Detection of Defoliation Injury in Peanut with Hyperspectral Proximal Remote Sensing

: Remote sensing can be applied to optimize e ﬃ ciency in pest detection, as an insect sampling tool. This e ﬃ ciency can result in more precise recommendations for decision making in pest management. Pest detection with remote sensing is often feasible because plant biotic stress caused by herbivory triggers a defensive physiological response in plants, which generally results in changes to leaf reﬂectance. Therefore, the key objective of this study was to use hyperspectral proximal remote sensing and gas exchange parameters to characterize peanut leaf responses to herbivory by Stegasta bosqueella (Lepidoptera: Gelechiidae) and Spodoptera cosmioides (Lepidoptera: Noctuidae), two major pests in South American peanut ( Arachis hypogaea ) production. The experiment was conducted in a randomized complete block design with a 2 × 3 factorial scheme (two lepidopterous species and 3 categories of injury). The injury treatments were: (1) natural infestation by third instars of S. bosqueella , (2) natural infestation by third instars of S. cosmioides , and (3) simulation of injury with scissors to mimic larval injury. We veriﬁed that peanut leaf reﬂectance is di ﬀ erent between herbivory by the two larval species, but similar among real and simulated defoliation. Similarly, we observed di ﬀ erences in photosynthetic rate, stomatal conductance, transpiration, and photosynthetic water use e ﬃ ciency only between species but not between real and simulated larval defoliation. Our results provide information that is essential for the development of sampling and economic thresholds of S. bosqueella and S. cosmioides on the peanut. S.P., R.P. O.F.; O.F.; acquisition, D.R. and O.F.


Experimental Design
The experiment was conducted in a randomized complete block design with a 2 × 3 factorial scheme (two lepidopterous species and 3 categories of injury [insect-related injury, simulated injury and control]). The injury treatments were: (1) Natural infestation by third instars of S. bosqueella, (2) Simulation of injury with scissors to mimic S. bosqueella larval injury, (3) Natural infestation by third instars of S. cosmioides, (4) Simulation of injury with scissors to mimic S. cosmioides injury, and (5) Non-Injured (control). Third instars were used because they cause representative and sufficient injuries that allow for measurement and evaluation (unpublished data). Each treatment consisted of ten 5-L pots (replicates) containing a peanut plant 30 days after emergence (DAE).

Test Plants and Insects
Studies were carried out with peanut cultivar IAC 503, which is a major peanut cultivar currently used in Brazil. IAC 503 belongs to the runner vegetative group with a "high oleic" characteristic [31]. For the experiments, peanut plants were grown in individual 5-L pots containing a mixture of soil, sand, and cattle manure in a 1:2:1 proportion. Initially, 5 seeds were placed in each pot, and seedlings were thinned after plant emergence, so that each pot contained only a single plant. The peanut plants were kept in an air-conditioned greenhouse facility (3.50 × 6.35 m in diameter, 25 ± 10 • C, RH 70 ± 10%, and natural light) at São Paulo State University, São Paulo, Brazil. Plants were irrigated daily using the same amount of water per plant.
In Brazil, both S. bosqueella and S. cosmioides prefer to feed on young peanut leaves, therefore, for insect colony establishment, planting was carried out weekly to ensure continuous availability of peanut leaflets. The S. bosqueella population (approximately 150 adults) was collected in a commercial peanut field located in Jaboticabal, São Paulo, Brazil, brought to the laboratory and kept in acrylic cages (30 × 30 × 30 cm) under controlled temperature conditions (12H:12D photoperiod, 25 ± 2 • C, and 60 ± 10% RH), according to Boiça Jr. et al. [32]. Therefore, the cultivar IAC 503 leaflets were used as feeding and oviposition substrates. The initial S. cosmioides colony (approximately 200 specimens) was provided by the Insect Biology Laboratory of the Superior School of Agriculture "Luiz de Queiroz", São Paulo University, Piracicaba, state of São Paulo, Brazil and it was kept in the same controlled temperature conditions as S. bosqueella. Also. similar to S. bosqueella, IAC 503 leaflets were used as feeding and oviposition substrates. For our study, the F2 generation was used.

Injury Treatments
Natural defoliation was performed 30 DAE based on the methodology of Delaney et al. [33]. Leaf cages composed of transparent plastic cylinders (6 cm in height × 4 cm in diameter) were used to restrict larvae to a single freshly opened leaflet and each cage was sealed with cardboard (5 × 5 cm in diameter) (Figure 1e). The same leaf position was adopted for all treatments. All larvae were removed after 24 h, while primary-leaf photosynthetic measurements were recorded 24 and 48 h after herbivores were introduced. The simulated herbivory (Figure 1b,d) was carried out by removing part of the peanut leaflet corresponding to the injury observed in the natural defoliation caused by the larvae. This simulated larval defoliation was performed after the removal of leaf cages with the aid of scissors in the same position as the injury caused by both S. bosqueella and S. cosmioides (Figure 1a,c). In addition, control plants were used for all treatments. All treatments received leaf cages to control for any cage-effects on the experiments.

Hyperspectral Remote Sensing
Hyperspectral remote sensing data were acquired from peanut plants (n = 50 leaflets) 72 h after treatment infestation. The plants were transferred to a climate-controlled room without artificial lighting (to ensure only camera lighting), where the hyperspectral imaging was performed using a benchtop system. Ambient climate conditions were 22-25 • C and 50-60% relative humidity. The acquisition of hyperspectral remote sensing data was performed from 8:00-11:00 a.m. using a hyperspectral camera (PIKA L, Resonon Inc., Bozeman, MT, USA) which acquires reflectance data in the 400-1000 nm spectral range. This camera has a 3.7 nm spectral resolution (Full Width Half Maximum-FWHM) and 281 spectral channels. The calibration surface used as a white reference was made of polyethylene plastic (Type 822, Spectronon Pro, Resonon, Bozeman, MT, USA). The software Spectronon Pro™ was used for camera control, data acquisition, and data processing.
We carefully dissected the portion of the plant where the injury occurred (Figure 1f), and hyperspectral imaging was completed within 1-2 min.
For this, with the aid of Spectronon Pro™ software, we manually collected the average reflectance by drawing a rectangle (Length 1 cm × Width 0.8) and mining the data only in the area on interested, thus ensuring greater homogeneity of data and the extraction of spectral responses. Thus, only the specific herbivory-injured leaflet information was evaluated per plant and used for data analysis.

Data Analysis
For hyperspectral remote sensing data, univariate feature selection based on ANOVA F-value scores was used to create a variable importance plot to highlight the spectral regions useful in distinguishing between injured and non-injured (control) leaves. For this analysis, we used the "sklearn.feature_selection_Select KBest" library in Scikit-learn: Machine Learning in Pythonv3.6. [34].
To test for treatment effects on peanut leaf reflectance, we used the near-infrared (NIR) spectral range (777.42-1000 nm) because it exhibited separation in sample reference spectra ( Figure 2) and is sensitive to variation in plant stressors ( Figure 3). For this, we performed comparative contrast analysis between groups using an F test (p < 0.05) (PROC MIXED) [35].
To evaluate the accuracy of prediction in discerning S. bosqueella, S. cosmioides, and uninfested leaves, a Random Forest Classification was performed using the Scikit-learn: Machine Learning in Pythonv3.6. [34]. Random Forest is one of the more successful ensemble methods for pattern detection in a dataset as it conducts a random selection of all possible explanatory variables (here, hyperspectral bands) at each split [36]. Due to the lack of experimental replicates in our study, the desire to reduce noise, and to respect the principles of parsimony and the Hughes Phenomenon [6,37], we used principal component analysis (PCA) along with Random Forest classification as a dimension reduction tool for all spectral bands (n = 150) collected from the hyperspectral data. The main purpose in using PCA was to summarize all spectral dimensions (features) into three principal components, without losing information. The Random Forest Classification parameters were selected by a standard 5-StratifiedKFold cross-validation with GridSearchCV from Sklearn.model_selection library using Pythonv3. 6. To evaluate the model robustness, we estimated the overall accuracy and the Kappa statistic from our Random Forest Classification model, using 70% of the data for training and 30% for validation [38]. standard deviation (Figure 2, Figure S1). The reflectance profiles for each of the treatment groups appeared to be similar in the visible wavelengths, although spectral deviations occurred above 700 nm in the near-infrared (NIR) region. The spectral profile for control leaves exhibited the lowest values between 777.42-1000 nm, in contrast to an increase in reflectance in the stressed treatment classes. We also observed that the stressed classes based on the actual injury corresponding to each larval species exhibited different spectral responses in the NIR region. However, the simulated injury appeared to be similar to the injury made by the real larvae. Furthermore, the variable importance plot based on ANOVA F-values scores confirmed that the 777.42-1000 nm portion of the spectrum was crucial for spectral distinction among treatment groups ( Figure 3).  However, because one of our main goals was to determine if leaf reflectance was different among larval species and injury categories, we performed contrast analysis with the 700-1000 nm spectral reflectance data. The contrast analysis showed that there was a difference between species' reflectance in the NIR region (777.42-1000 nm) (F5, 3354 = 167.46, p = < 0.0001). However, there was no difference in peanut leaf reflectance between simulated or real injury (F5, 3354 = 0.55, p = 0.4566). Conversely, peanut leaf reflectance changed when we compared injured vs non-injured leaves (F5, 3354 = 356.40, p = < 0.0001) ( Figure 2). To evaluate the effect of different larval species and herbivory-injury methods in peanut leaf gas exchange, we applied the multiple comparison Tukey test (p-value < 0.05) (PROC MIXED) [35]. Similar to the hyperspectral analysis, we applied comparative contrast analysis between groups with an F test (p < 0.05), using the gas exchange parameters as explanatory variables to represent the variation between classes (injury categories). Aiming to understand the peanut leaf gas exchange response to larval-induced herbivory, we developed a classification tree model with only injured and non-injured data as the response variable. We used the Sklearn.tree library to create the classification tree and a standard 5-StratifiedKFold cross-validation with GridSearchCV to define the maximum depth of the tree, the minimum number of samples required to split an internal node and the function to measure the quality of a split. The classification tree analysis, a non-parametric statistical modeling technique, is widely used to identify a set of characteristics (here, gas exchange variables) that best distinguish individuals based on a categorical outcome variable (here, larval-induced-herbivory) [39]. Therefore, we adopted the photosynthetic parameters assimilation/respiration (A-photosynthetic rate), stomatal conductance (GS), transpiration (E), and photosynthetic water use efficiency (WUE) as explanatory (predictor) variables.

Hyperspectral Images
We verified that the average reference spectral signature for the different treatment groups illustrated the spectral regions that were crucial in their distinction and classification with low standard deviation ( Figure 2, Figure S1). The reflectance profiles for each of the treatment groups appeared to be similar in the visible wavelengths, although spectral deviations occurred above 700 nm in the near-infrared (NIR) region. The spectral profile for control leaves exhibited the lowest values between 777.42-1000 nm, in contrast to an increase in reflectance in the stressed treatment classes. We also observed that the stressed classes based on the actual injury corresponding to each larval species exhibited different spectral responses in the NIR region. However, the simulated injury appeared to be similar to the injury made by the real larvae. Furthermore, the variable importance plot based on ANOVA F-values scores confirmed that the 777.42-1000 nm portion of the spectrum was crucial for spectral distinction among treatment groups ( Figure 3).
However, because one of our main goals was to determine if leaf reflectance was different among larval species and injury categories, we performed contrast analysis with the 700-1000 nm spectral reflectance data. The contrast analysis showed that there was a difference between species' reflectance in the NIR region (777.42-1000 nm) (F 5, 3354 = 167.46, p = < 0.0001). However, there was no difference in peanut leaf reflectance between simulated or real injury (F 5, 3354 = 0.55, p = 0.4566). Conversely, peanut leaf reflectance changed when we compared injured vs non-injured leaves (F 5, 3354 = 356.40, p = < 0.0001) (Figure 2).
Our second goal was to build a Random Forest model to discern S. bosqueella, S. cosmioides, and uninfested leaves. Therefore, we performed PCA to condense the multidimensional spectral data into a three-dimensional environment (Figure 4). Principal components 1, 2, and 3 explained most of the variation in the data (95.52%), with PC1 responsible for 74% of the explained variance ( Figure 4). For the Random Forest Classification model, using 70% of the dataset as training data and 30% as validation data, larval injury was classified with an overall accuracy of 92% and a Kappa statistic of 0.83 (Table 1). These results indicate that based on the 9 samples used for validation, our model correctly classified 8 of them.

Gas Exchange Responses to Herbivory Measurements
Neither injury treatment significantly affected peanut leaf electrical conductivity in the primaryleaf photosynthetic measurements recorded 24 h after infestation (Table 2). Also, 24 h after the infestation, the contrast analysis showed no difference in injury caused by S. bosqueella and S. cosmioides and between real herbivory and simulated leaf injury gas exchange responses parameters ( Table 2).
The treatment effects on peanut leaf gas exchange were only observed 48 h after the infestation (Table 3). Herbivory by S. bosqueella caused a low impact on peanut gas exchange and the photosynthetic rates were similar to the control. In contrast, injury by S. cosmioides reduced the photosynthetic rate by 3.97 μmol CO2 m −2 s −1 compared to non-injured leaves. The contrast results showed differences in photosynthetic rate, stomatal conductance, transpiration, and photosynthetic water use efficiency only between species (Table 3). Similar to the hyperspectral image analysis, we verified that real and simulated injuries presented similar effects on peanut leaf gas exchange.

Gas Exchange Responses to Herbivory Measurements
Neither injury treatment significantly affected peanut leaf electrical conductivity in the primary-leaf photosynthetic measurements recorded 24 h after infestation (Table 2). Also, 24 h after the infestation, the contrast analysis showed no difference in injury caused by S. bosqueella and S. cosmioides and between real herbivory and simulated leaf injury gas exchange responses parameters ( Table 2).
The treatment effects on peanut leaf gas exchange were only observed 48 h after the infestation (Table 3). Herbivory by S. bosqueella caused a low impact on peanut gas exchange and the photosynthetic rates were similar to the control. In contrast, injury by S. cosmioides reduced the photosynthetic rate by 3.97 µmol CO 2 m −2 s −1 compared to non-injured leaves. The contrast results showed differences in photosynthetic rate, stomatal conductance, transpiration, and photosynthetic water use efficiency only between species (Table 3). Similar to the hyperspectral image analysis, we verified that real and simulated injuries presented similar effects on peanut leaf gas exchange.    The classification tree showed that injury by S. bosqueella presented high stomatal conductance and similar photosynthetic rate and water use efficiency when compared to non-injured leaflets ( Figure 5). Alternatively, when infested with S. cosmiodes, peanut leaflets showed decreased photosynthetic rate and water-use efficiencies and high stomatal conductance compared to S. bosqueella and the control. The classification tree showed that injury by S. bosqueella presented high stomatal conductance and similar photosynthetic rate and water use efficiency when compared to non-injured leaflets ( Figure 5). Alternatively, when infested with S. cosmiodes, peanut leaflets showed decreased photosynthetic rate and water-use efficiencies and high stomatal conductance compared to S. bosqueella and the control.

Discussion
The main purpose of this study was to characterize relationships between peanut leaf defoliation and leaf reflectance and gas exchange parameters, focusing on two major South American peanut defoliators, S. bosqueella and S. cosmioides. We demonstrated that peanut leaf reflectance is different between herbivory by the two larval species, but similar among real and simulated defoliation. Similar results were obtained based on peanut leaf gas exchange, in which differences were noted 48 h after injury. Based on these results, the potential of using changes in leaf reflectance simultaneously with gas exchange parameters as indicators of biotic stress in the peanut is of particular interest to the study of plant-insect interactions and for incorporating remote sensing into peanut IPM programs.
We identified certain spectral regions that were especially sensitive to variation in plant stressors. Stress can be defined as "a departure from optimal physiological conditions" [40,41] and in our study, we observed that stressed leaves presented high reflectance in both the visible and NIR portions of the electromagnetic spectrum compared to non-stressed leaves. Generally, plant stress results in an increase in reflectance in the visible portion of the spectrum (400-700 nm) due to a decrease in chlorophyll a and b absorption; the opposite generally occurs in the NIR portion of the

Discussion
The main purpose of this study was to characterize relationships between peanut leaf defoliation and leaf reflectance and gas exchange parameters, focusing on two major South American peanut defoliators, S. bosqueella and S. cosmioides. We demonstrated that peanut leaf reflectance is different between herbivory by the two larval species, but similar among real and simulated defoliation. Similar results were obtained based on peanut leaf gas exchange, in which differences were noted 48 h after injury. Based on these results, the potential of using changes in leaf reflectance simultaneously with gas exchange parameters as indicators of biotic stress in the peanut is of particular interest to the study of plant-insect interactions and for incorporating remote sensing into peanut IPM programs.
We identified certain spectral regions that were especially sensitive to variation in plant stressors. Stress can be defined as "a departure from optimal physiological conditions" [40,41] and in our study, we observed that stressed leaves presented high reflectance in both the visible and NIR portions of the electromagnetic spectrum compared to non-stressed leaves. Generally, plant stress results in an increase in reflectance in the visible portion of the spectrum (400-700 nm) due to a decrease in chlorophyll a and b absorption; the opposite generally occurs in the NIR portion of the spectrum (750-2500 nm), with a decrease in reflectance driven by changes in the internal leaf structure [15,42]. The increase in NIR reflectance we observed in stressed peanut leaves could be related to variations in water content due to the infestation, which leads to a degeneration of the internal leaf structure at the cellular level [43,44]. Also, NIR reflectance is affected primarily by the leaf structure and is strongly associated with photosynthetic performance [45].
In this context, low photosynthetic rates (A) and changes in water use efficiency (WUE) observed in leaves injured by S. cosmioides may explain the high reflectance in the NIR portion of the spectrum and could be a useful indicator of photosynthetic potential. These results suggest that the leaf internal structure and leaf pigments are important to the classification of the peanut leaflets infested by the most important defoliator larvae that occur in South American peanut crops.
Similarly, ANOVA F-values analysis suggested that the NIR spectral data can be used to detect stress between S. bosqueella, S. cosmioides, and non-infested leaves. The importance of the NIR portion of the spectrum in characterizing biotic stress caused by defoliators was reported by Huang et al. [46] in rice leaves injured by Cnaphalocrocis medinalis (Lepidoptera: Crambidae) and Sudbrink et al. [47] in cotton (Gossypium hirsutum) injured by Spodoptera exigua and Trichoplusia ni (Lepidoptera: Noctuidae). Our finding is an important one because a wide variety of vegetation indices combine reflectance data in the NIR region to detect biotic stress, including, for example, the Normalized Difference Vegetation Index (NDVI) and the Green Normalized Difference Vegetation Index (GNDVI) [48].
Random Forest classification models of leaf reflectance provided a high accuracy of detection for S. bosqueella and S. cosmioides infestations. Also, the Kappa statistic values were greater than 0.80, representing a good agreement between the reference and predicted data [49]. Random Forest Classification models have been successfully applied to a wide range of hyperspectral data analyses [50][51][52][53][54][55]. It is important to emphasize that in our study we controlled for the effects of environmental factors and reflectance measurements were obtained from plant leaves and not plant canopies. With this in mind, in-field remote sensing presents additional challenges, including the leaf angle to the sensor and leaf location in the canopy with respect to shadowing, as well as other environmental variables that could affect radiometric resolution and consequently classification accuracy [12,50,56,57]. Also, due to recent advances in spatial classification algorithms, additional studies must evaluate spatial image classification for pest detection in peanut.
Similar to what has been observed in other studies [30,[58][59][60][61][62][63][64][65], we did not observe significant effects on either the assimilation/respiration (photosynthetic rate) or in stomatal conductance, intercellular CO 2 concentration, and transpiration at 24 h after defoliation by leaf-mass consumers. Differences in gas exchange of peanut leaves were only observed 48 h after infestation in the S. cosmioides injury treatment.
Therefore, in contrast to S. cosmioides, our results suggest that the principal effect of S. bosqueella larval-induced herbivory in peanut leaflets seems to be a leaf-mass reduction, not reduction of photosynthetic capacity of remaining leaf tissue. However, the injury by S. bosqueella is visually different. Rather than gross leaf-mass consumption like S. cosmioides, S. bosqueella leaves a layer of interveinal tissue, referred to as "windowpane injury" (Figure 1a). In addition, S. bosqueella feeds on meristematic tissues (buds), which can reduce plant development and alter plant architecture [20,66].
According to Higley et al. [39] and Peterson [18], insect herbivores can be grouped into categories known as injury guilds based on plant response, such as population or stand reduction, leaf-mass reduction, leaf photosynthetic-rate reduction, leaf senescence alteration, light reduction, assimilate removal, water-balance disruption, seed or fruit destruction, architecture modification, and phenological disruption. From an agronomic and pest management perspective, injury guilds can be used to develop multiple-species economic injury levels (EILs), as well contribute to accurate pest sampling programs [8, 18,67,68]. However, results from this study suggest that injury from S. bosqueella and S. cosmioides most likely cannot be placed in the same injury guild for peanut cultivars that belongs to the runner vegetative group.
Because peanut leaves respond similarly to both real and simulated injury, both spectrally and physiologically, it is possible to use simulated injury to evaluate leaf-mass consumption from S. cosmioides. This is important because simulating insect defoliation offers distinct advantages in evaluating plant primary physiology, growth, and yield [69]. Importantly, however, our data here are restricted to only a particular developmental stage (i.e., flowering). In addition, because our experiment was conducted in controlled conditions, other studies will need to assess if environmental factors could have an influence on peanut plant spectral and biotic responses in field conditions.
We determined that remote sensing can be used to differentiate injured and non-injured leaves. That is an important step to use remote sensing in peanut sampling programs. However, for decision making, more information is needed (e.g., inferences about the presence of insects which are the target of control measures). Consequently, new studies should address this to improve remote sensing-based sampling programs in the peanut.

Conclusions
Our results reported here support the hypothesis that remote sensing in the NIR (777.42-1000 nm) portion of the spectrum can be used for classification of S. bosqueella and S. cosmioides injury. Also, we verified that S. bosqueella and S. cosmioides should be considered separately in sampling programs, and studies with simulated injury can be used to evaluate defoliator-induced stress in peanut. This information will contribute to incorporating remote sensing as an instrument to leverage the adoption of IPM programs for the peanut.