Brown and green LAI mapping through spectral ... - Semantic Scholar

International Journal of Applied Earth Observation and Geoinformation 35 (2015) 350–358

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and Geoinformation journal homepage: www.elsevier.com/locate/jag

Brown and green LAI mapping through spectral indices Jesús Delegido ∗ , Jochem Verrelst, Juan P. Rivera, Antonio Ruiz-Verdú, José Moreno Laboratorio de Procesamiento de Imágenes, Universidad de Valencia, C/Catedrático José Beltrán, 2, 46980 Paterna (Valencia), Spain

a r t i c l e

i n f o

Article history: Received 23 August 2014 Accepted 3 October 2014 Keywords: Brown LAI Vegetation indices Hyperspectral Agroecosystem Senescent vegetation Sentinel-2

a b s t r a c t When crops senescence, leaves remain until they fall off or are harvested. Hence, leaf area index (LAI) stays high even when chlorophyll content degrades to zero. Current LAI approaches from remote sensing techniques are not optimized for estimating LAI of senescent vegetation. In this paper a two-step approach has been proposed to realize simultaneous LAI mapping over green and senescent croplands. The first step separates green from brown LAI by means of a newly proposed index, ‘Green Brown Vegetation Index (GBVI)’. This index exploits two shortwave infrared (SWIR) spectral bands centred at 2100 and 2000 nm, which fall right in the dry matter absorption regions, thereby providing positive values for senescent vegetation and negative for green vegetation. The second step involves applying linear regression functions based on optimized vegetation indices to estimate green and brown LAI estimation respectively. While the green LAI index uses a band in the red and a band in the red-edge, the brown LAI index uses bands located in the same spectral region as GBVI, i.e. an absorption band located in the region of maximum absorption of cellulose and lignin at 2154 nm, and a reference band at 1635 nm where the absorption of both water and dry matter is low. The two-step approach was applied to a HyMap image acquired over an agroecosystem at the agricultural site Barrax, Spain. © 2014 Elsevier B.V. All rights reserved.

Introduction Leaf area index (LAI), defined as half of the total developed area of leaves per unit of ground area, is one of the most important biophysical parameters of vegetation, providing valuable information on key processes such as plant production or plant capacity for carbon absorption (Duveiller et al., 2012; Kokaly et al., 2009; Jonckheere et al., 2004). However, LAI field measurements are cumbersome and unsuitable for estimating its spatial distribution, and therefore LAI estimation methods have been developed that rely on optical remote sensing technology (e.g., Bsaibes et al., 2009; Haboudane et al., 2004). Optical remote sensing methods can be categorized as either physically-based or statistically-based. Statistical retrieval methods typically relate a biophysical parameter of interest (e.g. LAI) with (transformed) spectral data through linear (e.g. based on vegetation indices) or nonlinear (e.g. machine learning) regression techniques (Delegido et al., 2013; Glenn et al., 2008; Verrelst et al., 2013). Their strengths lie in its simplicity and computational speed, however these methods are only valid for data ranges similar to those used for its development (Bsaibes et al., 2009; Li and Wang, 2013). Physically-based retrieval methods, which refer to the inversion of reflectance models against

∗ Corresponding author. Tel.: +34 963543112; fax: +34 963543261. E-mail address: [email protected] (J. Delegido). http://dx.doi.org/10.1016/j.jag.2014.10.001 0303-2434/© 2014 Elsevier B.V. All rights reserved.

remote sensing observations, on the other hand, are more generally applicable. However, these methods more often than not are computationally demanding. It usually requires quite some auxiliary information to enable the parameterization of the physical model and the description of the boundary conditions for which the model is valid (Bsaibes et al., 2009). Besides, the difficulty in describing canopy structure increases in heterogeneous scenes, such as mosaics of crops at different phenological stages or complex mixtures of woodlands and/or grasslands. Regardless whether the retrieval method originates from a statistical or physical nature, the large majority of these methods have been developed for retrieving LAI from green vegetation only. LAI of brown or senescent vegetation is usually ignored or the algorithm tends to perform poorly because is not calibrated for senescent vegetation (e.g., Kokaly et al., 2009; Delegido et al., 2013). In many applications it is also necessary to explicitly distinguish between dry or brown vegetation LAI and green vegetation LAI. For instance, in various agronomic or environmental studies, in which the objective is the estimation of photosynthesis and carbon fixation rates, authors define a green LAI (Broge and Leblanc, 2000; Haboudane et al., 2004; Delegido et al., 2011), or introduce the term Green Area Index (GAI) instead of LAI (Duveiller et al., 2012). While the role of green LAI is recognized, the role of brown LAI should not be ignored lightly. During summertime annual crops can reach a senescent state. This is especially the case for rainfed agroecosystems in temperate or Mediterranean climates. In these

J. Delegido et al. / International Journal of Applied Earth Observation and Geoinformation 35 (2015) 350–358

agroecosystems the warm dry summer months cause annuals to senescence, and sometimes this process starts as soon as on May. Although leaves remain on the plant until harvesting or falling off, it leads to degradation of chlorophyll content, i.e. reducing the greenness of the plant. Hence, contrary to chlorophyll content, LAI stays rather stable during senescing, even when their chlorophyll content dropped to zero and the plant is no longer photosynthesizing. These senescent leaves represent a significant amount of aboveground biomass and are a key factor in the carbon cycle (Vuichard et al., 2007). Quantification of brown LAI can thus help closing the aboveground carbon budget. Moreover, knowledge of brown LAI is of interest for the farmer because it may imply that crops died before being harvested, e.g. due to damage caused by drought or other natural hazards. Senescent fields are also more fire-prone than green fields; knowledge of brown LAI can contribute to fire risk assessments. Given this all, it would be incorrect to interpret senescent fields with a LAI close to zero, as is usually done when applying a green LAI approach (e.g., Haboudane et al., 2004; Delegido et al., 2013). Instead, an explicit separation between green LAI and brown LAI is strongly welcomed. One reason why a long history of green LAI assessment methods have been presented as opposed to brown LAI, is that green LAI is sensitive to the spectral region that is covered by traditional optical sensors. The majority of proposed methods make use of the visible part of the spectrum (between 350 and 700 nm), where most photosynthetic pigments absorb (Kokaly et al., 2009; Ustin et al., 2009), as well as the near-infrared (NIR) (Houborg et al., 2007) and specifically the red-edge region, in which the slope of the reflectance spectra is maximum (Delegido et al., 2013; Ustin et al., 2009). Hence, in the last decades various LAI-sensitive vegetation indices have been developed using combinations of these bands, enabling the remote sensing estimation of green LAI (e.g., Gu et al., 2013; Haboudane et al., 2004; Darvishzadeh et al., 2008; Zheng and Moskal, 2009). On the other hand, when moving away from the visible and near infrared (VNIR) towards shortwave infrared (SWIR), in the region between 1000 and 2400 nm other leaf components absorb and scatter light, such as water, cellulose, lignin and several other biochemical constituents (Curran, 1989; Brown et al., 2000; Kokaly et al., 2009; Thenkabail et al., 2013). This region of the vegetation spectrum allows the identification of vegetation stress due to drought (Serrano et al., 2000; Fensholt and Sandholt, 2003; Othman et al., 2014). The SWIR region also shows great potential for discriminating green vegetation from senescent vegetation (Daughtry et al., 2005), and for estimating the LAI of senescent vegetation. But so far the applicability of this region has hardly been studied in-depth. Only recently a few studies have demonstrated its potential for estimating (green) LAI (Gonsamo, 2011; Heiskanen et al., 2013). The reason why the SWIR region has long been ignored is that traditional optical sensors were limited to the silicon detector range (400–1000 nm). With the advent of new InGaAs array spectrometers with continuous radiometric data in the 350–2400 nm range in space, advanced tools for monitoring vegetation properties are provided. It is therefore hypothesized that imaging spectrometers operating in this full range possess capabilities to accurately estimate LAI not only over green but also over senescent crops. Until now imaging spectrometers that detect vegetation biophysical properties are mounted onto aircrafts (e.g. AVIRIS, CASI, HyMap, APEX) or onto experimental satellite missions (e.g. HYPERION, HICO, CHRIS). While these sensors do not have the capability of automated vegetation monitoring, in near future operational space missions are being planned for routinely monitoring land surfaces (e.g. the German’s EnMAP mission, NASA’s HyspIRI mission). Moreover, upcoming superspectral resolution sensors (more than 10 and less than 50 bands, i.e. in-between multispectral and hyperspectral resolution) onboard of new generation Earth observation

351

spacecrafts also sample the SWIR. The forthcoming Sentinel-2 constellation operated by the European Space Agency (ESA) has been configured with 13 bands covering the VNIR to SWIR spectral range and is, amongst others, optimized for agroecosystems monitoring applications (Drusch et al., 2012; Verrelst et al., 2012). Altogether, these upcoming spaceborne missions open an unprecedented data availability for vegetation properties monitoring. This brings us to the main objective of this study, which is the development of a spectral method based on hyperspectral data that enables the separation and estimation of both green and brown LAI in a single remote sensing image. Subsequently, the second objective is to optimize a spectral index for estimating brown LAI. The final objective is to map the separately derived green and brown LAI over an agroecosystem at Barrax, Spain.

Methods The most widely used index for distinguishing green vegetation from other land covers in remote sensing images is the Normalized Difference Vegetation Index (NDVI) (Rouse et al., 1973). The NDVI expresses the normalized ratio between the reflected energy in the red region and the reflected energy in the near infrared region, and provides an indicator of the vegetation ‘greenness’ (Koltunov et al., 2009; Delegido et al., 2013). However, NDVI values for sparse green vegetation are very similar to those of bare soils or dry vegetation due to the similarity of the spectrum in the VIS and NIR regions for bare soil and dry vegetation (Kokaly et al., 2009), which makes this index unsuitable for distinguishing green from brown LAI. When it comes to the development of an index capable of discriminate between green and brown vegetation and quantifying their LAI, this paper proposes a two-step approach. The first step involves identifying the spectral regions where the difference between both vegetation covers is maximized. This information will then serve to the development of a spectral index. Once having green vegetation from brown vegetation discriminated, as a next step, for both vegetation types an optimized LAI retrieval function will be developed. As for green vegetation a wide range of LAI retrieval methods have been presented in literature, we will make use of a recently developed index that proved to be successful over the agroecosystem Barrax, Spain (Delegido et al., 2013). The spectral regions that could serve for discriminating between green vegetation, dry vegetation and bare soils, were examined. In the visible region (400–700 nm) the profiles of green and brown vegetation are clearly distinct, but the profile between dry vegetation and soil is alike (see Fig. 1). In the SWIR region, significant differences can be observed around 2000 nm, which is a region of strong water absorption (Kokaly et al., 2009; Thenkabail et al., 2013). Brown vegetation has unique absorption near 2100 nm and 2300 nm due to its dry matter which is associated with cellulose and lignin (Jacquemoud et al., 2000; Daughtry, 2001; Kokaly et al., 2009; Feret et al., 2008). The specific absorption features of chlorophyll, water and dry matter, the latter mainly caused by cellulose and lignin is shown in Fig. 1 (Jacquemoud et al., 2000). These absorptions feature forms the basis for differentiation of brown vegetation from green vegetation or soils. By exploiting the differences in the spectral region around 2100 nm, various authors proposed a diversity of indices for estimating dry-matter contents (Daughtry et al., 2005; Romero et al., 2012; Wang et al., 2013). Green vegetation reflectance always increases in the 1900–2200 nm interval, whereas dry vegetation shows a relative maximum around 2000 nm and a relative minimum at 2100 nm. Conversely, the soil profile exhibits a smooth profile in this region. Given these different absorption features, it should be possible to discriminate green vegetation from brown vegetation by means of a simple spectral algorithm. The spectral region around 2000 nm

352

J. Delegido et al. / International Journal of Applied Earth Observation and Geoinformation 35 (2015) 350–358

Fig. 1. Specific absorption coefficient of chlorophyll (cm2 /␮g) (left scale) and water and dry matter (cm2 /g) (right scale). Adapted from Jacquemoud et al. (2000). Inside the figure: typical reflectance spectra of green vegetation, dry (brown) vegetation and bare soils (from Clark et al., 2003).

seems suitable for making such distinction. Accordingly, we propose the so-called ‘Green Brown Vegetation Index’ (GBVI): R2000 − R2100 GBVI = R2000

all possible two-band narrowband combinations according to the NDVI formulation (Stagakis et al., 2010; Delegido et al., 2011): NDIa–b =

(1)

where Ri is the reflectance at the band centred at a given wavelength i (in nm). In this index, dry vegetation is expected to reach positive values (since R2100 < R2000 ), green vegetation negative values and bare soils around zero. The difference is normalized by R2000 as reference, because 2000 nm is a wavelength where there is no absorption of cellulose and lignin (Kokaly et al., 2009). Hence, the GBVI can be used to distinguish green vegetation from brown vegetation. To support the functioning of GBVI in a more general setting, a simulation exercise was conducted using a physically-based plant reflectance model. The integrated Soil-Leaf-Canopy (SLC) model (Verhoef and Bach, 2007) was chosen because of its ability to simulate Top-Of-Canopy reflectance over the spectral between 400 and 2500 nm for vegetation consisting of either green or brown leaves or a mixture of both. For green vegetation, the following concentrations were fixed: chlorophyll content Chl = 50 ␮g/cm2 , equivalent water thickness = 0.05 cm, dry matter content = 0.005 g/cm2 and senescent material = 0.05. For brown canopies: chlorophyll content Chl = 5 ␮g/cm2 , equivalent water thickness = 0.002 cm, dry matter content = 0.012 g/cm2 and senescent material = 0.9. For both of green and brown vegetation the following variables were kept fixed: leaf structural parameter (N) = 1.25; spherical leaf angle distribution; soil moisture content = 5%; hot spot = 0.01; solar zenith angle = 30◦ ; observer zenith angle = 0◦ and azimuth angle = 0◦ . These simulations enable visualizing the different spectral responses of green and brown vegetation. Once both vegetation types have been successfully separated, the following step involves quantifying their LAI, i.e. both for green and brown vegetation independently. One of the recently proposed indices for the estimation of green LAI in heterogeneous areas, is the generic Normalized Difference Index (NDIa–b ). The NDI calculates

Rb − Ra Rb + Ra

(2)

where Ra,b are the reflectance values in the a and b bands over the entire spectral range. An earlier study using the same dataset found that the bands providing the best linear relationship between NDI and green LAI are a = 674 and b = 712 nm, located in the region of maximum chlorophyll absorption and in the red edge, respectively (Delegido et al., 2011). The NDI674–712 index was validated in several field campaigns (Delegido et al., 2013) and analysed by other authors using different experimental datasets (Frampton et al., 2013). In all cases, a high correlation was observed between NDI674–712 and green LAI. Moreover, these bands are positioned closely to the Sentinel-2 band B4 (665 nm) and the new red-edge band B5 (705 nm) band, which implies that green LAI can be estimated on an operational basis (Delegido et al., 2011). Similarly, in order to identify a spectral index with optimized sensitivity to brown LAI, it requires that again all possible band combinations are systematically analyzed. We therefore applied the NDI-type index (Eq. (2)) to search for the best performing twoband combination. For this purpose, the analysis is based on a brown LAI dataset that was collected in agricultural senescent fields that were overflown with an airborne hyperspectral sensor. Experimental dataset We used data from earlier European Space Agency (ESA) campaigns DAISEX (1999 and 2000) and SPARC (2003). The three campaigns were carried out in the Barrax agricultural test site (Southeast Spain) and encompassed biophysical field measurements, radiometry and remote sensing images acquired over the area with airborne and spaceborne sensors (ESA, 2004). In each campaign, LAI was measured in plots in fields of different crops and growing conditions, using a LICOR LAI-2000 (Welles and Norman, 1991). During the DAISEX campaigns, LAI was measured in dry barley, corn and sugar beet plots in the summer of 1999, and corn, sugar

J. Delegido et al. / International Journal of Applied Earth Observation and Geoinformation 35 (2015) 350–358

353

Fig. 2. SCL simulated green (a) and brown (b) vegetation spectra as a function of LAI.

beet and alfalfa plots in the summer of 2000. During the SPARC campaign (summer of 2003), LAI was measured in alfalfa (lucerne), corn, potato, sugar beet, garlic, onion, opium poppy and dry wheat plots (Gandía et al., 2004). LAI samples of senescent crops were collected only during DAISEX 99 and SPARC 03; DAISEX 00 only covered crops in vegetated state. The large variability of crops and growing conditions ensures that derived relationships with spectra data are applicable over complete agroecosystems, i.e. complete remote sensing images. In each of the three campaigns the study area was overflown with the airborne hyperspectral HyMap sensor (Cocks et al., 1998) operated by the German Aerospace Center (DLR), thereby acquiring images with 1 m spatial resolution and 128 bands covering the 450–2500 nm spectral region. The L1b calibrated images (radiances at sensor) were subsequently atmospherically corrected by DLR (Richter, 2001) and then georeferenced. From the joint set of field samples and image data, we subsequently obtained a database of 222 sampling points, which encompassed LAI measurements and corresponding hyperspectral observations obtained from HyMap pixels. Of these sampling points, 18 corresponded to bare soils (of different dates), 6 to wheat and 14 to barley, both at the end of their life cycle in senescent stage (i.e., a total of 20 brown LAI points) and 184 points of growing crops (i.e., green LAI). Results The difference between green and brown (senescent) vegetation is first illustrated based on simulated data. Fig. 2 shows SLC reflectance simulations of (a) green vegetation and (b) brown vegetation. It is not a surprise that large differences between green and brown can be observed in the visible part of the spectra, i.e. in the chlorophyll absorption range. The figure also supports the functioning of the GBVI: in the brown vegetation (Fig. 2b) a relative maximum at 2000 nm and a minimum around 2100 nm can be observed, while for green vegetation (Fig. 2a) the spectrum keeps steadily increasing between 2000 and 2100 nm. From the brown vegetation reflectance profiles it can be further deduced that the regions around 1500 nm and 2500 nm, i.e. where the maximum spectral differences due to LAI occurs, are the best candidates for applying a NDI-type index sensitive to LAI. The difference between green and brown (senescent) vegetation is subsequently illustrated based on experimental data. Fig. 3 shows HyMap spectral profiles of different green and brown (senescent) crops. Among the green crops, corn corresponds to a low LAI

plot (green LAI = 0.2), which shows a typical bare soil spectrum. The green crops show a reflectance maximum around 550 nm and a minimum around 650 nm, whereas the low green LAI (corn) spectra look similar to those of bare soil (not shown). In the infrared part spectra region, the brown crops have a maximum around 2000 nm, while in the green crops the reflectance increases in the 1900–2200 nm interval. Both observations correspond with comparable measurements reported by different authors (Kokaly et al., 2009; Clark et al., 2003) and with the SLC simulations (Fig. 2). Dry vegetation spectra (Fig. 3b) reveal two peaks centered at 2000 and 2200 nm due to absorption of cellulose and lignin around 2100 and 2300 nm. The next step involved calculating GBVI using HyMap reflectance spectra for pixels matching the LAI field measurements. The LAI measurements have been subsequently plotted against GBVI values in Fig. 4a, for the two campaigns (DAISEX 99 and SPARC 03) with data labeled for green vegetation, brown vegetation and bare soil (LAI = 0). In the scatterplot it can be observed that the LAI points corresponding to brown vegetation appear clearly separated as a cluster of positive GBVI values. Bare soil points concentrate around zero values, while green vegetation LAI points led to negative GBVI values. To compare with, in Fig. 4b the same LAI measurements have been plotted against the conventional greenness index NDVI as calculated from the HyMap reflectance spectra. The typical variation curve of green LAI points against NDVI values can be observed (Bsaibes et al., 2009) with NDVI saturation at high values (for NDVI greater than 0.8, green LAI varies between 2 and 6). More worrying is that NDVI values of around 0.2, correspond both to dry (brown) crops with high LAI and green crops with low LAI (in which the observed spectra is a mixture of a low proportion of green vegetation and a high proportion of bare soil). This observation confirms that NDVI is unable to distinguish between brown and green vegetation, especially when the vegetation coverage is low. Altogether, from the experimental observations we can conclude that the GBVI possesses the capability to differentiate between green and brown vegetation. The NDI674–712 , calculated from HyMap spectra using Eq. (2), was firstly plotted against all LAI measurements (including brown and green vegetation and bare soils) in Fig. 5. The figure shows that both the wheat points from SPARC 03 and the barley points from DAISEX 99 (brown vegetation) differs clearly from the green vegetation and bare soil points. Both data clusters (green and brown) can be independently fitted to a regression line, leading to a high goodness-of-fit as expressed by the coefficient of determination

354

J. Delegido et al. / International Journal of Applied Earth Observation and Geoinformation 35 (2015) 350–358

Fig. 3. Actual reflectance spectra extracted from HyMap images. (a) Green vegetation, (b) brown vegetation.

(R2 ) and root mean squared error (RMSE). For green vegetation the linear regression equation with LAI is: LAI green = −0.13 + 6.63 NDI674−712

R2 = 0.89 RMSE = 0.55 (3)

Although the brown vegetation points can be fitted to a linear regression line, its slope is very steep: y = −1.17 + 58.1 NDI674–712 (R2 = 0.82 and RMSE = 0.71). The risk of committing high estimation errors in brown LAI is therefore substantial, even for small estimation errors of NDI674–712 , which makes this index less suitable for brown LAI estimation.

Additionally, due to the different spectral features observed in green and brown vegetation, it is reasonable to expect that the most sensitive wavelengths to brown LAI estimation differ than those of green LAI estimation. Considering the analysis in the previous section, the regions around 800 nm and 2000 nm seem more suitable for estimating brown LAI than the red/red edge wavelengths used for estimating green LAI. Specifically the 2000–2400 nm region, characterized by cellulose and lignin absorption features, may be of key importance. In order to ascertain the best combination of wavelengths for optimizing the brown LAI NDIa–b index (Eq. (2)), we generated a correlation matrix of regression results between brown LAI field measurements and NDIa–b values as calculated by all possible

Fig. 4. LAI measurements plotted against GBVI values (a) and LAI plotted against NDVI values (b).

J. Delegido et al. / International Journal of Applied Earth Observation and Geoinformation 35 (2015) 350–358

7

355

5

6

4

5

LAI brown

3

LAI

4 3

2

1

2

Wheat-SP03 Barley-DAI99 Green-SP03 Green-DAI99 Green-DAI00 DRY GREEN

1 0

Regression Bare Soil Wheat-SP03 Barley-DAI99

0

-1 -0,05

0

0,05

0,1

0,15

0,2

0,25

NDI

-1 0

0,1

0,2

0,3

0,4

NDI

0,5

0,6

0,7

2154-1635

0,8

674-712

Fig. 5. Brown and green LAI measurements plotted against NDI674–712 values.

two-band combinations from HyMap data. Bare soil was added to the dataset in order to keep a reference at zero LAI. The correlation matrix is shown in Fig. 6. Only R2 results above 0.9 are highlighted in order to visually enhance the regions with maximal sensitivity. Distinct regions of excellent correlation can be observed. Two regions of high correlation appearing at the matrix top, share a common wavelength “a”: one around 2100 nm and the other around 2300 nm, while the value of “b” varies over a wide range from 800 to 2200 nm. Another region of high correlations appears in the lower left corner of the matrix, with “a” about 800 nm and “b” near 700 nm, which includes the red edge. To derive the most sensitive band combination that estimates brown LAI, we compared the 50 best performing two-band NDIa–b indices of the correlation matrix (R2 > 0.95). These indices revealed that the vast majority of them are characterized by an “a” band located between 2100 and 2170 nm and “b” band located between 1550 and 1700 nm. A second group of highly sensitive NDIa–b

Fig. 7. Brown LAI measurements plotted against NDI2154–1635 values.

indices appear with an “a” band located between 2290 and 2310 nm and “b” band located between 1550 and 2200 nm. A third small group of sensitive indices is characterized by an “a” band located between 840 and 870 nm and a “b” band located between 750 and 780 nm, i.e. the red edge. Altogether, the highest correlation with LAI measurements of brown vegetation is obtained with bands in the SWIR centred at 2154 and 1635 nm, and therefore the best expression of the index is NDI2154–1635 . Since in dry vegetation reflectance is more reduced at 2100 nm than at 1600 nm, the R1635 –R2154 values will result into positive index values. In Fig. 7, LAI measurements of brown vegetation and bare soils have been plotted against corresponding NDI2154–1635 values, which leads to the following linear regression equation: LAIbrown = −0.008 + 16.44 NDI2154−1635 RMSE = 0.34

(4)

Considering that the y intercept is close to 0 and the R2 is not affected when removing the constant, the equation can be simplified as: LAI brown =

Fig. 6. Correlation matrix (R2 ) between brown LAI measurements and two-band combinations from HyMap spectra according to NDIa–b index with a (y axis) and b (x axis) are reflectance bands. Correlations with R2 < 0.9 are color-scaled dark blue.

R2 = 0.96

16.44(R1635 − R2154 ) R1635 + R2154

(5)

It must be remarked that the regression line in Fig. 7 is highly biased by the values corresponding to bare soils (with LAI = 0), since sampling points of brown LAI between 0 and 2.5 were missing in the database. This is a typical situation in most field campaigns, in which LAI is not usually measured for sparse senescent vegetation. If we eliminate the bare soil points, a linear correlation between NDI2154–1635 and brown LAI can still be drawn, and the slope and y-intercept are kept almost equal (16.7 instead of 16.4 and −0.03 instead of −0.008) but the correlation coefficient decreases substantially (R2 from 0.96 to 0.45 and RMSE from 0.34 to 0.36). Also of interest is that the evaluated best performing bands 1635 nm and 2154 nm, are positioned closely to the Sentinel-2 bands B11 (1610 nm) and B12 (2190 nm). However, its performance need to be evaluated since these bands are broadbands with a band width of 90 and 180 nm, respectively. The brown LAI dataset was therefore resampled to Sentinel-2 band settings and the NDIa–b correlation matrix was calculated again. The Sentinel-2 B12-B11 normalized index was top performing with a R2 of 0.95 (RMSE = 0.37) and regression equation of: −0.455 + 19.91 NDIB12–B11 . This result thus

356

J. Delegido et al. / International Journal of Applied Earth Observation and Geoinformation 35 (2015) 350–358

Fig. 8. RGB composition (top left), green LAI map (top right) and joint green and brown LAI map of the study area of the SPARC 03 campaign (down).

suggests that brown LAI can also be estimated on an operational basis. The last step involves mapping brown and green LAI over an agroecosystem based on the two-step approach. The first step distinguishes green and dry vegetation by means of the proposed GBVI (Eq. (1)), and the second step then estimates green and brown LAI by means of a NDI-type equation. The GBVI was applied to the hyperspectral HyMap image acquired July 12th of 2003 over the Barrrax agricultural site (Spain) to classify the image into brown and green vegetation types. Fig. 8 top left map shows the RGB of a subset of the HyMap flightline. The green circular fields represent croplands in vegetative state that are watered by center pivot irrigation. The brown circular fields along with brown parcels represent croplands in senescent state (e.g. wheat, barley). These surfaces look very similar to fallow land and bare soils around these croplands. Effectively, when applying a conventional (green) LAI retrieval methods, these surface will hardly be distinguished, leading to an underestimation of LAI over senescent crops. That also occurred when applying only the LAI green algorithm (Eq. (3)) to the image (Fig. 8 top right). Instead, by applying Eq. (3) to the green vegetation pixels and Eq. (5) to the dry vegetation pixels an unbiased LAI map that covers both vegetation types (green and brown) and bare soils is produced. Fig. 8 down shows the LAI map generated from the HyMap image. Particularly the pronounced spatial variation of high green LAI clearly marks the irrigated circular fields with green biomass. These irrigated fields are characterized by a green LAI above 3 and

a high chlorophyll content (Verrelst et al., 2013). The brown color surfaces represent the brown LAI, i.e., fainfed, dried out senescent croplands. Parcels with low brown LAI represent harvested cereal fields while parcels with whitish colors represent bare soils.

Discussion For the last three decades a wide variety of LAI retrieval methods have been presented from optical remote sensing data. Although only a minority of proposed approaches explicitly refer to green LAI s (e.g., Broge and Leblanc, 2000; Myneni et al., 2002; Haboudane et al., 2004; Yang et al., 2006), in fact all current globally operational LAI retrieval methods are optimized to the retrieval of green LAI (e.g., Myneni et al., 2002; Chen et al., 2002; Deng et al., 2006; Baret et al., 2007). Apart from the importance of green LAI estimation methods in many vegetation monitoring applications, the success of these methods also lies in: (1) the pronounced sensitivity of optical data within the commonly exploited VNIR spectral range, and (2) the fact that experimental or simulated calibration data is typically restricted to green vegetation. Consequently current LAI retrieval methods typically confound senescent vegetation with bare soil, leading to an underestimation of non-green LAI (Haboudane et al., 2004). To the best of our knowledge we are not aware of any LAI estimation approach that is able to quantify the LAI of senescent crops.

J. Delegido et al. / International Journal of Applied Earth Observation and Geoinformation 35 (2015) 350–358

A two-step approach has been proposed to realize simultaneous and unbiased LAI mapping over green and senescent parcels. The first step separates green from brown vegetation based on a newly proposed index, being GBVI. This index exploits SWIR spectral bands centered at 2100 and 2000 nm. Having senescent vegetation identified, in the second step its LAI is quantified through a calibrated linear regression function based on an optimized spectral index. This index uses bands located in the same spectral region as GBVI: an absorption band located in the region of maximum absorption of cellulose and lignin at 2154 nm (Daughtry, 2001), and a reference band at 1635 nm where absorption of both water and dry matter is low. Interestingly, the proposed index is closely related to previously developed indices, such as the Cellulose Absorption Index (CAI) (Nagler et al., 2000; Daughtry et al., 2005) that makes use of three bands in the SWIR: 0.5 (R2000 − R2200 ) − R2100 , or the dry matter index (R2305 − R1495 )/(R2305 + R1495 ) proposed by Romero et al. (2012) that was developed based on radiative transfer modelling with PROSPECT. These indices have been applied to estimate dry matter content, and not, although related, for brown LAI estimation. Although the robustness of the proposed spectral indices (GBVI and green/brown LAI regression functions) may benefit from additional testing in more extreme situations (e.g. in other atmospheric conditions, complex topography, other crops), the developed algorithms find their strength in their simplicity. In principle it can be run continuously in near-real time over large agricultural areas without having to rely on auxiliary data. This simplicity constitutes an important advantage over inversion of radiative transfer models. Such techniques typically need information about the crop architectural characteristics for the generation of matching crop- and phenology-specific synthetic spectra, which is not always directly available (Rivera et al., 2013). The simple regression equation is intended for rapid prediction of LAI in a straightforward way. Empirical approaches are amongst the simplest way to predict biophysical parameters, however it does not escape our attention that they provide relationships that are significantly space, time and species dependent (Verrelst et al., 2008, 2010; Zheng et al., 2014). In this respect, it must be remarked that the objective of this paper was not the development of a regression equation itself, but rather the delivering of a brown LAI index that cope with senescent vegetation types. Given that calibration occurred across a broad range of crop types at various growth stages, obtained empirical relationships are expected to be sufficiently robust for precise green and brown LAI estimations. Assessment of the NDI2154–1635 proved that the intrinsic relationship between brown LAI and the index was essentially stable; nevertheless regression coefficients can vary depending on the local situation and sensor characteristics. Henceforth, regression coefficients can be easily recalibrated. The ability to monitor green and brown LAI will be greatly improved in the near future, as more high quality satellite data will be available to us. It is thereby important to note that the wavelength ranges where the brown LAI index yield accurate results, are relatively wide: 2000–2400 nm region for the absorption bands and 800–1700 nm for the reference bands. That implies that the proposed brown LAI index can be adapted to superspectral or multispectral sensors, given the disposal of bands in these SWIR regions. An excellent candidate in this respect would be the forthcoming Sentinel-2 constellation, which encompasses two bands in the SWIR (B11: 1610 nm and B12: 2190 nm). The first Sentinel-2, is envisaged to be launched in 2015 and aims to deliver data taken over all land surfaces at a spatial resolution of 10 m, 20 or 60 m (depending on the used bands) at a high revisiting time (each 5th day under cloud-free conditions) (Drusch et al., 2012). These Sentinels will permit the use of multiple sensors to provide frequent observation to capture rapid changes to agricultural lands (Zheng et al., 2014). The GBVI is unfortunately unapplicable to

357

Sentinel-2 data to discriminate brown from green vegetation because it lacks a band in the 2000 nm region. Nevertheless, there is no doubt that with classification or unmixing techniques Sentinel-2 images can be classified into photosynthetic and nonphotosynthetic vegetation (Byambakhuu et al., 2010; Zheng et al., 2013; Mishra and Crews, 2014). Moreover, it is expected that in the near future brown LAI will be estimated with improved accuracy from forthcoming operational spaceborne imaging spectroscopy missions such as EnMap (244 spectral bands from 420 to 2450 nm) or HyspIRI (over 200 bands from 380 to 2500 nm). Given these forthcoming unprecedented high-quality data flow, it is foreseen that the LAI quantification of joint green and senescent vegetation will open new agroecosystem monitoring application in the coming years, not in the least to capacitate unbiased aboveground biomass estimation. Conclusions LAI of senescent vegetation is typically underestimated by current LAI mapping algorithms. To account for senescent vegetation, a new methodology is proposed to quantify LAI of both brown and green vegetation based on the application of two LAI indices to a previously classified image. For sensors having the adequate bands in the SWIR, the vegetation classification could be easily made by applying a new “Green Brown Vegetation Index (GBVI)”, which evaluates the reflectance slope in the 2000–2100 nm spectral interval, effectively separating brown senescent vegetation from green vegetation. The proposed green LAI index uses red and red-edge bands, while the new brown LAI index uses two SWIR bands. The combined green and brown LAI mapping was demonstrated over the agricultural site Barrax (Spain) using HyMap data. The simplicity of the method facilitates easy portability to other sites and images, e.g. as acquired from the Sentinel-2 constellation. Acknowledgments This work was possible thanks to the project AYA201021432-C02-01, financed by the Spanish Ministry of Economy and Competitiveness. The authors would like to thank the two anonymous reviewers for their valuable comments. References ˜ Baret, F., Hagolle, O., Geiger, B., Bicheron, P., Miras, B., Huc, M., Berthelot, B., Nino, F., Weiss, M., Samain, O., Roujean, J.L., Leroy, M., 2007. LAI, fAPAR and fCover CYCLOPES global products derived from VEGETATION. Part 1: Principles of the algorithm. Remote Sens. Environ. 110, 275–286. Broge, N.H., Leblanc, E., 2000. Comparing prediction power and stability of broadband and hyperspectral vegetation indices for estimation of green leaf area index and canopy chlorophyll density. Remote Sens. Environ. 76, 156–172. Brown, L., Chen, J., Leblanc, S., Cihlar, J., 2000. A shortwave infrared modification to the simple ratio for LAI retrieval in boreal forests: an image and model analysis. Remote Sens. Environ. 71, 16–25. Bsaibes, A., Courault, D., Baret, F., Weiss, M., Olioso, A., Jacob, F., Hagolle, O., Marloie, O., Bertrand, N., Desfond, V., Kzemipour, F., 2009. Albedo and LAI estimates from FORMOSAT-2 data for crop monitoring. Remote Sens. Environ. 113 (4), 716–729. Byambakhuu, I., Sugita, M., Matsushima, D., 2010. Spectral unmixing model to assess land cover fractions in Mongolian steppe regions. Remote Sens. Environ. 114, 2361–2372. Chen, J.M., Pavlic, G., Brown, L., Cihlar, J., Leblanc, S.G., White, H.P., Hall, R.J., Peddle, D.R., King, D.J., Trofymow, J.A., Swift, E., Van Der Sanden, J., Pellikka, P.K.E., 2002. Derivation and validation of Canada-wide coarse-resolution leaf area index maps using high-resolution satellite imagery and ground measurements. Remote Sens. Environ. 80, 165–184. Clark, R.N., Swayze, G.A., Livo, K.E., Kokaly, R.F., Sutley, S.J., Dalton, J.B., McDougal, R.R., Gent, C.A., 2003. Imaging spectroscopy: earth and planetary remote sensing with the USGS Tetracorder and expert systems. J. Geophys. Res. 108 (E12), 5131. Cocks, T., Jenssen, R., Stewart, A., Wilson, I., Shields, T., 1998. The HYMAP airborne hyperspectral sensor: the system, calibration and performance. In: Proceedings of the First EARSeL Workshop on Imaging Spectroscopy, 6–8 October 1998, Remote Sensing Laboratories, University of Zurich, Zurich, Switzerland, pp. 37–42.

358

J. Delegido et al. / International Journal of Applied Earth Observation and Geoinformation 35 (2015) 350–358

Curran, P.J., 1989. Remote sensing of foliar chemistry. Remote Sens. Environ. 30, 271–278. Darvishzadeh, R., Skidmore, A., Atzberger, C., Wieren, S., 2008. Estimation of vegetation LAI from hyperspectral reflectance data: effects of soil type and plant architecture. Int. J. Appl. Earth Observ. Geoinform. 10, 358–373. Daughtry, C.S.T., 2001. Discriminating crop residues from soil by shortwave infrared reflectance. Agron. J. 93, 125–131. Daughtry, C.S.T., Hunt Jr., E.R., Doraiswamy, P.C., McMurtrey III, J.E., 2005. Remote sensing the spatial distribution of crop residues. Agron. J. 97, 864–871. Deng, F., Chen, J.M., Plummer, S., Chen, M., Pisek, J., 2006. Algorithm for global leaf area index retrieval using satellite imagery. IEEE Trans. Geosci. Remote Sens. 44, 2219–2228. Delegido, J., Verrelst, J., Alonso, L., Moreno, J., 2011. Evaluation of Sentinel-2 red-edge bands for empirical estimation of Green LAI and chlorophyll content. Sensors 11, 7063–7081. Delegido, J., Verrelst, J., Meza, C.M., Rivera, J.P., Alonso, L., Moreno, J., 2013. A red-edge spectral index for remote sensing estimation of green LAI over agroecosystems. Eur. J. Agron. 46, 42–52. Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B., Isola, C., Laberinti, P., Martimort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F., Bargellini, P., 2012. Sentinel-2: ESA’s optical high-resolution mission for GMES operational services. Remote Sens. Environ. 120, 25–36. Duveiller, G., Baret, F., Defourny, P., 2012. Remotely sensed green area index for winter wheat crop monitoring: 10-year assessment at regional scale over a fragmented landscape. Agric. Forest Meteorol. 152 (166–167), 156–168. ESA, 2004. SPARC data acquisition report. Contract no: 18307/04/nl/ff. University Valencia, Available from: http://earth.esa.int/campaigns/DOC/SPARC2004 Data Acquisition Report.pdf 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. Feret, J-P., Franc¸ois, C., Asner, G.P., Gitelson, A.A., Martin, R.E., Bidel, L.P.R., Ustin, S.L., Le Maire, G., Jacquemoud, S., 2008. PROSPECT-4 und 5: advances in the leaf optical properties model separating photosynthetic pigments. Remote Sens. Environ. 112, 3030–3043. Frampton, W.J., Dash, J., Watmough, G.R., Milton, E.J., 2013. Evaluating the capabilities of Sentinel-2 for quantitative estimation of biophysical variables in vegetation. J. Photogram. Remote Sens. 82, 83–92. Gandía, S., Fernández, G., García, J.C., Moreno, J., 2004. Retrieval of vegetation biophysical variables from Chris/Proba data in the SPARC campaing. In: Proceedings of the 2nd CHRIS/Proba Workshop, ESA/ESRIN, Frascati, Italy, Available from: http://earth.esa.int/workshops/chris proba 04/papers/12 Gandia.pdf Glenn, E.P., Huete, A.R., Nagler, P.L., Nelson, S.G., 2008. Relationship between remotely-sensed vegetation indices, canopy attributes and plant physiological processes: what vegetation indices can and cannot tell us about the landscape. Sensors 8, 2136–2160. Gonsamo, A., 2011. Normalized sensitivity measures for leaf area index estimation using three-band spectral vegetation indices. Int. J. Remote Sens. 32, 2069–2080. Gu, Y., Wylie, B.K., Howard, D.M., Phuyal, K.P., Ji, L., 2013. NDVI saturation adjustment: a new approach for improving cropland performance estimates in the Greater Platte River Basin, USA. Ecol. Indic. 30, 1–6. Haboudane, D., Miller, J.R., Pattey, E., Zarco-Tejada, P.J., Strachan, I.S., 2004. Hyperspectral vegetation indices and novel algorithms for predicting green LAI of crop canopies: modeling and validation in the context of precision agriculture. Remote Sens. Environ. 90, 337–352. Heiskanen, J., Rautiainen, M., Stenberg, P., Mottus, M., Vesanto, V.H., 2013. Sensitivity of narrowband vegetation indices to boreal forest LAI, reflectance seasonality and species composition. ISPRS J. Photogr. Remote Sens. 78, 1–14. Houborg, R., Soegaard, H., Boegh, E., 2007. Combining vegetation index and model inversion methods for the extraction of key vegetation biophysical parameters using Terra and Aqua MODIS reflectance data. Remote Sens. Environ. 106, 39–58. Jacquemoud, S., Bacour, C., Poilvé, H., Frangi, J.-P., 2000. Comparison of four radiative transfer models to simulate plant canopies reflectance – direct and inverse mode. Remote Sens. Environ. 74, 471–481. Jonckheere, I., Fleck, S., Nackaerts, K., Muysa, B., Coppin, P., Weiss, M., Baret, F., 2004. Review of methods for in situ leaf area index determination. Part I. Theories, sensors and hemispherical photography. Agric. For. Meteorol. 121, 19–35. Koltunov, A., Ustin, S.L., Asner, G.P., Fung, I., 2009. Selective logging changes forest phenology in the Brazilian Amazon: evidence from MODIS image time series analysis. Remote Sens. Environ. 113, 2431–2440. Kokaly, R.F., Asner, G.P., Ollinger, S.V., Martin, M.E., Wessman, C.A., 2009. Characterizing canopy biochemistry from imaging spectrometer data for studying ecosystem processes. Remote Sens. Environ. 113, S78–S91. Li, P., Wang, Q., 2013. Developing and validating novel hyperspectral indices for leaf area index estimation: effect of canopy vertical heterogeneity. Ecol. Indic. 32, 123–130. Mishra, N.B., Crews, K.A., 2014. Estimating fractional land cover in semi-arid central Kalahari: the impact of mapping method (spectral unmixing vs.

object-based image analysis) and vegetation morphology. Geocarto Int., http://dx.doi.org/10.1080/10106049.2013.868041. Myneni, R.B., Hoffman, S., Knyazikhin, Y., Privette, J.L., Glassy, J., Tian, Y., Wang, Y., Song, X., Zhang, Y., Smith, G.R., Lotsch, A., Friedl, M., Morisette, J.T., Votava, P., Nemani, R.R., Running, S.W., 2002. Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data. Remote Sens. Environ. 83, 214–231. Nagler, P.L., Daughtry, C.S.T., Goward, S.N., 2000. Plant litter and soil reflectance. Remote Sens. Environ. 71, 207–215. Othman, Y., Steele, C., VanLeeuwen, D., Heerema, R., Bawazir, S., Hilaire, R.S., 2014. Remote sensing used to detect moisture status of pecan orchards grown in a desert environment. Int. J. Remote Sens. 35, 949–966. Richter, R.,2001, July. Atmospheric correction methodology for imaging spectrometer data. In: Proceedings of the DAISEX Final Results Workshop, ESA SP-499. ESA Publication Division, ESTEC, The Netherlands. Rivera, J.P., Verrelst, J., Leoneko, G., Moreno, J., 2013. Multiple cost functions and regularization options for improved retrieval of leaf chlorophyll content and LAI through inversion of the PROSAIL model. Remote Sens. 5 (7), 3280–3304. Romero, A., Aguado, I., Yebra, M., 2012. Estimation of dry matter content in leaves using normalized indexes and PROSPECT model inversion. Int. J. Remote Sens. 33, 396–414. Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., 1973. Monitoring vegetation systems in the Great Plains with ERTS. In: Third ERTS Symposium, NASA SP-351, vol. 1, NASA, Washington, DC, pp. 309–317. Stagakis, S., Markos, N., Sykioti, O., Kyparissis, A., 2010. Monitoring canopy biophysical and biochemical parameters in ecosystem scale using satellite hyperspectral imagery: an application on a Phlomis fruticosa Mediterranean ecosystem using multiangular CHRIS/PROBA observations. Remote Sens. Environ. 114, 977–994. ˜ Serrano, L., Ustin, S.L., Roberts, D.A., Gamon, J.A., Penuelas, J., 2000. Deriving water content of chaparral vegetation from AVIRIS data. Remote Sens. Environ. 74, 570–581. Thenkabail, P.S., Mariotto, I., Gumma, M.K., Middleton, E.M., Landis, D.R., Huemmrich, K.F., 2013. Selection of hyperspectral narrowbands (HNBs) and composition of hyperspectral twoband vegetation indices (HVIs) for biophysical characterization and discrimination of crop types using field reflectance and hyperion/EO-1 data. IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens. 6, 427–439. Ustin, S.L., Gitelson, A.A., Jacquemoud, S., Schaepman, M.E., Asner, G.P., Gamon, J.A., Zarco-Tejada, P.J., 2009. Retrieval of foliar information about plant pigment systems from high resolution spectroscopy. Remote Sens. Environ. 113, S67–S77. Verhoef, W., Bach, H., 2007. Coupled soil–leaf-canopy and atmosphere radiative transfer modeling to simulate hyperspectral multi-angular surface reflectance and TOA radiance data. Remote Sens. Environ. 109, 166–182. Verrelst, J., Alonso, L., Rivera, J.P., Moreno, J., Camps-Valls, G., 2013. Gaussian process retrieval of chlorophyll content from imaging spectroscopy data. IEEE J. Select. Top. Appl. Earth Obs. Remote Sens. 6 (2 (Pt. 3)). ˜ Verrelst, J., Munoz, J., Alonso, L., Delegido, J., Rivera, J., Camps-Valls, G., Moreno, J., 2012. Machine learning regression algorithms for biophysical parameter retrieval: opportunities for Sentinel-2 and -3. Remote Sens. Environ. 118, 127–139. Verrelst, J., Schaepman, M.E., Koetz, B., Kneubuhler, M., 2008. Angular sensitivity analysis of vegetation indices derived from CHRIS/PROBA data. Remote Sens. Environ. 112, 2341–2353. Verrelst, J., Schaepman, M.E., Malenovsky, Z., Clevers, J.G.P.W., 2010. Effects of woody elements on simulated canopy reflectance: implications for forest chlorophyll content retrieval. Remote Sens. Environ. 114, 647–656. Vuichard, N., Soussana, J.-F., Ciais, P., Viovy, N., Ammann, C., Calanca, P., CliftonBrown, J., Fuhrer, J., Jones, M., Martin, C., 2007. Estimating the greenhouse gas fluxes of European grasslands with a process-based model: 1. Model evaluation from in situ measurements. Global Biogeochem. Cycles 21 (GB1004). Wang, L., Hunt, E.R., Qu, J.J., Hao, X., Daughtry, C., 2013. Remote sensing of fuel moisture content from ratios of narrow-band vegetation water and dry-matter indices. Remote Sens. Environ. 1129, 103–110. Welles, J.M., Norman, J.M., 1991. Instrument for indirect measurement of canopy architecture. Agron. J. 83, 818–825. Yang, W., Tan, B., Huang, D., Rautiainen, M., Shabanov, N.V., Wang, Y., Privette, J.L., Huemmrich, K.F., Fensholt, R., Sandholt, I., Weiss, M., Ahl, D.E., Gower, S.T., Nemani, R.R., Knyazikhin, Y., Myneni, R.B., 2006. MODIS leaf area index products: from validation to algorithm improvement. IEEE Trans. Geosci. Remote Sens. 44 (7), 1885–1896. Zheng, B., Campbell, J.B., Serbin, G., Daughtry, C.S.T., 2013. Multitemporal remote sensing of crop residue cover and tillage practices: a validation of the minNDTI strategy in the United States. J. Soil Water Conserv. 68 (2), 120–131. Zheng, B., Campbell, J.B., Serbin, G., Galbraith, J.M., 2014. Remote sensing of crop residue and tillage practices: present capabilities and future prospects. Soil Tillage Res. 138, 26–34. Zheng, G., Moskal, L.M., 2009. Retrieving leaf area index (LAI) using remote sensing: theories, methods and sensors. Sensors 9, 2719–2745.