Exploring the potential of PROCOSINE and close-range hyperspectral imaging to study the effects of fungal diseases on leaf physiology

Abstract

The detection of plant diseases, including fungi, is a major challenge for reducing yield gaps of crops across the world. We explored the potential of the PROCOSINE radiative transfer model to assess the effect of the fungus Pseudocercospora fjiensis on leaf tissues using laboratory-acquired submillimetre-scale hyperspectral images in the visible and near-infrared spectral range. The objectives were (i) to
assess the dynamics of leaf biochemical and biophysical parameters estimated using PROCOSINE inversion as a function of the disease stages, and (ii) to discriminate the disease stages by using a Linear
Discriminant Analysis model built from the inversion results. The inversion results show that most of the parameter dynamics are consistent with expectations: for example, the chlorophyll content
progressively decreased as the disease spreads, and the brown pigments content increased. An overall accuracy of 78.7% was obtained for the discrimination of the six disease stages, with errors mainly occurring between asymptomatic samples and frst visible disease stages. PROCOSINE inversion
provides relevant ecophysiological information to better understand how P. fjiensis afects the leaf at each disease stage. More particularly, the results suggest that monitoring anthocyanins may be critical
for the early detection of this disease.

Full text

1
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
www.nature.com/scientificreports
Exploring the potential of
PROCOSINE and close-range
hyperspectral imaging to study the
eects of fungal diseases on leaf
physiology
Julien Morel1,2, Sylvain Jay3, Jean-Baptiste Féret
4, Adel Bakache1, Ryad Bendoula1,
Francoise Carreel5 & Nathalie Gorretta1
The detection of plant diseases, including fungi, is a major challenge for reducing yield gaps of crops
across the world. We explored the potential of the PROCOSINE radiative transfer model to assess the
eect of the fungus Pseudocercospora jiensis on leaf tissues using laboratory-acquired submillimetre-
scale hyperspectral images in the visible and near-infrared spectral range. The objectives were (i) to
assess the dynamics of leaf biochemical and biophysical parameters estimated using PROCOSINE
inversion as a function of the disease stages, and (ii) to discriminate the disease stages by using a Linear
Discriminant Analysis model built from the inversion results. The inversion results show that most
of the parameter dynamics are consistent with expectations: for example, the chlorophyll content
progressively decreased as the disease spreads, and the brown pigments content increased. An overall
accuracy of 78.7% was obtained for the discrimination of the six disease stages, with errors mainly
occurring between asymptomatic samples and rst visible disease stages. PROCOSINE inversion
provides relevant ecophysiological information to better understand how P. jiensis aects the leaf at
each disease stage. More particularly, the results suggest that monitoring anthocyanins may be critical
for the early detection of this disease.
Plant diseases are a major issue for food and crop production, and can lead to yield gaps between potential and
actual productions1. One way to reduce those gaps is to develop ecient and accurate monitoring systems to
detect plant diseases. To date, most of those monitoring systems rely on identication based on visual assess-
ment and expert knowledge, which is subjective, time- and money-consuming. Other methods using molecu-
lar analyses techniques such as polymerase chain reaction (PCR) have recently gained importance2, but remain
time-consuming and highly disease-specic, making them inappropriate for processing numerous plant samples.
Remote sensing has gained tremendous importance in agriculture, providing objective data that can be used
to retrieve information on the status of a crop at various spatial scales. Estimation of crop yields3,4, evapotranspi-
ration within a eld5, or assessment of the leaf chlorophyll or carotenoids contents6–8 are examples of such use.
Remotely-sensed data are also used for detecting plant foliar diseases at various spatial scales and with dif-
ferent sensors and methods9–11. Methods developed for leaf disease detection are usually based on empirical
relationships between spectral indices and the presence or degrees of intensity of the disease. Spectral indices can
be computed easily and eciently, require a limited amount of spectral information, and usually provide indirect
information on the presence of the disease by capturing the eects of symptoms (such as chlorosis) on the leaf
optical properties. Gennaro et al.12 used unmanned aerial vehicles (UAVs) to acquire very high spatial resolu-
tion data of vineyards elds, from which leaves contaminated with grapevine leaf stripe disease (Phaeomoniella
1UMR ITAP, Irstea, Montpellier SupAgro, Univ. Montpellier, Montpellier, France. 2Department of Agricultural
Research for Northern Sweden, Swedish University of Agricultural Sciences, Umeå, Sweden. 3Aix-Marseille Univ.,
CNRS, Central Marseille, Institut Fresnel, Marseille, F-13013, France. 4UMR TETIS, Irstea, Univ. Montpellier,
Montpellier, France. 5UMR AGAP, Cirad, Montpellier, France. Correspondence and requests for materials should be
addressed to J.M. (email: julien.morel@irstea.fr)
Received: 10 April 2018
Accepted: 6 October 2018
Published: xx xx xxxx
OPEN
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
2
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
chlamydospora) could be discriminated from healthy leaves based on Normalized Dierence Vegetation Index
(NDVI) values. However, the authors outlined that such a method is only reliable if no other factors aect leaf
chlorophyll. Indeed, as various biotic and abiotic stresses may aect leaf chlorophyll content (e.g., nitrogen stress,
pests, etc.), the NDVI cannot discriminate a specic disease from other stresses. Dierent methods have been
tested to overcome this limitation. Lu et al.13 used a non-imaging spectrometer to compute spectral vegetation
indices and discriminate various tomato leaf diseases using a K-nearest neighbour classier. Although the method
showed promising results, it was limited by the diculty to specically measure the reectance on the disease
spots, creating mixed healthy and infected spectra. Mahlein et al.14 used hyperspectral signatures of healthy and
infected sugar beet leaves acquired in laboratory conditions with a non-imaging spectrometer to develop specic
disease indices. However, such indices may be aected by directional eects induced by the leaf surface, and the
obtained empirical relationships may show moderate extrapolative abilities, e.g., when applied to other crops or
dierent sensors.
Another widely used method to detect foliar diseases based on remote-sensing data relies on the development
of multivariate regression models based on the whole reectance spectrum. Such methods allow the full spectral
information contained in the visible and near infrared (VNIR) and/or the shortwave infrared spectral range
(SWIR) to be taken into account, instead of only a limited number of spectral bands, as when using spectral indi-
ces. Delalieux et al.15 acquired hyperspectral data of healthy and Venturia inaequalis-infected apple leaves with a
non-imaging spectroradiometer, and compared several multivariate statistical approaches. ey showed that the
SWIR range contains relevant information for the early discrimination of healthy and infected leaves. Yeh et al.16
used hyperspectral images acquired in a laboratory on healthy and Colletotrichum gloeosporioides-infected straw-
berry leaves, and compared various multivariate methods for discriminating between healthy and infected leaves.
eir results showed that it was possible to select optimal wavelengths in the VNIR range in order to discriminate
healthy leaves from those contaminated at incubation (i.e., early) and symptomatic stages.
Although multivariate regression models are powerful tools for the detection of plant diseases, these tech-
niques are limited by the exhaustiveness of the database used to calibrate the models. Furthermore, similarly to
spectral indices, such models remain potentially sensitive to light conditions and vegetation structure at various
scales, including leaf geometry. As a consequence, statistical models derived from spectral indices or full spectral
analyses may show limited generalisation abilities, because of the multiple factors inuencing reectance that
must be taken into account when producing a training dataset. Alternatively, physically-based models are prom-
ising tools to overcome these limitations, since their inversions allow disentangling the inuences of multiple fac-
tors on spectral information, including leaf biochemistry, leaf and canopy structures, and acquisition conditions.
Radiative transfer models (RTM) describe the interactions between light and matter. In the case of vegetation,
these interactions occur from molecular scale, due to the absorption of photons by leaf biochemical constituents
such as pigments, to larger scales, due to the scattering of photons within the leaf and canopy internal struc-
tures. Although RTMs are designed to simulate the interaction of light with healthy and senescent leaves, some
authors used such models for the detection of foliar diseases. Caldéron et al.17 used the PROSAIL model18–20 as a
validation tool to study the eects of varying chlorophyll content and leaf area index on simulated spectral indi-
ces, including NDVI and the R550/R670 ratio. ose simulated indices were compared with multispectral images
acquired with UAVs in poppy elds contaminated with downy mildew (Peronospora arborescens). Albetis et al.21
used multispectral images acquired with a UAV to detect avescence dorée disease (Candidatus Phytoplasma) in
vineyards. eir results showed that, in the case of red cultivars, the anthocyanin content derived from PROSAIL
inversion is a relevant proxy for the discrimination of avescence dorée symptomatic leaves against the asympto-
matic ones. However, to our knowledge, there have been no previous attempts to study the impact of a disease on
leaf structure and biochemistry based on physical modelling applied to remote sensing. Such knowledge would
help to better understand the interactions between the leaf host and the pathogen, which could then be used to
develop methods for, e.g., early disease detection (i.e., when symptoms are still not visible to the naked eye) or for
discriminating a disease-induced stress from other biotic or abiotic stresses.
In this paper, we used PROSPECT-D22 combined with COSINE23 (PROCOSINE) in order to estimate leaf bio-
chemical and structural parameters from close-range imaging spectroscopy of banana leaves, and to identify the
degree of infection from black leaf streak disease (hereinaer referred to as BLSD) a foliar disease caused by the
Dothideomycete fungus Pseudocercospora jiensis (previously Mycosphaerella jiensis). BLSD is a major disease
aecting bananas24, rst identied in the Sigatoka Valley in Fiji in 196325. It causes leaf necrotic lesions that even-
tually lead to (1) a yield reduction of up to 50%, due to the reduction of photosynthesis26,27, and (2) a premature
ripening of bunches, making the bananas inappropriate for export and commercialisation28,29. e objective of
this exploratory study was to assess the potential of the combined PROCOSINE model to (i) monitor the changes
in leaf structure and biochemistry caused by the development of the disease and (ii) identify the degree of infec-
tion as estimated by expert knowledge. Indeed, discriminating early infection stages from the later ones can help
to decide on the protection strategy to adopt, i.e., preventive (by using contact fungicides) for early stages or
curative (by using systemic fungicides) for later stages. Disks of banana leaves at dierent stages of infection from
Pseudocercospora jiensis were imaged in laboratory conditions with a VNIR hyperspectral camera. Biochemical
and biophysical foliar properties, including chlorophyll (Cab), carotenoids (Ccx), anthocyanins (Cant), and brown
pigments (Cbp) contents, structure parameter (N), light incident angle (θi) and specular reection (bspec) were then
estimated by model inversion on a pixel basis, leading to submillimeter maps of PROCOSINE parameters. Based
on these maps, the eect of BLSD on the ecophysiology of the leaves was studied. Finally, a linear discriminant
analysis was performed on the inverted parameters in order to discriminate the disease stages.
Results
Inuence of the disease on leaf reectance. e mean reectance spectra of asymptomatic and BLSD-
infected leaf tissues were rst computed based on pixels extracted from hyperspectral images (Fig.1).
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
3
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
e reectance spectra corresponding to asymptomatic pixels were consistent with our expectations: in the
VIS domain, the reectance was minimum in the blue and red spectral domains, and maximum in the green
spectral domain. It showed a sharp increase from 680 nm to 740 nm (the so-called red edge), before reaching a
plateau in the NIR domain (740 nm to 900 nm).
Overall, the reectance decreased in the VIS domain from 400 nm to 650 nm with the development of P.
jiensis. is decrease was more pronounced in the green spectral domain (around 550 nm) than in the blue
(around 450 nm) and red (around 650 nm) ones. A slight increase in reectance associated with the presence of
the disease was observed in the domain from 650 nm to 700 nm. For longer wavelengths and towards the NIR
domain, P. jiensis strongly impacted the measured reectance, as the sharp shoulder of the red edge around
750 nm observed for asymptomatic leaves tended to smoothen with increasing disease stage, and completely
disappeared for stages 5 and 6 leaves. In the NIR domain, the reectance showed a slight increase for stage 3, and
a more pronounced decrease for stages 4, 5 and 6.
Dynamics of the inverted parameters. PROCOSINE was then inverted for each pixel of the image data-
base. e estimated PROCOSINE parameters were then used to compute simulated spectra, from which was
computed the reconstruction error. Example maps of inverted parameters and reconstruction errors are pre-
sented in Fig.2 for illustrative purposes. Note that the results obtained for Cw and Cm are not shown here, due to
their limited inuence on reectance in the VNIR domain. Error spectra as a function of the wavelength are also
presented in Supplementary Fig.1.
BLSD spots did not appear on every map of inverted parameters. N, θi and bspec did not show any clear visible
variation on infected areas. Cab showed lower values on infected areas, while Cbp and, to a lesser extent, Ccx values
tended to increase on infected areas. Cant increased until stage 4, to nally decrease with stages 5 and 6. Low val-
ues (less than 0.015) obtained for reconstruction errors indicate that the modelled spectra were consistent with
measured ones. Although stage 6 showed a slight increase in reconstruction error on circumscribed areas, the
precision of the inversion of PROCOSINE was not notably aected by the disease. Finally, the error spectra also
showed low values (Supplementary Fig.1), although patterns were observed for dierent spectral regions. e
blue region (400 to 500 nm) of the spectrum showed a relatively high standard deviation. e shape of the error
spectra in the range of about 500–650 nm changed from oscillating at high frequency in stage 0 to lower frequency
and smoother oscillation pattern in the later stages. Around the red edge, the error increased between stage 0 and
stage 3 and decreased again.
Boxplots and distributions of the complete pixel dataset were computed to monitor the dynamic of each
PROCOSINE parameter according to the disease stage (Fig.3).
e mesophyll structure parameter (N) dynamic showed a slight increase as the disease spreads, ranging
from a median value of 1.52 at stage 0 (asymptomatic stage) to 1.82 at stage 5. An increase of the dispersion
Figure 1. Mean reectance spectra (solid lines) and standard deviation (dashed lines) of asymptomatic and
BLSD-infected leaf tissues at dierent stages of the disease (stage 0 to 6). Grey spectra show the reectance
values for the previous disease stage.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
4
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
of the values for stages 3, 5 and 6 is also evident. Cab showed a decrease starting between stages 3 and 4, and
median values ranged from 43.90 µg.cm−2 for asymptomatic leaves, to 0.69 µg.cm−2 for stage 6. Ccx increased
between stages 0 to 3, decreased between stages 3 and 5, and nally stabilised for stages 5 and 6. Median values
ranged between 9.06 and 24.15 µg.cm−2, and a strong dispersion was found for stages 3, 4, 5 and 6. Cant displayed
a bell-shaped pattern, i.e., it increased between asymptomatic stage and stage 3, before decreasing from stage 4
to stage 6. Median values ranged from 1.70 to 9.80 µg.cm−2, with a strong dispersion for all stages, excepted for
asymptomatic stage. Cbp displayed low values from asymptomatic stage to stage 3, then dramatically increased at
stage 4, with median values ranging from 0.01 to 4.99. Dispersion was also very pronounced for stages 4, 5 and 6.
θi displayed an increase aer stage 4. In particular, the values for stage 6 were higher than those obtained for other
stages (median values ranging from 24 to 44°), with a strong dispersion of θi values for all disease stages. Finally,
there was no clear dynamic for the specular term bspec, although stage 6 exhibited slightly higher values and an
increased dispersion. e median bspec values ranged between −0.01 and 0.00.
Discrimination of disease stages. A two-fold cross validation linear discriminant analysis (hereinaer
referred to as LDA) was used to classify the dierent disease stages based on the results of PROCOSINE’s inver-
sion. e overall accuracy of the model was then computed for a variable number of discriminant axes. Figure4a
shows that the rst two axes provided optimal performances, with 78.7% of correct classications, while avoid-
ing overtting. e confusion matrix associated with these classication results provides more insights into the
classier performance (Table1). e best performance was achieved for the discrimination of stage 5 (producer
accuracy: 95.7%). e identication of stages 2 and 4 pixels was more dicult (producer accuracies of 61.8 and
50.3%, respectively) due to confusions with stage 0 and 3, respectively.
A new LDA model was then calibrated using the complete dataset to compute the weighting coecients of the
parameters used to build the discriminant vectors. With a weighting coecient value three-times that of the other
ones on the rst discriminant axis, Cab was the main parameter for the classication of the disease stages. Cbp and
Figure 2. Examples of inverted PROCOSINE parameter maps. “RGB” are visualizations of the hyperspectral
images of the adaxial side of the leaf disk. N and bspec are unitless, Cab, Ccx,Cant are expressed in μg.cm−2, Cbp
is given in arbitrary units and θi is expressed in degrees. “Error” indicates the reconstruction error maps,
expressed in reectance factor (unitless). e scale next to the RGB images indicates the size of the leaf disks.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
5
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
Cant also appeared to carry critical information, with coecients higher than 3 in absolute value for the second
axis (Fig.4b). Figure5 shows that the rst axis mainly discriminated stages 3, 4, 5 and 6, with some overlapping
between stages 3 and 4. e second axis discriminated the asymptomatic stage from stages 2 and 3, again with
some overlapping between stages 0 and 2.
Discussion
Pixels belonging to stage 0 show reectance features characteristic of healthy green vegetation18,30,31. We observed
a global decrease of reflectance in the VIS and NIR spectral domains as the disease spreads, except for the
650–700 nm range (Fig.1). e results obtained from model inversion (Figs2 and 3) suggest that the associated
decrease in reectance may result from an increase in Ccx and Cant, as observed in various situations of stress32,33.
An increase in Ccx and Cant is physically sound to explain the decrease of reectance for the 500–650 nm range,
as it shows unsaturating reectance for the asymptomatic stage. However, considering the saturating reec-
tance (i.e., that does not change with a further increase in absorption) for the blue (400–500 nm) range, the
observed decrease is more likely to be due to changes in surface properties unaccounted for by the model. e
slight increase of reectance in the NIR domain for stage 3 can be explained by disorganisation of the structure
Figure 3. Boxplots (outliers, minimum, rst quartile, median, third quartile and maximum) and distribution of
estimated PROCOSINE parameters according to the stage of black leaf streak disease. e width of the boxplots
indicates the number of samples for each group. is number ranges between 445 (stage 3) and 1504 (stage 0).
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
6
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
of the mesophyll. e strong decrease of reectance measured in the NIR domain that starts from stage 4 can be
explained by the increasing amount of brown pigments or other denatured proteins absorbing the light below
1300 nm34,35, as conrmed by the results of model inversion.
Figure 4. Results of the linear discriminant function analysis (a: overall classication accuracy obtained
using two-fold cross-validation as a function of the number of discriminant axes; b: weighting coecients of
PROCOSINE parameters obtained for the rst two discriminant axes and computed over the complete data set).
Expert knowledge User
accuracy0 2 3 4 5 6
LDA results
0 1218 168 15 0 0 0 86.9
2 286 381 89 28 0 0 48.6
3 0 65 307 172 0 0 56.4
4 0 3 34 504 26 0 88.9
5 0 0 0 299 1344 45 79.6
6 0 0 0 0 34 923 96.5
Producer accuracy 81.0 61.8 69.0 50.2 95.7 95.4 78.7
Table 1. Confusion matrix computed from the LDA classication results obtained using two-fold cross-
validation.
Figure 5. Sample distribution in the LDA plane dened by the rst two linear discriminant vectors computed
on the complete data set.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
7
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
e increase of reectance in the 650–700 nm range is explained by a decrease in Cab which leads to decreased
absorption. For stages 4, 5 and 6, the decrease in absorption due to the reduction of Cab is compensated by the
increased amount of brown pigments that also absorb light in this spectral domain. N shows a slight increase as
the disease spreads. is parameter describes the number of theoretical homogeneous layers composing the mes-
ophyll, and can be considered a measurement of the optical complexity of the inner leaf medium. e increase
of N can be linked to the action of the fungus on the leaf. Indeed, P. jiensis is an hemibiotrophic fungus36, with
rst a biotrophic phase which only colonizes the intercellular spaces between mesophyll cells, hence initially
having little inuence on the global structure of the mesophyll. However, this rst step is followed by growth in
all intercellular mesophyll layers and air chambers, then by growth of conidiophores through the stomata and
then the rst alterations of mesophyll cells are observed (at the beginning of the necrotrophic phase), which may
explain the slight disorganisation of the vegetative medium37. e response of Cab is also consistent with expec-
tations, as the development of P. jiensis induces chlorosis36. Cab starts decreasing beyond stage 3, suggesting that
Cab is not signicantly aected at the rst disease stages. Ccx tends to increase at the rst stages of disease, more
particularly between stage 2 and 3 (Fig.3). Many studies have reported an increase of the carotenoids content as
a response to various environmental stresses33,38–40. However, to the best of our knowledge, there have been no
reports of Ccx increases induced by pathogen-linked specic stresses. e dynamic of Ccx can also be explained by
other factors unaccounted for by PROCOSINE, including biochemical constituents produced by the fungus or
the immune system of the plant, or changes in leaf surface properties or internal structure. As illustrated in Fig.3,
Cant increases from stage 0 to 3, before decreasing from stage 4 to 6. e increase in Cant observed during the early
stages of infection is also visible in Fig.2. Interestingly, these results are in agreement with previous studies that
reported an increase in Cant as a response to dierent types of stresses32,41, including the appearance of pathogens
in vegetative tissues42 using direct biochemical measurements. e increase in Cbp is consistent with expectations,
as brown pigments appear during senescence and diseases, with complexing of oxidized phenols43,44. Brown pig-
ments are absent in healthy leaves or during the rst stages of the infection. e slight increase observed for θi
might be explained both by the disorganisation of the medium while the fungus grows and spreads, and by the
progressive evaporation of the cell-contained water, leading eventually to the collapse of the cell. is might also
come from compensations between Cbp and θi as observed for necrotic tissues by Jay et al.23. e dynamic of the
specular reection bspec shows no particular trends, which suggests that this parameter is not aected by the devel-
opment of P. jiensis, or that the current parameterisation of the model or the spatial resolution used in this study
does not take into account potential variations in bspec.
e dierences between simulated and measured spectra (supplementary Fig.1) can be explained by dierent
factors according to the spectral range: in the blue domain, the error probably comes from the inherent noise of
the sensor at such wavelengths. e variations with high frequencies observed for the mean error in the 500–
750 nm range may be explained by a slight shi of the spectral calibration of the sensor. Indeed, the dierences
between measured and simulated spectra tend to become more important when occurring in spectral regions
with rapid changes of reectance, such as for the green or red edge regions. e concomitant reduction of those
dierences when reaching late stages of the disease would support this hypothesis, with the observed slopes of the
green and red edge regions decreasing. Finally, the variations with low frequencies observed in the visible and the
NIR regions for stage 3 and later stages might come from the fact that the pathogen is pushing the model toward
the limits of its domain of validity, especially with the potential appearance of new chemical compounds that are
not taken into account by PROSPECT.
e discrimination of stages 3, 4, 5 and 6 is based on the rst linear discriminant axis (Fig.5), with Cab as the
most important variable for this axis (Fig.4b). e high accuracy achieved for the classication of stages 5 and 6
is consistent with the results presented in Fig.2 and further illustrated in Fig.6, as clear dierences in chlorophyll
content appear between stages 4, 5 and 6. However, confusion occurs in discriminating between stages 3 and 4.
is is probably due to the fact that stages 3 and 4 are dierentiated by the geometry of the disease spot, which is
not taken into account here.
As mentioned earlier, the LDA results show that the rst stages of the disease (i.e., stages 2 and 3) are di-
cult to discriminate. More particularly, confusion occurs between pixels related to stage 0 and stage 2. Figure5
shows that the separation of stages 0 and 2 is mainly explained by the second discriminant vector. Variables that
contribute the most to this axis are Cab, Cant and Cbp (Fig.4b). Among those parameters, only Cant shows dierent
values between stages 0 and 2 (Fig.6), despite some overlap of the two distributions (Fig.3) that might explain the
poorer accuracy for discriminating stages 0 and 2. e confusion between stage 0 and 2 can also be explained by
the size and spatial distribution of early stages spots (i.e., stage 2) that can lead to the inclusion of asymptomatic
pixels into the stage 2 subset (see Supplementary Fig.2). However, with an overall accuracy of 78.7%, the use of a
LDA model applied to the inversion results of PROCOSINE is a relevant approach for discriminating the disease
stages of P. jiensis-infected banana leaves. Moreover, the increase in Cant during the rst stages of the spreading
of the fungus makes it a promising proxy for early disease detection.
PROSPECT is a physically-based model describing the interactions between light and leaf tissues in order
to simulate leaf optical properties resulting from foliar biochemical and structural properties. However, models
are based on the simplication of complex systems. In the case of PROSPECT, one of the main advantages of
the model is its straightforwardness: it does not seek to exhaustively describe the complexity of the leaf, and
simplications such as its generic refractive index and the simplistic description of the leaf/air interface are well
known22,45. When considering the physiological impact of foliar diseases on leaf tissues, it is possible that the
unusual plant responses to the disease development may be beyond the original domain of validity of PROSPECT,
and lead to “exotic combinations” of parameters. For example, PROSPECT-D only includes Cab, Ccx, Cant and Cbp
to describe the absorptive eects of leaf constituents in the optical domain, and the presence of secondary pig-
ments (e.g., avonoids) or other optically-active chemical constituents that might be induced by the development
of the fungus are not considered. Moreover, the three families of pigments accounted for by PROSPECT (i.e.,
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
8
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
chlorophylls, carotenoids and anthocyanins) also rely on assumptions. ese pigment families encompass a large
variety of dierent molecules, and the stoichiometry of these molecules is supposed to be identical among veg-
etation, as they are dened by a unique specic absorption coecient. erefore, changes in specic absorption
of anthocyanins induced by the variations of pH are not taken into account46. Pathogens may induce changes in
the pH of part or all of leaf cells, resulting in changes in the optical activity of anthocyanins, which, in our case,
can be interpreted as a variation in content of Cant. Finally, PROSPECT is a plate model, describing the leaf as a
stack of several homogeneous layers, described by the N parameter. e fungus, although supposed to have little
impact on the inner structure of the medium, will introduce some heterogeneity while penetrating the inter-cell
space and activating its defences, which may push the model toward its limits from a physical point of view.
However, the error spectra remain low (Supplementary Fig.1), and the reconstruction error (Fig.2) do not show
any clear pattern depending on the disease stage. Consequently, values derived from PROSPECT inversion can
be considered as indicators of biophysical and chemical changes in leaf tissues. However, those values cannot be
used as absolute values for the leaf properties included in PROSPECT, as long as no proper validation has been
performed. Indeed, working at sub-millimetric scales is a prerequisite for the early detection of diseases47. Due to
the size of the investigated areas, the experimental setup required to perform biochemical analysis on destructive
samples that could be used as absolute values go beyond our exploratory study. Despite these limitations, the
results shown in previous sections demonstrate that PROCOSINE (i) provides dynamics that are consistent with
physiological expectations and (ii) is useful to investigate the eects of P. jiensis and, more generally, of plant
diseases on leaf tissues.
e exploratory results underline that the ecophysiological processes inferred from PROCOSINE inversion
can help understand how banana leaves are aected by black leaf streak disease. Indeed, the rst stages show no to
little variations in brown pigments and chlorophyll contents, the latter being a key variable for most of the current
vegetation indices. Consequently, this suggests that the classical approaches based on such vegetation indices may
be not suciently relevant for the early detection of black leaf streak disease. Instead, particular emphasis should
be paid to the use of anthocyanin-sensitive indices such as those proposed by Gitelson et al.48. One potential per-
spective would be to try to detect the BLSD at the eld level using a drone-mounted multispectral camera with the
relevant spectral bands to estimate the anthocyanins content. Another interesting prospect would be to acquire
images in the SWIR range, as this would enable inclusion of the leaf dry matter and water contents, which can be
interesting indicators of the development of the fungus.
More generally, the inversion of radiative transfer models such as PROCOSINE appears to be an interesting
tool to investigate the eects of diseases on foliar systems from a physiological point of view. Such an approach
could guide the selection of appropriate vegetation indices, e.g., for the early detection of plant diseases or to dis-
criminate disease-related stresses from water, nitrogen or other biotic stresses. e inversion of radiative transfer
models for detecting and discriminating disease stages should be further assessed with a larger dataset, including
various cultivars and image acquisition conditions. is would allow a rigorous comparison between a purely
statistical method (e.g., LDA) and the hybrid method used in this study (PROCOSINE and LDA), and also to
conrm the current performance with an independent dataset. One issue to assess is that, depending on the dis-
ease’s type and stage, the inversion performance may be limited by the fact that the biochemical and biophysical
conditions are out of the boundaries for which the model was validated (e.g., with a strong necrosis of the leaf).
Further investigation should be done in order to provide proper validation and conrm or rene our analysis in
the case of the inuence of the pathogen P. jiensis on banana leaves of dierent ages, from dierent accessions
and with dierent level of resistance, and new tests with larger datasets should be performed to assess the robust-
ness of the method. e use of PROCOSINE based on high spatial resolution images of plant canopies would also
Figure 6. Interactions between P. jiensis and the most sensitive PROCOSINE parameters according to the
disease stage. Cab and Cant are expressed in μg.cm−2, Cbp is given in arbitrary units.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
9
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
be an interesting prospect to explore, although the model has not been designed to properly handle the radiative
transfer conditions that prevail in such situations, i.e., with a strong inuence of the pixel environment.
Conclusion
e coupled PROSPECT-D + COSINE model (PROCOSINE) was used to study the response of structure and
chemistry of banana leaves when infected with black leaf streak disease (BLSD) caused by the fungus P. jiensis.
For this purpose, asymptomatic and BLSD-infected pixels were rst extracted from laboratory acquired hyper-
spectral images and were given a disease stage based on expert knowledge. Foliar chemistry and structure were
then retrieved based on PROCOSINE inversion for each pixel. Finally, a linear discriminant analysis was per-
formed to estimate the disease stage of the pixels of our database based on their inverted PROCOSINE parameter
values.
Our results show that most of the dynamics of estimated PROCOSINE parameters are consistent with expec-
tations. Cab showed a progressive decrease, starting at stage 4, while Cbp and, to a lesser extent, N and θi tended
to increase with the spread of the disease between stages 3 and 4. Ccx and Cant showed an interesting dynamic,
increasing between stage 0 and 3, then decreasing. The dynamic of bspec did not show any specific pattern.
Interestingly, Cab, Cant and Cbp appear to be particularly sensitive to the development of the fungus. ese dynam-
ics can be related to the action of the fungus on the leaf, for example the induced progressive chlorosis and brown
pigment appearance, or the increase in anthocyanins as a response to the spreading of the disease. Furthermore,
the classication results show a good overall accuracy 78.7%). e latest stages are the easiest ones to be retrieved,
while more confusion occurs between stages 0, 2 and 4. Of particular interest is the anthocyanins content, which
appears to be particularly sensitive to the rst infection stages. Further investigation on the detection of the very
early infection phase (i.e. stage 1) should be performed using intermediate values of Cant and Cab to those obtained
for stages 0 and 2.
Methods
Plant material and reference measurements of disease stage. Five symptomatic and asymptomatic
banana leaves (Musa AAA cv. Williams, Cavendish subgroup) were used in this study. All leaves were contam-
inated by natural inoculation and harvested aer the appearance of symptoms of BLSD caused by the fungus
Pseudocercospora jiensis. Observed symptoms ranged from stage 2, corresponding generally to a brown strip on
the adaxial surface of the leaf, to stage 6 (necrotic leaf tissue) with grey dry spots surrounded by a black ring as
described by Fouré49. No stage 1 BLSD spots were used in this study due to the diculty for experts to identify
it, as it is only visible on the abaxial side of the leaf49. A total of twelve disks corresponding to various stages of
infection and possibly including several disease stages (Fig.2), were cut o at one time from leaves using a cork
borer (22-mm diameter) and imaged the same day. Two disks obtained from dierent leaves were used for each
disease stage (including the asymptomatic stage), and based on expert identication, one area of a specic stage
was selected on each disk (Supplementary Fig.2). In total, the twelve areas comprised 5941 pixels.
Reectance measurements. e adaxial side of leaf disks were imaged in horizontal position using a
HySpex VNIR-1600 hyperspectral camera (Norsk Elektro Optikk, Norway). is camera was used to sample
the reected radiation in 160 bands ranging from 415 to 994 nm, with a 3.7 nm spectral sampling interval and
a 4.5 nm spectral resolution. e camera was placed at 30 cm above the samples and acquired successive scans
of 1600 pixels, with a resulting spatial resolution of approximately 0.07 mm. A halogen lamp equipped with a
converging lens in order to focus the light beam on the viewing zone of the sensor was used to illuminate the
sample with an incident angle of 40° relatively to the vertical direction. A 50% diuse reectance reference panel
(Spectralon®, Labsphere) was placed horizontally at the same distance from the sensor as the leaves to estimate
the incoming irradiance while limiting possible saturation of the sensor. For each pixel and each band, the bidi-
rectional reectance factor (BRF)50 was computed by dividing the signal measured over the target by that meas-
ured over the reference panel and multiplying by the reectance value provided by Labsphere (assuming the panel
to be Lambertian). Eventually, the image background was removed by image processing in order to keep only
vegetation-related pixels for further analysis.
Model inversion. e coupled models PROSPECT-D22 and COSINE23 were used to estimate the structure
and chemistry of the leaves. PROSPECT is a generalisation of the “plate model” proposed by Allen et al.51, and
simulates the leaf directional-hemispherical reectance and transmittance50 in the optical domain ranging from
400 to 2500 nm, based on a limited number of input parameters20. In this study we used PROSPECT-D22, which
is the latest release of the model. is version includes a structure parameter linked to the foliar anatomy, usually
named N, that corresponds to the number of homogeneous layers stacked in order to represent the leaf, separated
by N−1 air spaces. N is unitless and assumed to vary between 1 and 3 for the majority of non-senescent leaves20.
Low N values (close to 1) correspond to leaves with a relatively simple anatomy such as monocotyledons or juve-
niles, while higher N values correspond to more complex leaf structures characteristic of dicotyledons, including
dierentiated parenchyma and intercellular gaps that increase multiple scattering of the light within the leaf.
PROSPECT-D also simulates the eects of the three main families of foliar pigments on leaf optical properties,
i.e., chlorophylls (Cab), carotenoids (Ccx) and anthocyanins (Cant), whose contents are expressed in μg.cm−2. ese
constituents are optically-active in the visible (VIS) domain and are expected to have a particularly strong inu-
ence on our reectance measurements. Brown pigments content (Cbp, arbitrary unit per surface unit) is another
input parameter of PROSPECT-D. It corresponds to the complexing of oxidized phenols with proteins during
senescent or disease stages of the leaf, although their synthesis is not fully claried44,52. Due to the optical activity
of brown pigments in the VNIR spectral domain, we also expect this constituent to contribute to the banana leaf
reectance at the latest stages of the disease. Finally, the two remaining input parameters of PROSPECT-D are the
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
10
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
equivalent water thickness (Cw), expressed in cm, and the dry matter content (Cm), expressed in g.cm−2. e abil-
ity of PROSPECT-D to simulate leaf optical properties in direct mode, and to accurately estimate leaf chemistry
in inverse mode, has been validated for a wide range of species and phenology, from juvenile to senescent stages22.
Close-range imaging spectroscopy allows the retrieval of the pseudo bidirectional reectance factor23, which
means that PROSPECT cannot be used directly to simulate such spectral data and to retrieve leaf biochemis-
try based on model inversion. Indeed, the anisotropic eects induced by leaf surface properties (including the
presence of waxes and trichomes) do not allow considering the leaf as an isotropic surface, and require taking
into account leaf orientation and specularly-reected radiation53. To overcome this limitation, Jay et al.23 devel-
oped the COSINE (ClOse-range Spectral ImagiNg of lEaves) model that estimates the pseudo BRF from the
directional-hemispherical reectance factor simulated by PROSPECT. e coupling of PROSPECT and COSINE
(hereinaer referred to as PROCOSINE) thus relates the leaf pseudo BRF to the leaf structure and biochemis-
try. COSINE requires two specic parameters to compute the pseudo BRF from the reectance simulated by
PROSPECT: the light incident angle (θi), i.e., the angle between the light source and the normal to the leaf,
expressed in degrees, and the specular term (bspec) which describes the amount of specularly-reected radiation
(arbitrary unit).
PROCOSINE can be used either in forward or in inverse mode. When used in inverse mode, part or all of the
input parameters are inferred from the measured reectance factor. Various methods can be used in order to infer
these input parameters, including multivariate statistical analysis, machine learning, look-up tables or iterative
optimization54–56. In our study, iterative optimization was used in order to nd the optimal combination of input
parameters Θ that minimizes the merit function J. e latter relates the modelled and measured reectance fac-
tors and is here given by the spectral mean squared error dened as
∑
ΘΘ=−
λ
λλ
=
λ
JRR() (())
(1)
n
meas mod
1
,,
2
where nλ is the number of spectral bands, Rmeas,λ and Rmod,λ are, respectively, the measured and modelled reec-
tance factors at waveband λ, and Θ is the [9 × 1] vector of PROCOSINE parameters to be estimated. Among
all the numerical methods available to perform iterative optimization and to estimate leaf properties from
PROCOSINE inversion, we selected the trust-region reective algorithm implemented within the lsqcurvet
function of MATLAB (version 8.6.0, e MathWorks Inc., Natick, MA, 2015), and applied it for each leaf pixel
of the hyperspectral images. An iterative optimization was performed over the 415–900 nm spectral range, aer
removing the longer wavelengths because of the high noise level in this domain (using these wavebands for model
inversion actually did not improve the results, not shown). For each model input parameter to be estimated, the
initial value as well as the lower and upper bounds are given in Table2. Finally, the reconstruction error was com-
puted as the root mean squared error between measured and modelled spectra:
=∑−
λλλ
=
λ
RMSE RR
n
()
(2)
nmeas mod
1,,
2
where n is the number of samples, nλ is the number of spectral bands, Rmeas,λ is the measured reectance factor
and Rmod,λ the modelled reectance factor at the λth waveband.
Feature extraction and classication. For each leaf disk, areas of interest were delimited over asympto-
matic zones and relevant BLSD spots using the MATLAB soware (version 8.6.0, e MathWorks Inc., Natick,
MA, 2015) and true colour visualisation of the hyperspectral images. For each of the 5941 pixel included in these
areas of interest, the parameter values estimated from model inversion were extracted and associated with their
disease stage. Note that Cw and Cm were not used here, as their inuence on leaf reectance in the VNIR range is
very limited: the resulting Cw and Cm estimates thus provide little relevant information to discriminate the disease
stages.
e lda multiclass Linear Discriminant Analysis (LDA) function from the MASS package57 of the R soware
(version 3.3.3, R Foundation for Statistical Computing, Vienna, Austria, 2017) was used to estimate the disease
severity stage of our database’s pixels based on their inverted PROCOSINE parameter values. Linear discriminant
analysis is a statistical technique that aims to describe, explain and predict membership of predened groups (dis-
ease stage in our case) of a set of observations based on a series of predictive variables. It consists of maximizing
the ratio of the between-class variance to the within-class variance using at most n−1 linear discriminant vec-
tors, where n is the number of classes to be discriminated. Using these linear discriminant vectors, variables are
projected into a sub-space where their assignment to a given class is determined based on the distance between
Parameter N Cab Ccx Cant Cbp CwCmθibspec
Units — μg.cm−2μg.cm−2μg.cm−2a.u.cm g.cm−2° —
Initial value 1.5 50 10 1 0.5 0.002 0.002 20 0
Lower bound 1 0 0 0 0 0.0005 0.001 0 −0.2
Upper bound 3 100 30 10 5 0.1 0.02 90 0.6
Table 2. Initial values, lower and upper bounds used for PROCOSINE inversion. a.u.: arbitrary unit. —:
unitless parameter.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
11
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
the projected individual and the class centroids. Using the coecients of the linear discriminant vectors, one
can physically interpret the results of the LDA (in our case, the ecophysiological response of the banana leaf to
the spread of the disease). Before performing the LDA, the explanatory variables (i.e., inverted parameters) were
normalized using the following equation:
=
−
−
VN
Vmin
maxmin (3)
iiV
VV
where, for a given explanatory variable V, VNi is the normalized value of V for the ith pixel, Vi the value of V for
the ith pixel, and minV and maxV are, respectively, the minimum and maximum values of V over all the 5941 pixels
available in the database. Due to the structure and the size of our database, it was not possible to perform a ran-
dom sampling to create the training and validation subsets, as pixels from the same disk would have been highly
correlated. Consequently, the database was split into two independent subsets S1 (pixels belonging to the rst disk
of a given disease stage) and S2 (pixels belonging to the second disk of a given disease stage) of similar size (3294
pixels for S1 and 2647 for S2). We used a two-fold cross validation to compute the classication performances of
the LDA for a variable number of discriminant axes, with a rst model calibrated on S1 and validated on S2, and
a second model calibrated on S2 and validated on S1.
e classication accuracy was assessed with producer accuracy, user accuracy and overall accuracy58. e
producer accuracy (PA) of a given class is calculated as
=+
PA
TP
TP FN() (4)
where TP is the number of true positives for the considered class (each class is a disease stage), i.e., the number
of pixels correctly classied into the considered class, and FN is the number of false negatives, i.e., the number
of pixels belonging to the considered class but misclassied into another one. e user accuracy (UA) of a given
class is given by
=+
UA
TP
TP FP() (5)
where FP is the number of false positives, i.e., the number of pixels incorrectly classied into the considered class.
Finally, the overall classication accuracy is computed as follows:
=
∑
accuracy
TP
n(6)
s
with ns the number of samples in our dataset.
Finally, another LDA model was calibrated using the complete dataset to estimate more accurately the relative
importance of each PROCOSINE parameter on the discriminant vectors.
Data Availability
All data and codes used in this study are available upon request.
References
1. Oere, E.-C. Crop losses to pests. J. Agric. Sci. 144, 31–43 (2006).
2. Martinelli, F. et al. Advanced methods of plant disease detection. A review. Agron. Sustain. Dev. 35, 1–25 (2015).
3. Battude, M. et al. Estimating maize biomass and yield over large areas using high spatial and temporal resolution Sentinel-2 lie
remote sensing data. emote Sens. Environ. 184, 668–681 (2016).
4. Morel, J. et al. Coupling a sugarcane crop model with the remotely sensed time series of fIPA to optimise the yield estimation. Eu r.
J. Agron. 61, 60–68 (2014).
5. ullberg, E. G., DeJonge, . C. & Chávez, J. L. Evaluation of thermal remote sensing indices to estimate crop evapotranspiration
coecients. Spec. Issue Improv. Agric. Water Product. Ensure Food Secur. Chang. Environ. Overseen Brent Cloth. 179, 64–73 (2017).
6. Jay, S. et al. Estimating leaf chlorophyll content in sugar beet canopies using millimeter- to centimeter-scale reectance imagery.
emote Sens. Environ. 198, 173–186 (2017).
7. Schlemmer, M. et al. emote estimation of nitrogen and chlorophyll contents in maize at leaf and canopy levels. Int. J. Appl. Earth
Obs. Geoinformation 25, 47–54 (2013).
8. Zhou, X. et al. Assessment of leaf carotenoids content with a new carotenoid index: Development and validation on experimental
and model data. Int. J. Appl. Earth Obs. Geoinformation 57, 24–35 (2017).
9. Boc, C. H., Poole, G. H., Parer, P. E. & Gottwald, T. . Plant Disease Severity Estimated Visually, by Digital Photography and
Image Analysis, and by Hyperspectral Imaging. Crit. ev. Plant Sci. 29, 59–107 (2010).
10. Mahlein, A.-. Plant Disease Detection by Imaging Sensors – Parallels and Specic Demands for Precision Agriculture and Plant
Phenotyping. Plant Dis. 100, 241–251 (2016).
11. Sanaran, S., Mishra, A., Ehsani, . & Davis, C. A review of advanced techniques for detecting plant diseases. Comput. Electron.
Agric. 72, 1–13 (2010).
12. Gennaro, S. F. D. et al. Unmanned Aerial Vehicle (UAV)-based remote sensing to monitor grapevine leaf stripe disease within a
vineyard aected by esca complex. Phytopathol. Mediterr., https://doi.org/10.14601/Phytopathol_Mediterr-18312 (2016).
13. Lu, J., Ehsani, ., Shi, Y., de Castro, A. I. & Wang, S. Detection of multi-tomato leaf diseases (late blight, target and bacterial spots)
in dierent stages by using a spectral-based sensor. Sci. ep. 8 (2018).
14. Mahlein, A.-. et al. Development of spectral indices for detecting and identifying plant diseases. emote Sens. Environ. 128, 21–30
(2013).
15. Delalieux, S., van Aardt, J., eulemans, W., Schrevens, E. & Coppin, P. Detection of biotic stress (Venturia inaequalis) in apple trees
using hyperspectral data: Non-parametric statistical approaches and physiological implications. Eur. J. Agron. 27, 130–143 (2007).
16. Yeh, Y.-H. et al. Strawberry foliar anthracnose assessment by hyperspectral imaging. Comput. Electron. Agric. 122, 1–9 (2016).
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
12
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
17. Calderón, ., Montes-Borrego, M., Landa, B. B., Navas-Cortés, J. A. & Zarco-Tejada, P. J. Detection of downy mildew of opium
poppy using high-resolution multi-spectral and thermal imagery acquired with an unmanned aerial vehicle. Precis. Agric. 15,
639–661 (2014).
18. Jacquemoud, S. et al. POSPECT + SAIL models: A review of use for vegetation characterization. emote Sens. Environ. 113,
S56–S66 (2009).
19. Verhoef, W. Light scattering by leaf layers with application to canopy reectance modeling: e SAIL model. emote Sens. Environ.
16, 125–141 (1984).
20. Jacquemoud, S. & Baret, F. POSPECT: A model of leaf optical properties spectra. emote Sens. Environ. 34, 75–91 (1990).
21. Albetis, J. et al. Detection of Flavescence dorée Grapevine Disease Using Unmanned Aerial Vehicle (UAV) Multispectral Imagery.
emote Sens. 9, 308 (2017).
22. Féret, J.-B., Gitelson, A. A., Noble, S. D. & Jacquemoud, S. POSPECT-D: Towards modeling leaf optical properties through a
complete lifecycle. emote Sens. Environ. 193, 204–215 (2017).
23. Jay, S., Bendoula, ., Hadoux, X., Féret, J.-B. & Gorretta, N. A physically-based model for retrieving foliar biochemistry and leaf
orientation using close-range imaging spectroscopy. emote Sens. Environ. 177, 220–236 (2016).
24. de Lapeyre de Bellaire, L., Fouré, E., Abadie, C. & Carlier, J. Blac Leaf Strea Disease is challenging the banana industry. Fruits 65,
327–342 (2010).
25. Leach, . A new form of Banana leaf spot in Fiji. Blac leaf strea. Wor ld Cro ps 16, 60–64 (1964).
26. Stover, . Somaclonal variation in Grand Naine and Saba bananas in the nursery and eld. In Banana and plantain breedi ng strategies
21, 136–139 (ACIA Proc. 21, ppl36-139, 1987).
27. Castelan, F. P. et al. elation between the severity of Sigatoa disease and banana quality characterized by pomological traits and
fruit green life. Crop Prot. 50, 61–65 (2013).
28. C hillet, M. & de Lapeyre de Bellaire, L. Elaboration de la qualité des bananes au champ. Détermination des critères de mesure (1995).
29. Chi llet, M., Abadie, C., Hubert, O., Chilin-Charles, Y. & de Lapeyre de Bellaire, L. Sigatoa disease reduces the greenlife of bananas.
Crop Prot. 28, 41–45 (2009).
30. Ustin, S. L. et al. etrieval of foliar information about plant pigment systems from high resolution spectroscopy. emote Sens.
Environ. 113, S67–S77 (2009).
31. Wu, C., Niu, Z., Tang, Q. & Huang, W. Estimating chlorophyll content from hyperspectral vegetation indices: Modeling and
validation. Agric. For. Meteorol. 148, 1230–1241 (2008).
32. Gould, . S. Nature’s Swiss Army nife: e Diverse Protective oles of Anthocyanins in Leaves. J. Biomed. Biotechnol. 2004,
314–320 (2004).
33. Solovcheno, A. Photoprotection in Plants. 14, (Springer Berlin Heidelberg, 2010).
34. Baret, F. & Fourty, T. Estimation of leaf water content and specic leaf weight from reectance and transmittance measurements.
Agronomie 17, 455–464 (1997).
35. Peñuelas, J. & Filella, I. Visible and near-infrared reectance techniques for diagnosing plant physiological status. Trends Plant Sci.
3, 151–156 (1998).
36. Harelimana, G., Lepoivre, P., Jijali, H. & Mourichon, X. Use of Mycosphaerella jiensis toxins for the selection of banana cultivars
resistant to Blac Leaf Strea. Euphytica 96, 125–128 (1997).
37. Churchill, A. C. L. Mycosphaerella jiensis, the blac leaf strea pathogen of banana: progress towards understanding pathogen
biology and detection, disease development, and the challenges of control. Mol. Plant Pathol. 12, 307–328 (2011).
38. Baricman, T. C., opsell, D. A. & Sams, C. E. Abscisic acid increases carotenoid and chlorophyll concentrations in leaves and fruit
of two tomato genotypes. J. Am. Soc. Hortic. Sci. 139, 261–266 (2014).
39. Havaux, M. Carotenoid oxidation products as stress signals in plants. Plant J. 79, 597–606 (2014).
40. Nisar, N., Li, L., Lu, S., hin, N. C. & Pogson, B. J. Carotenoid Metabolism in Plants. Mol. Plant 8, 68–82 (2015).
41. Chaler-Scott, L. Environmental Signicance of Anthocyanins in Plant Stress esponses. Photochem. Photobiol. 70, 1–9 (1999).
42. Heim, D. Etiolated Maize Mesocotyls: A Tool for Investigating Disease Interactions. Phytopathology 73, 424 (1983).
43. Claussen, . A. & Pepper, E. H. An examination of the brown pigments from barley leaves. Cereal Chem. 45, 124–132 (1968).
44. Spurr, H. W. Brown-Pigment Formation in Tobacco Leaves Infected with Alternaria. Phytopathology 64, 738 (1974).
45. Féret, J.-B. et al. POSPECT-4 and 5: Advances in the leaf optical properties model separating photosynthetic pigments. emote
Sens. Environ. 112, 3030–3043 (2008).
46. Fossen, T., Cabrita, L. & Andersen, O. M. Colour and stability of pure anthocyanins inuenced by pH including the alaline region.
Food Chem. 63, 435–440 (1998).
47. Mahlein, A.-., Steiner, U., Hillnhütter, C., Dehne, H.-W. & Oere, E.-C. Hyperspectral imaging for small-scale analysis of
symptoms caused by dierent sugar beet diseases. Plant Methods 8, 3 (2012).
48. Gitelson, A. A., Merzlya, M. N. & Chivunova, O. B. Optical Properties and Nondestructive Estimation of Anthocyanin Content
in Plant Leaves. Photochem. Photobiol. 74, 38–45 (2001).
49. Fouré, E. Varietal reactions of bananas and plantains to blac leaf strea disease. In Banana and plantain breeding strategies 21,
110–113 (ACIA Proc. 21, ppl36-139, 1987).
50. Schaepman-Strub, G., Schaepman, M. E., Painter, T. H., Dangel, S. & Martonchi, J. V. eectance quantities in optical remote
sensing—denitions and case studies. emote Sens. Environ. 103, 27–42 (2006).
51. Allen, W. A., Gausman, H. W., ichardson, A. J. & omas, J. . Interaction of Isotropic Light with a Compact Plant Leaf. J. Opt. Soc.
Am. 59, 1376 (1969).
52. Vaughn, . C. & Due, S. O. Function of polyphenol oxidase in higher plants. Physiol. Plant. 60, 106–112 (1984).
53. Comar, A., Baret, F., Viénot, F., Yan, L. & de Solan, B. Wheat leaf bidirectional reflectance measurements: Description and
quantication of the volume, specular and hot-spot scattering features. emote Sens. Environ. 121, 26–35 (2012).
54. Darvishzadeh, ., Sidmore, A., Schlerf, M. & Atzberger, C. Inversion of a radiative transfer model for estimating vegetation LAI
and chlorophyll in a heterogeneous grassland. emote Sens. Environ. 112, 2592–2604 (2008).
55. Durbha, S. S., ing, . L. & Younan, N. H. Support vector machines regression for retrieval of leaf area index from multiangle
imaging spectroradiometer. emote Sens. Environ. 107, 348–361 (2007).
56. Fang, H. etrieving leaf area index using a genetic algorithm with a canopy radiative transfer model. emote Sens. Environ. 85,
257–270 (2003).
57. Venables, W. N. & ipley, B. D. Modern Applied Statistics with S. (Springer, 2010).
58. Congalton, . G. A review of assessing the accuracy of classications of remotely sensed data. emote Sens. Environ. 37, 35–46
(1991).
Acknowledgements
is work was funded by the CASDAR PHENOBET project (CASDAR, France), and current results will be
applied to disease detection for sugar beet crops. Jean-Baptiste Féret was funded by the HyperTropik project
(TOSCA program grant of the French Space Agency, CNES). e authors thank Dr David Parsons for proof-
reading the article.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

www.nature.com/scientificreports/
13
SCIENTIFIC REPORtS | (2018) 8:15933 | DOI:10.1038/s41598-018-34429-0
Author Contributions
A.B. provided the leaf material. J.M. and A.B. prepared the experiment. J.M. acquired hyperspectral images and
ran models. J.M., S.J., J.-B.F. R.B., F.C. and N.G. analysed the results. J.M. wrote the manuscript. All authors
reviewed the nal manuscript.
Additional Information
Supplementary information accompanies this paper at https://doi.org/10.1038/s41598-018-34429-0.
Competing Interests: e authors declare no competing interests.
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional aliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Cre-
ative Commons license, and indicate if changes were made. e images or other third party material in this
article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons license and your intended use is not per-
mitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the
copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
© e Author(s) 2018
Content courtesy of Springer Nature, terms of use apply. Rights reserved

1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com