (ASD, Inc., Boulder, Co., USA) with the ASD FieldSpec FR. The reflectance was measured in the ‘‘reflectance’’ mode against a black background. Fully expanded leaves near the tops, middle and bottom of sampling plants were respectively excised for leaf reflectance measurements, for a total of 113 leaf samples. The reflectance of a white Spectralon (BaSO4) panel was measured before every reflectance was taken, and reflectance was then calculated as the ratio between energy reflected by the leaf and energy incident on the leaf. Every reflectance was an average of ten repeated scans that were automatically acquired by the FieldSpec. 2.4. Specific waveband selection A rigorous and exhaustive approach is adopted in computing and evaluating hyperspectral indices. Because the ratio type of vegetation index (RVI, ration vegetation index) and the normalized difference type of vegetation index (NDVI, normalized difference vegetation index) are the classical vegetation index type, specific waveband selections for EWT and FMC were based on these two types of vegetation indices. The two-band ratio vegetation index (Eq. (3)) and two-band normalized difference vegetation index (Eq. (4)) have the following forms, respectively:
RVI ¼ k1 =k2
ð3Þ
NDVI ¼ ðk1 k2 Þ=ðk1 þ k2 Þ
ð4Þ
Two different hyperspectral reflectances were used for the different band combinations. In the first strategy, all raw spectral reflectances for wavelength from 350 to 2500 nm were directly used for the computation of RVI and NDVI band combinations. In the second strategy, the first order derivative spectral reflectances (DR, first derivative reflectance) based on raw reflectance were used for computation of RVI and NDVI band combinations. The first order derivative spetra was calculated by the following equation:
qðkiþ1 Þ ¼
qðkiþ1 Þ qðki1 Þ 2Dk
ð5Þ
All possible two-pair combinations of 2150 wavelengths (4, 622, 500 combinations) were used in Eqs. (3) and (4) and correlation analysis were performed in order to determine the correlation coefficient (r). All the r values were plotted in a matrix plot and the plot revealed a characteristic pattern with a number of ‘‘hot spot’’ with relatively high correlation coefficients. The optimal water indices were selected by choosing the wavelength combinations that showed the highest correlation coefficients. 3. Results and discussion 3.1. Single narrow band reflectance relationships with water content As a first step, raw reflectance and the first derivative reflectance from 350 to 2500 nm was correlated with EWT and FMC (Fig. 1). Maximum negative correlation coefficients (r) were mostly centered around 1550–1850 nm. This is a region of high water absorption. The maximum negative r values for relationship between raw reflectance and EWT (Fig. 1a) was at 1725 nm (r = 0.668), for FMC it was centered at 722 nm (r = 0.470). The maximum negative r values for relationship between the first derivative reflectance and EWT (Fig. 1b) was at 1482 nm (r = 0.693), for FMC it was at 1675 nm (r = 0.690). Recent investigations have revealed that reflectance is related to changes in leaf EWT rather than to changes in FWC (Ceccato et al., 2001; Datt,
1999; Davidson et al., 2006; Maki et al., 2004). FWC in fact, depends on two independent leaf variables; EWT and leaf dry mass area (LMA), both affecting leaf optical properties (Ceccato et al., 2001; Danson and Bowyer, 2004). 3.2. Published water indices relationships with water content In order to illustrate the efficiency of the new band combination indices, the published and most widely used ratio type of water indices and normalized differences type of water indices were adopted (Table 1). The performance of every water indices was evaluated by computing their correlations with water parameters (Table 1). As seen in Table 1, the correlationships between water indices and EWT were generally better than FMC. For EWT, the significant difference of almost all seventeen water indices reached 99% confidence level, but for FMC, the significant difference for almost all water indices just reached 95% confidence level. Compared with FMC, the better relationship between water indices and EWT was proved again. Besides, comparing the performance of published water indices with single narrow band reflectance, the maximum correlation coefficients values for EWT and FMC were both reduced. Among these published water indices, q900/min(q930–980 nm) and (q858 q2130)/(q858 + q2130) were best correlated with EWT and FMC, with r was 0.659 and 0.370, respectively. 3.3. RVI and NDVI band combinations relationship with water content Correlation coefficients r between all possible two-band ratio type of vegetation indices and water content, as well as two-band normalized type of vegetation indices and water content, were determined. The results of this comprehensive analysis were illustrated in matrix plots of r values, in Fig. 2. Matrix plots shown the correlation coefficients between water content and all RVI and NDVI band combinations for 2150 narrow bands spread across wavelength1 (350–2500 nm) and wavelength2 (350–2500 nm). A number of ‘‘hot spot’’ (Fig. 2) with high correlation coefficients were revealed. As seen in Fig. 1, the coefficients distributions of the matrix plots of normalized type vegetation index were completely symmetrical, but it was not true for ratio type vegetation index. Furthermore, the numbers of ‘‘hot spots’’ in plots for raw spectra were much more than in plots for the first derivative spectra. Analysis of the band centers and bandwidth in both directions were performed basing on these ‘‘hot spots’’. Band centers (k1 and k2) and band widths (Dk1 and Dk2) that combine to form the best indices were identified for EWT and FMC. This can be determined through matrix plots. For example, R-NDVI–EWT (NDVI band combinations based on raw reflectance relationship with EWT) plot, the index that has highest correlation coefficient (r = 0.693) has band centers (k1 and k2) and band widths (Dk1 and Dk2) extract from the plot range of 0.65–0.70, same to RRVI–EWT. Similar procedure was adopted for determining k1, k2, Dk1 and Dk2 for indices of other band combinations. It was noted that band centers (k1 and k2) and band widths (Dk1 and Dk2) for DR-NDVI–EWT, and DR-RVI–EWT extract from the plots range of 0.7–0.75, for R-NDVI–FMC, R-RVI–FMC, DR-NDVI–FMC, and DRRVI–FMC extract from the plots range of 0.75–1. Band centers and bandwidths for all optimal indices were tabulated in Table 2. As seen in Table 2, bands with center wavelengths in near infrared region from 950 nm to 1100 nm, and 1650 nm to 1700 nm were represented almost all selected bands for both EWT and FMC. The correlation coefficients for band combinations based on the first derivative reflectance were generally higher than these band combinations based on raw reflectance, for both EWT and FMC, the greatest correlation coefficients were acquired from the
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151
147
Fig. 1. Correlation coefficients (r) between water parameters and spectra reflectance. (a) Raw spectral reflectance and (b) the first derivative reflectance. r 0.05 and r 0.01 indicate significant differences at 95% and 99% confidence levels, respectively.
ratio type of band combinations based on the first derivative reflectance. Among all type of band combinations, DR-RVI index DR1647/DR1133 has the greatest coefficient (r = 0.748) with EWT, and DR1653/DR1687 has the greatest negative coefficient (r = 0.821) with FMC. Through above analysis, an interesting phenomenon was worth mentioning. Just as mentioned above, either for single narrow band reflectance or for the published water indices, their correlationships with EWT were generally better than with FMC, but for band combination indices, the coefficients for FMC were higher than EWT. Compared the correlationships between FMC and three types of reflectance, i.e. single narrow band reflectance, published water indices and new band combination indices, the coefficients between FMC and new band combination indices were greatly improve by increasing the maximum r from 0.690 for single band reflectance and 0.370 for published indices to 0.821. However, the improvement of correlation between EWT and the new band combination indices was not as obvious as FMC.
sions based on the best performance indices of all three types of independent variables were presented here. 3.5. Model validation Validation of the models was performed by comparing differences in R2 and the relative RMSE (REP). REP correlated with another widely used test criterion, i.e. root mean square error (RMSE). Because the differences among the values of RMSE were too small to effectively reflect the differences of the models’ performance, so REP was adopted. RMSE and REP values were calculated according to Eqs. (6) and (7), respectively:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn 2 ~ i¼1 ðyi yi Þ RMSE ¼ n
~i and yi were the predicted and measured water parameters, where y respectively, and n is the number of samples (for cross validation n = 75, for independent validation n = 38).
3.4. Model development In order to have the water parameter values equally distributed for model calibration and validation, the grouping of data for calibration and validation was accomplished. It was noted that here 75 samples out of 113 samples were used as calibration data set for model calibration and the remaining 38 samples were used as validation data set to examine the validity of the models. The statistics of EWT and FMC values for model calibration and validation were described in Table 3. Three different input strategies were used for the development of EWT and FMC estimation models. In the first strategy, single narrow bands of raw spectra and first derivative spectra that strongly correlated with EWT and FMC were used as independent variables. Basing on the above correlation analysis, raw reflectance at 1725 nm and 722 nm, and the first derivative reflectance at 1482 nm and 1675 nm were respectively selected as input variables for the EWT and FMC estimation model development. In the second strategy, among seventeen published water indices, those that significantly correlated with EWT and FMC were selected as independent variables. Then in the third strategy, the optimal indices based on band combinations analysis were used as independent variables. The model expressions for all water indices were presented in Table 4. For clarity, only the model expres-
ð6Þ
100 REP ¼ y
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn 2 ~ 100 i¼1 ðyi yi Þ RMSE ¼ y n
ð7Þ
was the mean value of the measured values of EWT and FMC. y In order to well demonstrate the performance of all indices, the REP and R2 were calculated with both calibration and validation data sets for all developed models, and the results were calculated, but only the evaluation results of the best performance models were presented here (Table 5). As can been seen, compared to the models based on the single narrow band reflectance, i.e. DR1482 and DR1675, or the published water indices, i.e. (q858 q2130)/(q858 + q2130) and q900/min(q930–980), the models based on the new band combination indices, i.e. DR1653/ DR1687and DR1647/DR1133, resulted in generally lower REP values and higher R2 values, which indicated good performance of the new band combination indices. Finally, in order to make the comparison results more convictive and more visual, the measured EWT and FMC values against the estimated values obtained by the best indices were plotted in Fig. 3. Ideally, when the regression line entirely overlaps the 1:1 line, it should be a perfect match. The slope and correlation coefficients between measured and estimated EWT and FMC values were also presented in the figure. It was obviously from the figure
148
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151
Fig. 2. Correlation coefficients for the relation of all band combinations (bands spread across 350–2500 nm) of RVI and NDVI based on raw reflectance (R) and the first derivative reflectance (DR) against EWT and FMC, respectively. (a) R-NDVI–EWT (NDVI band combinations based on raw reflectance relationship with EWT); (b) R-NDVI– FMC; (c) R-RVI–EWT; (d) R-RVI–FMC; (e) DR-NDVI–EWT; (f) DR-NDVI–FMC; (g) DR-RVI–EWT and (h) DR-RVI–FMC. Similar naming method for the rest plots.
149
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151 Table 2 Optimal band combinations for EWT and FMC. Band centers (k1 and k2) and bandwidths (Dk1and Dk2) in nm for all optimal band combinations. Type of index
EWT
FMC
R-NDVI
k1 Dk 1 954 882–959 Index (R954 R985)/(R954 + R985)
R-RVI
k1 954 Index R954/R985
k2 985
Dk 2 977–994 r 0.693
k1 Dk1 1693 1690–1694 Index (R1693 R1645)/(R1693 + R1645)
k2 985
Dk 2 977–994 r 0.693
k1 1693 Index R1693/R1645
DR-NDVI
k1 Dk 1 k2 1647 1646–1651 954 Index (DR1647 DR954)/(DR1647 + DR954)
DR-RVI
k1 Dk 1 1647 1645–1650 Index DR1647/DR1133
Dk 1 882–959
k2 1133
k2 1645
Dk 2 1642–1647 r 0.775
k2 1645
Dk 2 1642–1647 r 0.776
Dk 2 953–955 r 0.743
k1 Dk1 k2 1693 1691–1696 1660 Index (DR1693 DR1660)/(DR1693 + DR1660)
Dk 2 1660–1663 r 0.804
Dk 2 1127–1155 r 0.748
k1 Dk1 1653 1646–1661 Index DR1653/DR1687
Dk 2 1678–1696 r 0.821
EWT (cm) Total Calibration Validation FMC (%) Total Calibration Validation
Number
Std. deviation
Min.
Max.
Mean
113 75 38
0.0006 0.0008 0.0009
0.0131 0.0145 0.0131
0.0432 0.0432 0.0421
0.0287 0.0286 0.0288
113 75 38
9.595 13.22 8.108
240.31 240.31 261.76
762.16 762.16 518.70
369.13 390.71 326.53
Table 4 Expressions of estimation models for EWT and FMC. Independent variables
EWT models
DR1482 DR1675 (q858 q2130)/ (q858 + q2130) q900/min(q930–980) R954/R985 DR1647/DR1133 R1693/R1645 DR1653/DR1687
y = 15.274 +0.0483
k2 1687
to FMC model, the difference between the evaluation results for both calibration and validation set was small.
Table 3 Descriptive statistics of water content for model calibration and validation. Group
Dk1 1690–1695
FMC models y = 2E + 06 +585.68 y = 808.24 51.237
y = 0.4996 0.4915 y = 1.4195 1.4113 y = 0.1117 +0.0531 y = 17521 17065 y = 380.93 +141.37
that the FMC estimation model based on the new band combination index DR1647/DR1133 was the most efficient when evaluated by the calibration data set, but when evaluated by validation data set, the result was not so satisfied. Additionally, it was apparent that although the EWT model was a little less predictive compared
4. Conclusions In this study, in order to identified the optimal index for the estimation of cotton water content, i.e. EWT and FMC, the narrow band ratio type of vegetation index (RVI) and normalized difference type of vegetation index (NDVI) involving all possible twoband combinations of 2150 bands (from 350 nm to 2500 nm) were performed on cotton leaf reflectance, and the band combinations performed on both raw spectral reflectance and the first derivative reflectance. Then the predictive ability of the best combination of narrow wavebands for RVI and NDVI type of vegetation indices were compared with the single narrow band reflectance and some widely used published water indices. Correlation coefficients between all possible two-band combinations and water content were determined and the results of this comprehensive analysis were presented by matrix plots. Band centers (k1 and k2) and band widths (Dk1 and Dk2) that combine to form the best indices were identified for EWT and FMC through matrix plots. Bands with center wavelengths in near infrared region from 950 nm to 1100 nm, and 1650 nm to 1700 nm were represented almost all selected bands for both EWT and FMC. Furthermore, the performance of the best band combination indices based on the first derivative reflectance were always better than these based on the raw spectra. Among all type of band combinations, DR1647/DR1133 and DR1653/DR1687 were considered as the optimal indices for EWT and FMC estimation, respectively. Additionally, among the two water content, i.e. EWT and FMC, the correlation between EWT and spectral reflectance were
Table 5 Results of model performance analysis by calibration and validation data sets. Water content Water indices DR1482 DR1675 (q858 q2130)/(q858 + q2130) q900/min(q930–980) R954/R985 DR1647/DR1133 R1693/R1645 DR1653/DR1687
EWT REPcv 15.20
16.91 15.58 14.79
FMC 2
R cv 0.562
0.458 0.540 0.585
REP 17.72
16.49 17.16 14.57
2
R 0.343
REPcv
R2cv
REP
R2
19.78 26.12
0.551 0.205
22.53 28.78
0.191 0.116
17.53 16.74
0.642 0.674
17.74 12.91
0.363 0.472
0.424 0.402 0.523
REPcv and R2cv represent the cross-validation results based on the calibration data set; REPcv and R2cv represent the validation results based on the validation data set.
150
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151
Fig. 3. Scatter plots of measured versus estimated EWT and FMC for linear regression based on the optimal band combination indices. (a) Calibration data set and (b) validation data set. The dash line represents the 1:1 line, and the thick solid line represents the regression line.
generally better than that of FMC. However, the correlation between FMC and reflectance can be significantly improved by the best band combination indices, to some extent, which indicated that the selected published water indices used in the study were not suitable for FMC estimation. Finally, comparing the predictive power of the three types of hyperspectral predictors, all selected published water indices were found to be the worst performance, even worse than the single band reflectance. However, as expected, although the low REP and high R2cv were obtained from the best single narrow band, the REP and R2 were unsatisfied, which reflected the poor predictive ability of the single narrow band. The bad performance of the single narrow band reflectance and the published water indices once again demonstrate the necessity of the optimal vegetation index selection for different crop and different crop variables. However, this was the first attempt to estimate the cotton water content through two-band combination index, and the study should further our understanding of the relationships between cotton water content and vegetation indices. Although encouraging performance of the two new band indices, i.e. DR1647/DR1133 and DR1653/DR1687, for the estimation of EWT and FMC, was proven, more research is needed to test the band centers and band widths proposed in the study. Acknowledgements The work supported by West Light Foundation of the Chinese Academy of Sciences (XBBS200902), Major State Basic Research
Development Program (2009CB421301), National Natural Science Foundation of China (41101430), Knowledge Innovation Project (KZCX2-YW-BR-12), and China Postdoctoral Science Foundation (20100471681). The work assistances given by colleagues in Xinjiang Institute of Ecology and Geography Chinese Academy of Sciences are highly appreciated. References Blackburn, G.A., 1998. Spectral indices for estimating photosynthetic pigment concentrations: a test using senescent tree leaves. Int. J. Remote Sens. 19, 657– 675. Burgan, R.E., 1996. Use of remotely sensed data for fire danger estimation. EARSeL Adv. Remote Sens. 4 (4), 1–8. Carlson, J.D., Burgan, R.E., 2003. Review of user needs in operational fire danger estimation: the Oklahoma example. Int. J. Remote Sens. 24, 1601–1620. Carter, G.A., 1993. Responses of leaf spectral reflectance to plant stress. Am. J. Bot. 80, 239–243. Ceccato, P., Flasse, S., Tarantola, S., Jacquemoud, S., Gregoire, J.M., 2001. Detecting vegetation leaf water content using reflectance in the optical domain. Remote Sens. Environ. 77, 22–33. Chen, D.Y., Huang, J.F., Thomas, J.J., 2005. Vegetation water content estimation for corn and soybeans using spectral indices derived from MODIS near- and shortwave infrared bands. Remote Sens. Environ. 98, 225–236. Chuvieco, E., Cocero, D., Aguado, I., Palacios-Orueta, A., Prado, E., 2004. Improving burning efficiency estimates through satellite assessment of fuel moisture content. J. Geophys. Res. – Atmos. 109, 1–8. Curran, P.J., Dungan, J.L., Peterson, D.L., 2001. Estimating the foliar biochemical concentration of leaves with reflectance spectrometry: testing the Kokaly and Clark methodologies. Remote Sens. Environ. 76, 349–359. Danson, F.M., Bowyer, P., 2004. Estimating live fuel moisture content from remotely sensed reflectance. Remote Sens. Environ. 92, 309–321. Danson, F.M., Steven, M.D., Malthus, T.J., Clark, J.A., 1992. Highspectral resolution data for determining leaf water content. Int. J. Remote Sens. 13 (3), 461–470.
Q.-x. Yi et al. / Computers and Electronics in Agriculture 90 (2013) 144–151 Datt, B., 1999. Remote sensing of water content in eucalyptus leaves. Aust. J. Bot. 47, 909–923. Davidson, A., Wang, S., Wilmshurst, J., 2006. Remote sensing of grassland– shrubland vegetation water content in the shortwave domain. Int. J. Appl. Earth Obs. Geoinf. 8, 225–236. Dawson, T.P., Curran, P.J., Plummer, S.E., 1998. The biochemical decomposition of slash pine needles from reflectance spectra using neural networks. Int. J. Remote Sens. 19, 1433–1438. Fensholt, R., Sandholt, I., 2003. Derivation of a shortwave infrared water stress index from MODIS near- and shortwave infrared data in a semiarid environment. Remote Sens. Environ. 87, 111–121. Gao, B.C., 1996. NDWI-a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 58, 257–266. Hansen, P.M., Schjoerring, J.K., 2003. Reflectance measurement of canopy biomass and nitrogen status in wheat crops using normalized difference vegetation indices and partial least squares regression. Remote Sens. Environ. 86, 542–553. Hardisky, M.A., Klemas, V., Smart, R.M., 1983. The influences of soil salinity, growth form, and leaf moisture on the spectral reflectance of Spartina alterniflora canopies. Photogramm. Eng. Remote Sens. 49, 77–83. Hunt Jr., E.R., 1991. Airborne remote sensing of canopy water thickness scaled from leaf spectrometer data. Int. J. Remote Sens. 12, 643–649. José, R.R., David R., Eli C., Susan, U., David, R.S., 2007. Evaluation of hyperspectral reflectance indexes to detect grapevine water status in vineyards. Am. J. Enol. Vitic. 58 (3), 302–317. Kimes, D.S., Markham, B.L., Tucker, C.J., 1981. Temporal relationships between spectral response and agronomic variables of a corn canopy. Remote Sens. Environ. 11, 401–411. Knipling, E.B., 1970. Physical and physiological basis for the reflectance of visible and near-infrared radiation from vegetation. Remote Sens. Environ. 1, 155–159. Maki, M., Ishiahra, M., Tamura, M., 2004. Estimation of leaf water status to monitor the risk of forest fires by using remotely sensed data. Remote Sens. Environ. 90, 441–450. Peñuelas, J., Filella, I., Biel, C., Serrano, L., Save, R., 1993. The reflectance at the 950– 970 nm region as an indicator of plant water status. Int. J. Remote Sens. 14, 1887–1905. Peñuelas, J., Gamon, J.A., Fredeen, A.L., Merino, J., Field, C.B., 1994. Reflectance indices associated with physiological changes in nitrogen and water limited sunflower leaves. Remote Sens. Environ. 48, 135–146.
151
Peñuelas, J., Pinol, J., Ogaya, R., Filella, I., 1997. Estimation of plant water concentration by the reflectance water index WI (R900/R970). Int. J. Remote Sens. 14, 1887–1905. Pu, R., Ge, S., Kelly, N.M., Gong, P., 2003. Spectral absorption features as indicators of water status in coast live oak (Quercus agrifolia) leaves. Int. J. Remote Sens. 24, 1799–1810. Rock, B.N., Vogelmann, J.E., Williams, D.L., Vogelmann, A.F., Hoshizaki, T., 1986. Remote detection of forest damage. J. Biosci. 36, 439–445. Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W. 1974. Monitoring vegetation systems in the Great Plains with ERTS. In: Proceedings, Third Earth Resources Technology Satellite-1 Symposium. Greenbelt: NASA SP-351, pp. 301–317. Stimson, H.C., Breshears, D.D., Ustin, S.L., Kefauver, S.C., 2005. Spectral sensing of foliar water conditions in two co-occurring conifer species: pinus edulis and Juniperus monosperma. Remote Sens. Environ. 96, 108–118. Strachan, I.B., Pattey, E., Boisvert, J.B., 2002. Impact of nitrogen and environmental conditions on corn as detected by hyperspectral reflectance. Remote Sens. Environ. 80, 213–224. Thenkabail, P.S., Smith, R.B., Pauw, E.D., 2000. Hyperspectral vegetation indices and their relationships with agricultural crop characteristics. Remote Sens. Environ. 71, 158–182. Thomas, J.R., Namken, L.N., Oerther, G.F., Brown, R.G., 1971. Estimating leaf water content by reflectance measurements. Agron. J. 63, 845–847. Tian, Q., Tong, Q., Pu, R., Guo, X., Zhao, C., 2001. Spectroscopic determinations of wheat water status using 1650–1850 nm spectral absorption features. Int. J. Remote Sens. 22, 2329–2338. Tucker, C.J., 1980. Remote sensing of leaf water content in the near infrared. Remote Sens. Environ. 10, 23–32. Ustin, S.L., Roberts, D.A., Pinzón, J., Jacquemoud, S., Gardner, M., Scheer, G.C., 1998. Estimating canopy water content of chaparral shrubs using optical methods. Remote Sens. Environ. 65, 280–291. Van Niel, T.G., McVicar, T.R., Fang, H., Liang, S., 2003. Calculating environmental moisture for per-field discrimination of rice crops. Int. J. Remote Sens. 24, 885–890. Zarco-Tejada, P.J., Miller, J.R., Noland, T.L., Mohammed, G.H., Sampson, P.H., 2001. Scaling-up and model inversion methods with narrow-band optical indices for chlorophyll content estimation in closed forest canopies with hyperspectral data. IEEE Trans. Geosci. Remote Sens. 39 (7), 1491–1507. Zarco-Tejada, P.J., Rueda, C.A., Ustin, S.L., 2003. Water content estimation in vegetation with MODIS reflectance data and model inversion methods. Remote Sens. Environ. 85, 109–124.