Enhancing Kiwi Bacterial Canker Leaf Assessment: Integrating Hyperspectral-Based Vegetation Indexes in Predictive Modeling ^†

Abstract

The potential of hyperspectral UV–VIS–NIR reflectance for the in-field, non-destructive discrimination of bacterial canker on kiwi leaves caused by Pseudomonas syringae pv. actinidiae (Psa) was analyzed. Spectral data (325–1075 nm) of twenty kiwi plants were obtained in vivo and in situ with a handheld spectroradiometer in two commercial kiwi orchards in northern Portugal over 15 weeks, resulting in 504 spectral measurements. The suitability of different vegetation indexes (VIs) and applied predictive models (based on supervised machine learning algorithms) for classifying non-symptomatic and symptomatic kiwi leaves was evaluated. Eight distinct types of VIs were identified as relevant for disease diagnosis, highlighting the relevance of the Green, Red, Red-Edge, and NIR spectral features. The class prediction was achieved with good model metrics, achieving an accuracy of 0.71, kappa of 0.42, sensitivity of 0.67, specificity of 0.75, and F1 of 0.67. Thus, the present findings demonstrated the potential of hyperspectral UV–VIS–NIR reflectance for the non-destructive discrimination of bacterial canker on kiwi leaves.

1. Introduction

The Bacterial Canker of Kiwi (BCK) disease, caused by Pseudomonas syringae pv. actinidiae (Psa), is accountable for numerous epidemics in kiwi orchards annually [1,2]. Scouting- and laboratory-based techniques (e.g., Polymerase Chain Reaction—PCR—and Enzyme-linked Immunosorbent assay—ELISA) are currently applied as diagnostic procedures. While insightful, these methods are hindered by their labor intensiveness, time requirements, complex sampling, and unsuitability for rapid real-time field decisions, thus limiting their use in disease monitoring and field mapping [3,4].

Early diagnosis, especially before symptoms’ visible appearance, is of paramount importance in plant disease diagnosis. This proactive approach allows for timely and targeted interventions, reducing the spread of the disease and minimizing crop damage. It also enables more efficient resource allocation and cost-effective management strategies, safeguarding agricultural productivity and food security.

Recently, Hyperspectral Spectroscopy (HS) techniques have alternatively been applied as an innovative indirect plant disease diagnostic tool capable of retrieving relevant information about the host–pathogen interactions related to the host’s biochemical and biophysical modifications. Briefly, changes promoted by pathogens related to plants’ pigment concentrations and physiological processes (e.g., photosynthesis) produce changes in the quantitative and qualitative patterns of plants’ spectral behavior, namely in the visible region of the electromagnetic spectrum (VIS, 400–700 nm). In turn, modifications in leaf water levels, chemical composition (i.e., lignin and protein content), structure, and internal scattering processes impact the spectral signatures in infrared wavelengths (IR, 800–2500 nm) [5,6]. Hence, HS could be successfully applied for the detection of pests [7,8], fungi [9,10], bacteria [11], and viruses [12] affecting different crops, even at non-symptomatic stages [13].

Nevertheless, data collected from HS frequently present redundant information from proximal bands. Hence, only a few spectral wavelengths might help to assess plant disease [14,15]. Approaches including statistical signal processing, mathematical combinations of different bands, and applied predictive models may be computed to extract meaningful information, reduce data dimensionality, and/or select relevant features [16,17]. Vegetation Indices (VIs) exemplify these techniques, as they are numerical measures derived from parametric formulations of the different spectral bands or wavelengths associated with essential plant biophysical parameters like photosynthetic pigments, structural molecules, and water content. These indices are widely employed because of their simplicity and comprehensiveness, users’ limited knowledge requirements, fast processing, and computational inexpensiveness [18]. VIs formalizations can be combinations of two bands (most frequent case), three bands, and four or more bands (combination of two VIs) [18]. Among the most frequently computed VIs are the Normalized Difference Vegetation Index (NDVI) [19,20] and the Enhanced Vegetation Index (EVI) [21,22], which are effective at assessing parameters related to the plant’s status and structure. VIs developed specifically for parameter estimation (e.g., leaf’s photosynthetic pigment and water levels) are frequently employed. Some examples include the Anthocyanin reflectance index (ARI) [23], Browing Reflectance Index (BRI) [24], Chlorophyll Green (Chlgreen) [25], and Coloration Index (CI) [26], among others. Furthermore, Vegetation Indices (VIs) can undergo band optimization procedures, enhancing their spectral sensitivity to the target parameters and enabling a more comprehensive analysis of the variable under consideration [27].

The present research aims to compare the suitability of VIs and classification modeling for discriminating non-symptomatic and BCK symptomatic kiwi leaves in-field, using ground-level UV–VIS hyperspectral reflectance assessments.

2. Methods

2.1. Experimental Site

Two commercial orchards cultivated with kiwi plants (Actinidia deliciosa) were monitored in 2020, both located in Guimarães, Portugal: one situated in Caldas das Taipas (CT; 41°29′09.8″ N 8°21′54.3″ W) and the other in Briteiros (BT; 41°30′53.3″ N 8°19′20.5″ W). Twelve feminine kiwi plants of the variety Bo.Erika^® in CT and eight in BT were chosen, identified with tape and classified according to the absence or presence of typical BCK visual symptoms (i.e., minor greasy dark lesions which turn from brown to black over time, and are usually randomly spread on leaves’ surface). Visual phenotyping was performed on both the adaxial and abaxial sides of the leaves.

2.2. Ground-Based Hyperspectral Reflectance Acquisition

A portable spectroradiometer (ASD FieldSpec^® HandHeld 2, ASD Instruments, Boulder, CO, USA) was used for leaf spectra capturing between May and August 2020 (9 visits), ending when the full development of Psa symptoms was reported in the plants’ growing season. More details of the spectra measurement procedure can be found in [28].

A total of 504 spectral averaged signatures were collected at both test sites, and the dataset was balanced regarding class distribution (

Table 1. Number of test sites, visits, plants, and leaves assessed per location of experimental sites [28].

Experimental Site	Sites	Visits	Plants	Non-Symptomatic Leaves	Symptomatic Leaves	Total Measurements
Briteiros (BT)	1	9	8	89	127	216
Caldas das Taipas (CT)	1	8	12	192	96	288
Total	2	17	20	281	223	504

).

2.3. Data Modeling

Spectral pre-processing was performed via the computation of a multiplicative scatter correction (MSC) [29]. A total of 751 wavelength predictors were considered (325–1075 nm). Due to the overlapping nature of hyperspectral data and multi-scale interference, auto-correlated signals may arise across various scales [30]. Thus, techniques capable of identifying the most relevant wavelengths or bands for discrimination and not considering redundant information are essential.

In this regard, the reflectance data were processed into 32 spectral VIs, resulting in 41 distinct band combinations (

Table A1. Spectral Vegetation Indices (VIs) computed in this study.

Vegetation Indices	Formula	Ref.
Ashburn Vegetation Index (AVI)	$2.0 \times NIR - RED$	[36,37]
Anthocyanin reflectance index (ARI)	$\frac{1}{GREEN}-\frac{1}{RED}$	[23]
Blue Green Pigment Index (BGI)	$\frac{BLUE}{GREEN}$	-
Browning Reflectance Index (BRI)	$\frac{\frac{1}{GREEN}-\frac{1}{RED}}{NIR}$	[24]
Chlorophyll Green (Chlgreen)	${\left(\frac{NIR}{GREEN}\right)}^{(-1)}$	[25]
Coloration Index (CI)	$\frac{RED-BLUE}{RED}$	[26]
Chlorophyll Index Green (CIgreen)	$\frac{NIR}{GREEN}-1$	[22,38,39]
Chlorophyll Index Red Edge (CIrededge)	$\frac{NIR}{REDEDGE}-1$	[22,38,39]
Chlorophyll Vegetation Index (CVI)	$NIR\times \frac{RED}{{GREEN}^2}$	[40]
Double Difference Index (DD)	$(749nm-720nm)-(701nm-672nm)$	[41,42]
Enhanced Vegetation Index (EVI)	$2.5\times \frac{NIR-RED}{\left(NIR+6RED-7.5BLUE\right)+1}$	[21,22]
Green Atmospherically Resistant Vegetation Index (GARI)	$\frac{NIR-(GREEN-(BLUE-RED\left)\right)}{NIR-(GREEN+(BLUE-RED\left)\right)}$	[39,43]
Green-Blue NDVI (GBNDVI)	$\frac{NIR-\left(GREEN+BLUE\right)}{NIR+\left(GREEN+BLUE\right)}$	[44]
Global Environment Monitoring Index (GEMI)	$\left(n\times \left(1-0.25n\right)-\frac{RED-0.125}{1-RED}\right)$ $n=\frac{2\times \left({NIR}^2-{RED}^2\right)+1.5\times NIR+0.5\times RED}{NIR+RED+0.5}$	[45]
Simple Ratio Greenness Index (GI)	$\frac{GREEN}{RED}$	[42,46]
Green Normalized Difference Vegetation Index (GNDVI)	$\frac{NIR-GREEN}{NIR+GREEN}$	[22,38]
Tasselled Cap—Vegetation (GVI)	$\begin{array}{ll}-0.2848\times Blue& -0.2435\times Green-0.5436\times Red\\ & +0.7243\times NIR+0.0840\times SWIR\\ & -0.1800\times SWIR\end{array}$	[47,48]
Infrared Percentage Vegetation index (IPVI)	$\frac{\frac{NIR}{NIR+RED}}{2}\times (NDVI+1)$	[49,50]
Log Ratio (LogR)	$log\left(\frac{NIR}{RED}\right)$	-
Misra Green Vegetation Index (MGVI)	$-0.386\times GREEN-0.530\times RED+0.535\times REDEDGE+0.532\times NIR$	[37,51]
Modified NDVI (mNDVI)	$\frac{NIR-RED}{NIR+RED-2\times BLUE}$	[46,52]
Modified Simple Ratio (mSR)	$\frac{NIR-BLUE}{RED-BLUE}$	[46,50]
Modified Simple Ratio 2 (mSR2)	$\left(\frac{NIR}{RED}\right)-\frac{1}{\sqrt{\left(\frac{NIR}{RED}\right)+1}}$	[53]
Normalized Difference NIR / Red Normalized Difference Vegetation Index (NDVI)	$\frac{NIR-RED}{NIR+RED}$	[19,20]
Normalized Green (NG)	$\frac{GREEN}{NIR+RED+GREEN}$	[54]
Normalized Near Infrared (NNIR)	$\frac{NIR}{NIR+RED+GREEN}$	[54]
Hyperspectral Perpendicular VI (PVIhyp)	$\frac{NIR-a\times 807-b}{{(1+a^2)}^{0.5}}$ $a=1.17,b=3.37$	[48]
Plant Senescence Reflectance Index (PSRI)	$\frac{RED-BLUE}{NIR}$	[55,56]
Reflectance at the Inflexion Point (Rre)	$\frac{RED+NIR}{2}$	[57]
Red-Edge Stress Vegetation Index (RVSI)	$\frac{718+748}{2}-733$	-
Structure Intensive Pigment Index (SIPI)	$\frac{NIR-BLUE}{NIR-RED}$	[41,46]
Simple Ratio (SR)	$\frac{NIR}{RED}$	-

). To calculate them, the wavelengths considered were: (i) the ones enumerated in their original formula (as indicated in Table A1) or (ii) default values chosen by the authors, namely 450 nm (representing the Blue region of the electromagnetic spectrum), 550 nm (Green), 680 nm (Red), 700 nm (Red Edge), and 800 nm (NIR).

Applied predictive modeling was then performed using a model with a built-in Feature Selection (FS) method called Flexible Discriminant Analysis (FDA). Leaf symptomatology was used as a binary variable in the models tested, taking the values of ‘No’ (asymptomatic) and ‘Yes’ (symptomatic). The dataset was split into training (70% of random observations) and validation data (the remaining 30% of the observations), following a holdout method. A resampling approach was performed, followed by a repeated cross-validation strategy using repeated 10-fold cross-validation to estimate the model evaluation criteria. The confusion matrix (CM), accuracy score, kappa coefficient, and F1-score were considered to determine the model’s performance. A detailed description of these metrics applied and the R packages used can be found in [28].

3. Results

The model results showed the capacity of classifying the kiwi leaf measurements into ‘Non-symptomatic’ and ‘Symptomatic’ with a 0.71 accuracy (proportion of correctly classified instances), 0.42 Cohen’s kappa (agreement between predicted and actual classes beyond random occurrence), 0.67 sensitivity (ability to identify diseased measurements), 0.75 specificity (ability to identify healthy assessments), and 0.67 F1 score (harmonized measure of precision and recall) for the test set (

Table 2. Classification results of the Flexible Discriminant Analysis (FDA) model computed for the train and test datasets. Legend: Acc.—Accuracy, Kap.—Kappa coefficient, Sen.—Sensitivity, Spe.—Specificity, Pre.—Precision, Rec.—Recall, and F1—F1 score.

Modeling Approach		Acc.	Kap.	Sen.	Spe.	Pre.	Rec.	F1
FDA	Train	0.76	0.48	0.68	0.80	0.73	0.68	0.70
	Test	0.71	0.42	0.67	0.75	0.68	0.67	0.67

). The confusion matrix (CM) results (

Table 3. Vegetation Index (VI) importance for class discrimination and Confusion Matrix (CM) results according to Flexible Discriminant Analysis. Legend: Pred—Predicted, ‘No’—Non-symptomatic, and ‘Yes’—Symptomatic.

VI	Wavelength (nm)	Importance (a.u.)	CM Train
Chlgreen	553, 800	100	Pred	‘No’	‘Yes’
mSR	705, 750	67.15	‘No’	157	50
CI	450, 700	52.94	‘Yes’	40	107
GI	554, 677	44.45
BRI	450, 690	40.55	CM Test
AVI	400, 994	33.71	Pred	‘No’	‘Yes’
PVIhyp	800, 1000	24.46	‘No’	63	22
Chlgreen	530, 730	19.65	‘Yes’	21	44
Rre	670, 780	16.46

Chlgreen—Chlorophyll Green, mSR—Modified Simple Ratio, CI—Coloration Index, GI—Simple Ratio Greenness Index, BRI—Browning Reflectance Index, AVI—Ashburn Vegetation Index, PVIhyp—Hyperspectral perpendicular VI, and Rre—Reflectance at the inflexion point.

) demonstrated that 63 samples were correctly classified as non-symptomatic (True Negatives), and 44 as symptomatic (True Positives). Nevertheless, 21 measurements were wrongly classified as symptomatic (False Positives), and 22 as non-symptomatic (False Negatives). Thus, the model performed better at predicting non-symptomatic assessments than symptomatic measurements. These findings indicate a reasonably effective model performance, with an overall ability to distinguish between classes and make accurate predictions.

The built-in Feature Selection tool highlighted eight distinct VIs for sample discrimination, namely the Chlorophyll Green (Chlgreen), modified Simple Ratio (mSR), Coloration Index (CI), Simple Ratio Greenness Index (GI), Browning Reflectance Index (BRI), Ashburn Vegetation Index (AVI), Hyperspectral perpendicular VI (PVIhyp), and Reflectance at the inflexion point (Rre). These VIs are mostly based in the NIR, Red, and Green regions of the electromagnetic spectrum (Table 2).

4. Discussion

Eight distinct VIs (nine wavelength combinations) were identified as highly relevant for disease discrimination. They mostly considered the NIR, Green, and Red spectral regions. These findings present biological significance since they are coherent with the impact of Pseudomonas syringae pv. actinidiae (Psa) in kiwi leaves. Briefly, these pathogens cause modifications in pigment concentration and physiological processes (e.g., photosynthesis), resulting in changes in plants’ spectral behavior in the VIS wavelengths (Blue, Green, and Red). Furthermore, they cause changes in leaf water levels, chemical composition (namely lignin and protein content), structure, and internal scattering processes, which impact the NIR features [5,6]. Similar spectral regions were also identified as relevant for late blight, target, and bacterial spots detection in tomato leaves [31], and for the assessment of Cercospora leaf spot, sugar beet rust, and powdery mildew in sugar beet plants [32]. The model evaluation metrics also supported the model’s ability in discriminating non-symptomatic from symptomatic samples. The model’s performance may be enhanced by further fine-tuning, particularly in addressing the model’s sensitivity and minimizing the occurrence of false negatives.

Hyperspectral data may have redundant information in adjacent bands, and only a few wavelength features might be interesting in classifying a diseased plant [15]. For this reason, in crop remote sensing (ground-, aerial-, and satellite-based solutions), spectral VIs are still the most common approaches studied for identifying and managing biotic stresses in different crops [33]. Despite its substantial inherent potential, the discernment of the responsiveness of this extensive array of VIs to the target variable remains occasionally ambiguous. Furthermore, concerns related to the susceptibility to disturbances from confounding elements can arise, mostly encompassing fluctuations in leaf or canopy properties, background soil reflectance, solar illumination, and atmospheric composition. Such a confluence of factors can generate instabilities in the spectral attributes of surfaces [34]. Furthermore, VIs were developed when the first applications of broadband sensors occurred, when only a small set of spectral bands were available and the computational power was limited. With the development of narrowband devices (i.e., with a few hundred spectrally narrow bands), these VIs may use the available information within the spectral observation range inefficiently, often relying on only a partial spectral subset. Algorithms for extracting optimized band information were thus created, utilizing well-established index formulations such as simple ratios and normalized differences. These algorithms involved the correlation of all the potential band combinations to generate 2D correlation matrices, allowing for the visual identification of the most effective band combinations. Nevertheless, this approach can optimize indices which are strongly case-specific, successfully optimized for local applications but not to generic cases [18]. FS non-parametric methods, which evaluate all the spectral wavelengths provided by hyperspectral sensors, constitute an interesting option for disease assessment, providing more robust and customized information for modeling data class characteristics and a greater model performance [28,35]. Thus, future research is needed to better explore the different information extraction (e.g., modeling) approaches suitable for comprehending plant–pathogen interactions and their effects on host spectral behavior.

5. Conclusions

The present work aimed to apply hyperspectral reflectance in-field measurements for the diagnosis of bacterial canker of kiwi (BCK) disease, which is caused by the bacteria Pseudomonas syringae pv. actinidiae (Psa). Different vegetation indices were computed and later used to classify symptomless and symptomatic kiwi leaves’ signatures. Chlgreen, mSR, CI, GI, BRI, AVI, PVIhyp, and Rre were signed as the most relevant for disease discrimination, highlighting the Green, Red, Red Edge, and NIR regions of the electromagnetic spectrum. These findings are in line with the metabolic and structural changes promoted by the pathogen in the host tissues. Classification modeling allowed for disease discrimination with fair model metrics, showing the suitability of this approach for disease assessment. Nevertheless, further research exploring different Feature Selection methods considering a broader range of wavelengths is advised.