Characterization of Maize, Common Bean, and Avocado Crops under Abiotic Stress Factors Using Spectral Signatures on the Visible to Near-Infrared Spectrum

Abstract

Abiotic stress factors can be detected using visible and near-infrared spectral signatures. Previous work demonstrated the potential of this technology in crop monitoring, although a large majority used vegetation indices, which did not consider the complete spectral information. This work explored the capabilities of spectral information for abiotic stress detection using supervised machine learning techniques such as support vector machine (SVM), random forest (RF), and neural network (NN). This study used avocados grown under various water treatments, maize submitted to nitrogen deficiency, and common beans under phosphorous restriction. The spectral characterization of the crops subjected to abiotic stress was studied on the visible to near-infrared (450 to 900 nm) spectrum, identifying discriminative bands and spectral ranges. Then, the advantages of using an integrated approach based on machine learning to detect abiotic stress in crops were demonstrated. Instead of relying on vegetation indices, the proposed approach used several spectral features obtained by analyzing the discriminative signature shape, applying a spectral subset band selection algorithm based on similarity, and using the minimum redundancy maximum relevance (MRMR), F-test and chi-square test ranks for feature selection. The results showed that supervised classifiers applied to the spectral features outperform the accuracies obtained from vegetation indices. The best common bean results were obtained using SVM with accuracies up to 91%; for maize and avocado, NN obtained 90% and 82%, respectively. It is noted that detection accuracy depends on various factors, such as crop type, genotype, and level of stress.

1. Introduction

A technology gaining strength for crop monitoring is spectral sensors, such as spectrometers and multispectral and hyperspectral cameras [1,2]. Spectral sensors measure the energy emitted or reflected by a surface across the electromagnetic spectrum. Spectrometers typically measure the radiance in a high spectral resolution within hundreds or thousands of bands [3]. Instead, multispectral and hyperspectral cameras combine the capabilities of spectrometers with imagery, collecting both spatial and spectral information [3,4]. The spectral response of vegetation showcases a varied yet distinctive pattern that changes according to the phenological stage, health, and environmental conditions [5,6]. Due to chlorophyll and other pigments, the visible region has a low reflectance with absorption peaks around 490 and 660 nm [7]. An important feature within the vegetation’s spectral signature lies in the red edge, spanning from 690 to 720 nm. Here, a diminished red reflectance sharply increases around 800 nm, attributed to internal leaf structure and water content [7]. Moving into the near-infrared spectrum (700 to 1300 nm), plants exhibit a heightened reflectance, the intensity of which is contingent upon internal leaf structure [7].

One potential application of spectral data analysis is detecting limiting factors or stresses [8,9,10]. Limiting factors in vegetation are classified as biotic or abiotic according to their nature. Biotic factors are associated with adverse effects caused by living beings, such as fungi and pests; abiotic factors are related to nutritional and water deficiencies, and environmental impacts caused by rain, drought, and wind [11]. For instance, water stress closes stomata and impedes photosynthesis and transpiration, leading to changes in leaf color and temperature. On the other hand, nutritional stress is directly related to the photosynthetic process [11].

Water and nutritional stress, which are often related, can be detected using spectral data through a vegetation index, biophysical model, or nonparametric model based on machine learning. Vegetation indices (VIs) commonly use up to three bands to compute ratios; the most frequently used bands are red, green, and near-infrared [12,13]. These ratios are related to phenological phenomena or stress factors with low computational cost [12]. An extensive review of VIs is presented in [13]. Vegetation indices are utilized in agriculture to predict leaf area, chlorophyll content, plant water content, biomass, and yield. Using VIs requires careful consideration of the strengths and limitations of estimating plant conditions [14]. For instance, the normalized difference vegetation index (NDVI) is widely used in modern agriculture. This index is based on the red and near-infrared reflectance with ranges from −1 to 1, where positive values indicate increased green cover and negative values reveal non-vegetable features [12,13,14]. Other indices commonly used in agriculture are related to nitrogen contents, such as the ratio vegetation index (RVI) [15], normalized difference red edge index (NDRE) [15], canopy chlorophyll content index (CCCI) [15], and carotenoid reflectance index (CRI) [16]. Others are related to water contents, such as the Medium Resolution Imaging Spectrometer (MERIS) terrestrial chlorophyll index (MTCI) [17].

An alternative to vegetation indices is biophysical models, which explore the interaction of light with leaves. Biophysical models describe spectral variation as a function of land cover, leaf, and soil characteristics [18]. Estimating biophysical characteristics depends on an inversion process, which minimizes the difference between actual and simulated data. Among the biophysical models are the SAIL model [18] for bidirectional reflectance of the vegetation cover; PROSPECT [18], which models the optical properties of the leaves; and Hydrus-1D [19], a one-dimensional finite element model for simulating the movement of water, heat, and multiple solutes in soil. Despite the advances in these models, previous studies have shown that the inversion from spectral data is computationally expensive.

Unlike vegetation index and biophysical models, nonparametric models do not require assumptions about relationships between bands, knowledge of the data distribution, or the inclusion of physical parameters. In this sense, new approaches for characterizing crops based on regression [20,21], random forest [15], neural networks [22], k-nearest neighbors [23], and support vector machines [22] are found in the literature. Previous studies show the potential of spectral data, including spectral signatures and images, for stress detection in different crops [20,21,22,23]. However, most research is based on vegetation indices and linear regression.

In addition, it has proven difficult to generalize a model or method to different types of crops or stresses [24]. Therefore, this paper explores machine learning techniques for detecting stress in three crops: maize (Zea mays), common bean (Phaseolus vulgaris), and avocado (Persea americana cv. Hass). In the case of maize and common beans, this paper presents the stress characterization caused by nitrogen and potassium deficiency, respectively. In the case of avocado, the effects of excess and deficiency of water are analyzed. A feature selection process is carried out for stress detection using discriminative bands and vegetation indices. Three supervised classifiers are compared: support vector machines (SVM), random forests (RF), and neural networks (NN).

This paper seeks to establish the capabilities and limitations of machine learning-based approaches (feature extraction and supervised classifiers) to detect or quantify stress levels in different crops. In the following, we present in detail the experimental design for spectral data collection in crops (Section 2.1), the process for identifying relevant bands and regions from spectral signatures (Section 2.2), and the detection or quantification of stress using feature extraction methods and supervised classifiers (Section 2.3).

2. Materials and Methods

2.1. Spectral Library of Crops under Abiotic Stress

The spectral signatures used in this study consist of data collected from a spectrometer and a hyperspectral camera. The following sections describe the vegetal materials, the experimental setup to induce water and nutritional stress, and the data collection and preprocessing procedures.

2.1.1. Vegetal Materials and Experimental Design

This study used maize, common beans, and avocado crops under semi-controlled conditions. In the case of maize, two datasets were considered. The first one, available in [25], includes spectral signatures from ten maize genotypes subjected to four levels of nitrogen fertilization (three with deficits), corresponding to 25%, 50%, 75%, and 100% of the optimum nitrogen dose established at AGROSAVIA’s La Selva Research Center in Rionegro, Antioquia, Colombia. We used the dataset obtained on 23 June 2022, containing 2033 spectral signatures, referred to as Maize La Selva [25].

The second maize dataset under nitrogen deficiency stress includes two experimental genotypes (V117 and V121). The maize crops were established at the Turipana Research Center of AGROSAVIA in Cerete, Cordoba, Colombia (08°51′04″ N; 75°49′05″ W, 14 m.a.s.l.). This area has an average annual temperature of 29.7 °C, precipitation of 1280 mm, air relative humidity of 80%, daylight of 2190 h yr⁻¹, and evapotranspiration of 1329 mm [26]. Unlike the dataset published in [25], the second dataset was captured under different environmental conditions, where Cerete has a higher temperature and lower precipitation than Rionegro. Trials for the second maize dataset were established on 28 September 2022, using two levels of nitrogen fertilization treatments: 100% (TN100) and 50% (TN50) of the optimum nitrogen dose for maize. The optimum dose was 193.8 kg/ha of N (nitrogen), 74.6 kg/ha of P (phosphorus), and 178.8 kg/ha of K (potassium), which was determined before seeding, considering crop requirements, genotypes, and initial chemical soil conditions. The experimental design was arranged in a randomized complete block with a split-plot arrangement and three replications, where nitrogen doses were the main plots and genotypes were the sub-plots (Figure 1a). The experiments included 12 plots within 300 plants of the same genotype, with six furrows per plot, each 10 m long and spaced 0.2 m apart. The spacing between each plot was 2 m. The distribution of genotypes within each replication was randomly arranged.

The common bean crop was also established at the Turipana Research Center of AGROSAVIA, including two commercial genotypes (Mungo and Caupí). The plants were under two phosphorus fertilization treatments: 100% (TP100) and 50% (TP50) of the optimum phosphorus dose established by a prior chemical soil analysis. In this case, the optimum fertilization was 180 kg/ha of N, 92 kg/ha of P, and 148 kg/ha of K. The applied fertilization was divided into two applications: 50% during seeding and 50% after 15 days. The experimental design was like the maize trial; however, the main plot corresponded to phosphorus doses. The trial consisted of 12 plots, each with 300 plants of the same genotype and six furrows of 10 m in length, spaced 0.2 m apart (Figure 1c). The crop was established on 22 October 2022.

Finally, an avocado plant experiment was established in a greenhouse at La Selva Research Center of AGROSAVIA (Rionegro, Antioquia, Colombia). For this experiment, two rootstocks of local avocado landraces were used (ANNS88 and ANGUI52). The trial involved grafting cv. Hass avocados onto these rootstocks at 13 months; these were chosen for their adaptation to climatic conditions and biological irrigation in the region. The plants were grown under various water treatments, including 50% (TW50), 75% (TW75), 125% (TW125), 150% (TW150), and 100% (TW100) of the optimum level as determined by [27] for the production zone of Antioquia, a mainly avocado producer in Colombia. The optimum level for water was determined to be 222 mL per tree per day. The treatments were applied one month after the plants’ acclimatization to the greenhouse. All treatments were laid out in a split-plot arrangement under a randomized complete block design with nine replications, where the whole plot and sub-plots were rootstocks and water regimes, respectively, for a total of 90 plots (Figure 1a). The measurements were collected on 18 November 2021.

2.1.2. Data Collection

The spectral signatures of maize and common beans were collected using a FLAME S VIR NIR spectrometer (Ocean Optics, Orlando, FL, USA), as described in [25]. The spectrometer collects 2049 bands from 350 to 1000 nm. A customized clamp [28] was used to integrate the spectrometer and a light source, holding the fiber 5 mm from the surface at an angle of 45° between the surface and the light. The spectrometer fiber was positioned perpendicularly to the surface. The captured spectrum resulted from an average of 10 measurements, reducing the effects of the leaf structure or disturbances. For each experiment, the spectra from black and white standards were collected before the measurements on the leaves. OceanView 2.0 software (https://www.oceanoptics.com/software/oceanview/, accessed on 21 August 2024) was used to capture and calibrate the spectral signatures.

On the other hand, a hyperspectral image of avocados was obtained using the HySpex Mjolnir VS-620. PARGE^® software (Version 3.5) was used for orthorectification, and DROACOR (Version 2.0) for atmospheric correction and reflectance retrieval (https://parge.com/, accessed on 21 August 2024). The experiment was located at 2140 m.a.s.l. and the image was approximately 40 m from the surface. The pixel size obtained was 0.02 m. Each plant was manually segmented and labeled to extract spectral signatures from the hyperspectral image. The pixels identified within a leaf were considered the spectral signatures.

2.1.3. Data Preprocessing

This study used bands between 450 and 950 nm for all signatures (maize, common bean, and avocado). A 10-point sliding window filter was applied to all datasets to improve the signal-to-noise ratio. Finally, outliers were removed using a criterion based on three standard deviations. The final spectral library for each crop is described in

Table 1. Number of spectra acquired from maize, bean, and avocado crops per genotype and treatment after preprocessing.

Treatment	ID	Dose (%)	Spectral Genotype
Maize Turipana
			V117	V121
Nitrogen (N) stress	TN100	100	309	318
	TN50	50	318	325
Common bean Turipana
			Mungo	Caupí
Phosphorus (P2O5) stress	TP100	100	310	301
	TP50	50	339	324
Avocado La Selva
			ANNS88	ANGUI52
Water stress	TW50	50	1137	933
	TW75	75	861	1245
	TW125	125	1166	1002
	TW150	150	1207	1226
	TW100	100	610	731

.

2.2. Spectral Characterization and Abiotic Stress Detection

First, characteristic points of vegetation signatures for each crop were identified, including the red edge region, chlorophyll absorption, and green peak. Additionally, due to the redundancy among adjacent bands, relevant bands were selected using a spectral band subset selection algorithm. Subsequently, several vegetation indices were explored as features for stress detection. Figure 2 presents a flow chart about the proposed methodology.

2.2.1. Vegetation Signature Feature Identification

The signature characterization was performed by defining spectral bands and ranges where a specific behavior was expected for vegetation [5]. These bands and ranges can change according to vegetation type, genotype, and health status.

The first derivative of the average signature from healthy plants was computed. Local maxima and minima were established in the regions where the characteristic behaviors of the vegetation signature were expected. The local minimum in the blue range identified one of the maximum absorption points of chlorophyll (MAC). Similarly, the local minimum of the red range was identified as the other MAC. The green peak was selected as the local maximum in the visible range, and the first local maximum after the red edge was determined as the beginning of the NIR plateau. The start and end points of the red edge were detected using the maximum slope.

2.2.2. Spectral Band Subset Selection

Band selection in spectral data improves the accuracy and quality of analysis by reducing noise and highlighting specific features. It also optimizes data processing by reducing dimensionality, and facilitates the interpretation and visualization of relevant information [29]. The similarity-based unsupervised method is used for band reduction and demonstrates the desired behavior even without prior information. This method identifies the most distinctive and informative bands based on the measurement of band similarity. This algorithm was selected because it performs significantly better information preservation and class separability than other algorithms [29].

The algorithm was implemented in MATLAB and applied to each crop’s spectral library, selecting the most discriminant bands and identifying regions of interest (regions with the highest density of discriminant bands).

2.2.3. Vegetation Indices

Vegetation indices are ratios of different bands that can identify specific vegetation features.

Table 2. Vegetation indices were used in this study to detect abiotic stress.

Index	Equation
Normalized Difference Vegetation Index (NDVI)	$NDVI={\displaystyle \frac{R_{NIR}-R_R}{R_{NIR}+R_R}}$
MERIS Terrestrial Chlorophyll Index (MTCI)	$MTCI={\displaystyle \frac{R_{NIR}-R_{RE}}{R_{RE}+R_R}}$
Chlorophyll Estimation Green Index (CIG)	${CI}_G={\displaystyle \frac{R_{NIR}}{R_G}}-1$
Chlorophyll Estimation Red Edge Index (CIRE)	${CI}_{RE}={\displaystyle \frac{R_{NIR}}{R_R}}-1$
Ratio Vegetation Index (RVI)	$RVI={\displaystyle \frac{R_{NIR}}{R_R}}$
Normalized Difference Red Edge Index (NDRE)	$NDRE={\displaystyle \frac{R_{NIR}-R_{RE}}{R_{NIR}+R_{RE}}}$
Canopy Chlorophyll Content Index (CCCI)	$CCCI={\displaystyle \frac{NDRE}{NDVI}}$
Carotenoid Reflectance Index	$CRI={\displaystyle \frac{1}{R_G}}+{\displaystyle \frac{1}{R_{NIR}}}$

NIR: Near Infrared; RE: Red Edge; G: Green; R: Red.

summarizes the eight vegetation indices selected for this study. To obtain the reflectance in the green (R_G), red (R_R), red edge (R_RE), and near-infrared (R_NIR) bands, the average value was computed using the following wavelength ranges: 540 to 570 nm for R_G, 650 to 680 nm for R_R, 690 to 710 nm for R_RE, and 780 to 900 nm R_NIR [13,15].

2.3. Feature Selection and Classification

A new representation space was built using the relevant spectral regions, discriminative bands, and vegetation indices. The most suitable feature subsets were selected using feature selection approaches to improve abiotic stress detection. Feature selection is a process where the most relevant variables are chosen from a more extensive set of features [30,31,32]. In this study, three techniques were compared for feature selection. The first technique was the minimum redundancy maximum relevance (MRMR) rank feature [30]. This algorithm identifies the most optimal feature set that is mutually and maximally dissimilar and can effectively represent the classification variable. The second technique, F-tests univariate feature ranking for regression (FSRF), assesses the importance of each predictor using statistical data [31]. The final technique was FSCC, a univariate feature ranking for classification using chi-square tests to evaluate the data for each predictor [32].

Once the feature spaces were obtained, supervised classifiers were applied to detect or quantify stresses. Classification techniques included in this work were support vector machine (SVM), random forest (RF), and fully connected neural network (NN), specifically the multi-layer perceptron; these methods obtain desired classification results in datasets with high dimensionality.

The implementations of MATLAB 2022a were employed for the three classifiers. First, the three classifiers were used with the similarity-based unsupervised method to obtain the optimal number of bands for each crop according to stress detection accuracy. In this first experiment, the hyperparameters for each classifier were manually adjusted. The templateSVM function was used for the SVM, with a 5th-degree polynomial kernel with an offset equal to 0.1. The RF parameters were set with a minimum leaf size of 40, PCA as the categorical predictor, a maximum tree depth of 50, and a maximum number of splits of 9. Finally, the fitcnet function was employed for the NN, with 25 fully connected layers.

Finally, a comparison was performed among the different representation spaces: the discriminative bands established by the similarity-based algorithm, the regions with the most discriminative bands, these spaces combined with vegetation indices, the vegetation index, and the full spectra signature. In this case, the hyperparameters of the classification methods were optimized. Bayesian optimization was used to mitigate the adjusted error for each result. For each crop, the dataset was divided into 70% of samples for training and 30% for testing. The experiments were conducted on a single computer LENOVO ThinkStation (Beijing, China),with Intel (R) Core (TM) i9-9900K CPU @ 3.60 GHz, 64.0 GB RAM HDD, 1 TB; operating system: Windows 10 Pro 64-bit, MATLAB R2022. The comparison was performed using the overall accuracy and F1 Score. The overall accuracy was calculated as the average of each class’s accuracy, and the F1 Score was the weighted average according to the number of samples per class.

3. Results

3.1. Characterization of Spectral Signatures

Figure 3 and

Table 3. Spectral characteristics for maize, bean, and avocado crops.

Feature		Maize La Selva	Maize Turipana	Common Bean Turipana	Avocado La Selva
MAC blue	W (nm) R	500 0.05068	493 0.04774	496 0.01611	492 0.01216
MAC red	W (nm) R	673 0.04619	671 0.04598	664 0.01688	669 0.01260
Green peak	W (nm) R	554 0.12593	556 0.11229	554 0.09254	553 0.04336
Red edge	W (nm)	[693–748]	[696–748]	[694–744]	[693–753]
NIR plateau	W (nm)	[763–950]	[765–950]	[761–950]	[788–900]

MAC: Maximum absorption points of chlorophyll. W: Wavelength. R: Reflectance.

present the characteristic points for the four datasets selected using the procedure outlined in Section 2.1. The four crops (Maize La Selva, Maize Turipana, Common Bean La Selva, and Avocado La Selva) have similar wavelengths for the maximum absorption points of chlorophyll (MAC) in the blue and red regions and the green peak. In detail, MAC blue is between 492 and 500 nm, MAC red is between 664 and 673 nm, and the green peak is between 553 and 554 nm. However, the reflectance value varies according to the crop. For both Maize La Selva and Turipana, the amplitude of MAC blue is around 0.05, MAC red is around 0.04, and the green peak is 0.11 and 0.12, respectively. In addition, the red-edge range and the NIR plateau also have similar wavelengths for the four crops. The red edge is between 693 and 753 nm, and the NIR plateau is between 761 and 950 nm. Despite the different crops, genotypes, and environmental conditions for each experiment, there is a consistency in the characteristic points within the spectral signatures.

3.2. Spectral Band Selection

The similarity-based unsupervised band selection algorithm was applied to automatically select both the most discriminative bands and the representative spectral regions. Using the most dissimilar bands, the spectra range can be detected where discriminative bands are found most frequently, i.e., regions with a larger density of relevant bands.

Table 4. Relevant spectral regions for maize, common bean, and avocado crops were determined based on the distribution of the selected bands using the similarity-based unsupervised algorithm.

Regions		Maize La Selva	Maize Turipana	Common Bean Turipana	Avocado La Selva
R1	W (nm)	[780–800]	[930–950]	[770–775]	[880–915]
R2	W (nm)	[750–770]	[710–735]	[700–720]	[805–840]
R3	W (nm)	[450–465]	[460–470]	[470–485]	[740–770]
R4	W (nm)	[705–725]	[685–690]	[895–950]	[715–725]
R5	W (nm)	[940–950]	[785–790]	[550–555]	[520–540]

W: Wavelength.

summarizes the selected region bands for each crop, arranged based on their density, where R1 has a higher density than R5. It is important to note that these regions were selected unsupervised, using the similarity metric in the subset band selection algorithm. These regions represent the most dissimilar spectral bands in the crop signature. A range of wavelengths was present in the two maize crops: 785 to 790 nm, 460 to 465 nm, 710 to 725 nm, and 930 to 950 nm. For all crops, it was observed that wavelengths over 700 nm were more frequently selected than the bands in the visible range.

In addition, the spectral-band selection algorithm was used to identify the relevant wavelengths for each crop and genotype. The bands were organized according to their relevance, and the three classification methods were applied iteratively to sets consisting of 2 to 20 bands. The optimal number of bands was selected as the set that obtained the highest accuracy among the three classification methods.

Table 5. Selected bands (nm) using the similarity-based unsupervised algorithm and the overall accuracy classification for genotype and crop (Part A).

Band	Maize Turipana		Common Bean Turipana		Avocado La Selva
	V117	V121	Mungo	Caupí	ANSS88	ANGUI52
B1	946	946	766	769	876	864
B2	706	730	714	714	724	716
B3	457	467	472	467	913	768
B4	732	785	943	945	899	913
B5	827	902	--	898	768	902
B6	925	685	--	936	887	887
B7	948	942	--	551	469	518
B8	685	604	--	--	835	879
B9	937	934	--	--	745	736
B10	905	931	--	--	710	812
B11	921	--	--	--	538	788
B12	913	--	--	--	803	899

and

Table 6. Selected bands (nm) using the similarity-based unsupervised algorithm and the overall accuracy classification for genotype and crop (Part B).

	Maize La Selva 1
	G1	G2	G3	G4	G5	G6	G7	G8	G9	G10
B1	886	927	450	751	927	783	927	784	927	722
B2	720	474	711	713	718	710	450	573	715	927
B3	483	946	792	456	450	450	721	751	450	765
B4	793	783	799	783	780	785	782	779	784	784
B5	808	798	798	802	798	798	802	805	808	806
B6	946	767	767	888	768	768	768	767	767	946
B7	767	450	912	858	863	941	847	847	859	450
B8	863	847	751	767	909	847	877	879	905	847
B9	738	751	729	792	738	751	751	450	751	738
B10	555	731	705	723	557	731	731	731	731	704

¹ The genotypes used are described in [25].

summarize up to twelve bands chosen by the similarity-based algorithm for each crop and genotype.

In the case of V117 of Maize Turipana, ANSS88, and ANGUI52 of Avocado La Selva, and all genotypes of Maize La Selva [25], the process using twelve bands achieved higher overall accuracy. However, there were only ten bands for V121 of Maize Turipana, four for Mungo, and seven for Caupí of Bean Turipana.

The results show that some specific or close bands are recurrently selected as relevant despite the different genotypes. For example, the 946 nm band is chosen for both the genotypes V117 and V121 of Maize Turipana, and G1 and G2 of Maize La Selva; the 714 nm is selected for the Mungo and Caupi genotypes of common bean; and the 913, 899, and 887 nm for the ANSS88 and ANGUI52 avocado genotypes.

3.3. Feature Extraction

New feature representation spaces were built for each crop using the discriminative spectral bands (Table 5 and Table 6) and the vegetation indices (Table 2). The average score obtained from the MRMR, FSRF, and FSCC selection algorithms identified the most relevant features for each crop and genotype.

Table 7. Selected features (bands nm and vegetation indices) for genotype and crop (Part A).

Feature	Maize Turipana		Common Bean Turipana		Avocado La Selva
	V117	V121	Mungo	Caupí	ANSS88	ANGUI52
F1	946	NDRE	CIG	CIG	CIRE	CIRE
F2	457	CCCI	CCCI	769	MTCI	MTCI
F3	732	CIRE	NDRE	936	724	CIG
F4	CRI	MTCI	714	945	736	576
F5	827	902	RVI	898	NDRE	CRI
F6	706	931	NDVI	CCCI	710	716
F7	RVI	946	766	RVI	CCCI	518
F8	NDVI	467	943	NDRE	538	727

NDVI: Normalized Difference Vegetation Index; MTCI: MERIS Terrestrial Chlorophyll Index; CIG: Chlorophyll Estimation Green Index; CIRE: Chlorophyll Estimation Red Edge Index; RVI: Ratio Vegetation Index; NDRE: Normalized Difference Red Edge Index; CCCI: Canopy Chlorophyll Context Index; CRI: Carotenoid Reflectance Index.

and

Table 8. Selected features (bands nm and vegetation indices) for genotype and crop (Part B).

	Maize La Selva
	G1	G2	G3	G4	G5	G6	G7	G8	G9	G10
F1	CCCI	NDRE	CCCI	CIG	CCCI	CIG	NDRE	NDRE	NDRE	CCCI
F2	NDRE	CCCI	NDRE	450	NDRE	NDRE	CCCI	CCCI	CCCI	NDRE
F3	751	751	CIG	459	798	CCCI	MTCI	450	905	704
F4	738	731	705	NDRE	780	NDVI	751	779	784	555
F5	863	767	711	556	768	705	768	847	767	CRI
F6	739	783	CRI	713	CIG	710	CRI	767	808	CIG
F7	450	798	912	MTCI	459	459	721	751	CIG	765
F8	793	MTCI	450	792	MTCI	452	877	805	405	806

NDVI: Normalized Difference Vegetation Index; MTCI: MERIS Terrestrial Chlorophyll Index; CIG: Chlorophyll Estimation Green Index; CIRE: Chlorophyll Estimation Red Edge Index; RVI: Ratio Vegetation Index; NDRE: Normalized Difference Red Edge Index; CCCI: Canopy Chlorophyll Context Index; CRI: Carotenoid Reflectance Index.

summarize the eight most relevant features of each crop. Note that, for each crop, the selected features constitute a combination of bands and vegetation indices, which change not only by crop but also according to the genotype. The results also indicate that selecting the NDRE, CI, and CIRE indices as relevant features is more frequent than NDVI. Finally, bands over 750 nm were commonly identified as the most suitable, as shown in the results in Table 4.

3.4. Stress Detection Using Classification

Finally, the spectral features previously identified were compared for stress detection and quantification using classification. The feature spaces compared were (i) the relevant bands selected by the similarity-based algorithm shown in Table 5 and Table 6 (Select Bands), (ii) the featured selected by MRMR, FSRF, and FSCC shown in Table 7 and Table 8 (Features), (iii) the region bands presented in Table 4 (Region Bands), (iv) the region bands combined with the best VI according to Table 8 (Region Bands and Best VI), (v) the NDVI, (vi) the best VI according to Table 8, and (v) the full spectra (All Bands).

Table 9. Overall accuracy and F1 Score for stress detection in Maize Turipana.

Accuracy/ F1 Score	V117			V121
	SVM	RF	NN	SVM	RF	NN
Select Bands	67.7/80.3	62.7/77.4	62.2/77.2	57.7/73.2	58.6/73.9	57.7/73.5
Features	66.0/78.8	63.2/77.4	64.8/77.5	64.7/77.9	63.7/77.2	61.1/76.0
Region Bands	63.4/76.8	66.7/79.6	63.6/77.1	58.6/73.7	60.1/75.5	57.7/73.3
Region Bands and Best VI	66.0/79.9	65.6/79.8	64.6/78.9	55.5/69.1	57.4/72.4	61.1/76.1
NDVI	48.6/64.0	49.8/65.7	48.6/63.9	50.1/63.6	54.0/70.6	52.1/68.9
Best VI	62.0/76.9	55.3/71.4	60.8/76.2	49.3/59.5	53.8/70.5	56.0/72.4
All Bands	65.1/78.0	56.9/72.2	69.9/82.5	65.7/79.8	61.6/76.5	68.6/81.8

The values in bold indicate the best accuracy and F1 Score for the proposed approach.

,

Table 10. Overall accuracy classification results for stress detection in Common Bean Turipana.

Accuracy/F1 Score	Mungo			Caupí
	SVM	RF	NN	SVM	RF	NN
Select Bands	95.4/97.7	84.4/91.8	93.8/96.9	96.9/98.5	89.4/94.7	95.6/97.7
Features	92.9/96.3	89.2/94.3	93.4/96.7	96.0/98.0	93.4/96.7	96.0/98.1
Region Bands	95.7/97.8	93.1/96.6	94.5/97.2	95.4/97.6	85.9/92.4	95.6/97.7
Region Bands and Best VI	96.3/98.1	91.8/95.7	96.1/98.0	97.4/98.7	91.0/95.3	96.3/98.1
NDVI	45.5/34.7	57.4/72.8	52.9/68.7	55.5/69.7	56.4/72.6	51.8/67.3
Best VI	71.9/83.6	61.3/76.2	68.4/81.2	71.8/84.0	71.4/84.8	75.1/86.6
All Bands	95.2/97.6	87.6/93.7	95.0/97.4	98.2/99.1	91.6/95.7	98.7/99.4

The values in bold indicate the best accuracy and F1 Score for the proposed approach.

,

Table 11. Overall accuracy classification result for stress quantification in Avocado La Selva.

Accuracy/ F1 Score	ANNS88			ANGUI52
	SVM	RF	NN	SVM	RF	NN
Select Bands	49.8/66.7	53.0/69.6	76.5/86.8	51.8/67.3	47.2/64.0	72.8/84.3
Features	43.9/61.9	48.9/66.5	65.3/79.5	45.1/61.2	48.6/66.2	64.6/79.4
Region Bands	42.1/60.2	46.6/64.1	60.0/75.6	40.1/52.9	44.5/62.2	57.2/73.8
Region Bands and Best VI	42.3/60.7	45.6/63.4	61.4/76.9	40.4/46.1	46.3/64.2	57.9/74.4
NDVI	27.4/40.6	31.8/44.9	32.1/53.8	19.7/31.5	38.9/32.5	40.9/56.6
Best VI	42.3/60.7	45.6/63.4	61.4/76.9	29.5/46.6	32.7/51.3	32.5/48.2
All Bands	71.5/84.2	37.1/54.2	71.3/82.5	59.5/74.4	54.2/70.9	83.5/91.3

The values in bold indicate the best accuracy and F1 Score for the proposed approach.

,

Table 12. Overall accuracy classification results for stress detection for Avocado La Selva only T1 and T4.

Accuracy/ F1 Score	ANNS88			ANGUI52
	SVM	RF	NN	SVM	RF	NN
Select Bands	66.8/80.1	63.8/ 78.0	76.7/87.0	61.3/74.3	69.7/81.9	78.3/87.7
Features	68.8/81.4	68.3/ 81.8	67.4/80.2	67.3/79.2	71.7/83.1	82.0/89.9
Region Bands	67.6/ 81.1	67.9/ 80.4	77.2/87.4	59.4/56.0	70.2/81.8	73.9/84.3
Region Bands and Best VI	68.6/ 81.3	66.6/80.2	78.1/87.8	64.9/77.0	66.7/79.8	78.3/87.7
NDVI	51.5/ 68.6	52.8/69.8	53.3/70.2	56.8/73.0	56.8/73.0	57.5/59.2
Best VI	54.0/58.1	65.7/ 48.7	66.2/79.3	56.8/73.0	67.2/73.0	68.0/79.9
All Bands	79.5/ 88.7	69.1/82.0	80.1/89.0	74.4/85.1	74.7/85.2	90.8/95.1

The values in bold indicate the best accuracy and F1 Score for the proposed approach.

and

Table 13. Overall accuracy classification result for stress detection in Maize La Selva.

Accuracy/ F1 Score	G1			G2
	SVM	RF	NN	SVM	RF	NN
Select Bands	80.0/79.8	58.3/57.4	86.7/86.5	75.4/74.5	47.5/46.6	75.4/74.4
Features	66.7/66.3	50.0/46.2	61.7/61.5	67.2/66.6	49.2/47.4	63.9/63.2
Region Bands and Best VI	80.0/79.8	78.3/77.0	80.0/79.8	80.3/80.6	55.7/54.2	85.2/85.0
NDVI	35.0/47.5	35.0/34.2	46.7/45.3	31.1/34.7	23.0/23.7	32.8/30.0
All Bands	78.3/78.2	48.3/47.3	80/79.6	82.0/81.7	37.7/34.0	78.7/79.0
	G3			G4
	SVM	RF	NN	SVM	RF	NN
Select Bands	88.3/88.2	60.0/60.2	83.3/83.3	69.4/69.4	37.1/36.4	67.7/67.9
Features	73.3/73.2	48.3/49.5	80.0/80.3	54.8/49.3	38.7/38.3	59.7/59.3
Region Bands and Best VI	71.7/71.5	60.0/58.5	80.0/80.1	72.6/70.5	43.5/39.7	72.6/71.8
NDVI	53.3/61.2	45.0/62.9	43.3/43.4	30.6/29.8	27.4/25.6	30.6/30.0
All Bands	83.3/82.7	45.0/45.1	90.0/89.9	82.3/82.2	41.9/42.3	79.0/79.1
	G5			G6
	SVM	RF	NN	SVM	RF	NN
Select Bands	80.0/80.0	53.3/53.0	80.0/79.8	82.0/81.2	57.4/54.9	82.0/81.8
Features	75.0/74.5	61.7/61.6	70.0/69.3	60.7/59.7	45.9/45.3	70.5/69.9
Region Bands and Best VI	68.3/68.7	50.0/50.0	71.7/71.9	73.8/72.6	39.3/39.9	70.5/69.9
NDVI	35.0/46.3	26.7/26.8	31.7/29.5	52.5/51.5	44.3/42.5	36.1/38.5
All Bands	86.7/86.7	60.0/59.7	86.7/86.6	90.2/90.2	55.7/54.1	83.6/83.7
	G7			G8
	SVM	RF	NN	SVM	RF	NN
Select Bands	75.4/75.3	32.8/30.2	65.6/65.3	74.6/74.5	52.5/52.5	72.9/72.8
Features	73.8/73.4	59.0/58.4	63.9/63.9	66.1/65.8	61.0/60.0	64.4/64.3
Region Bands and Best VI	85.2/85.1	54.1/52.9	80.3/80.5	78.0/78.3	57.6/57.7	89.8/89.8
NDVI	32.8/44.1	39.3/44.9	36.1/32.9	49.2/52.3	57.6/74.3	64.4/56.9
All Bands	83.6/83.7	45.9/46.5	83.6/83.5	84.7/84.7	50.8/50.2	93.2/93.0
	G9			G10
	SVM	RF	NN	SVM	RF	NN
Select Bands	60.7/60.4	54.1/53.2	72.1/72.0	91.8/91.8	52.5/52.5	90.2/90.1
Features	63.9/63.8	57.4/54.6	59.0/59.1	77.0/76.8	57.4/57.3	85.2/85.3
Region Bands and Best VI	77.1/77.2	63.9/59.3	83.6/83.4	82.0/81.7	59.0/58.6	80.3/80.5
NDVI	36.1/38.1	34.4/33.1	41.0/38.1	39.3/39.7	37.7/42.3	37.7/49.4
All Bands	77.0/77.2	49.2/48.6	73.8/74.0	91.8/92.0	66.1/73.3	90.2/90.2

The values in bold indicate the best accuracy and F1 Score for the proposed approach.

present the overall accuracy and F1 Score obtained for each classifier, crop, and genotype. We include the comparison with the NDVI because this is the most frequently used index in the literature. In the case of Maize Turipana and Bean Turipana, the classifiers were used as detectors since these had only two levels of nitrogen and phosphorus treatments. However, Maize La Selva and Avocado La Selva had four nitrogen and five water treatments, respectively; the classifiers were then used for stress quantification.

Table 9 presents the overall accuracy classification and mean F1 Score for the Maize Turipana dataset; the results are presented by genotype. Figure 4 shows a graphical comparison. The best result for genotype V117 was obtained by the NN classifier using all bands, i.e., without including feature selection, with an up to 69.9% overall accuracy and an 82.5% F1 Score. Similarly, the best result for V121 was the NN classifier with an up to 68.6% accuracy and an 81.8% F1 Score. This result was expected since it takes into account the entire spectrum. For both genotypes, the worst performance was obtained using the NDVI index. Despite the accuracies and F1 Score obtained with the reduced spaces being lower than the full spectra, these are consistent between genotypes. There are interesting results, such as those obtained by the SVM using only the relevant bands for V117 (67.7% overall accuracy and 80.3% F1 Score) and using the selected features (Table 7) for V121 (64.7% overall accuracy and 77.9% F1 Score). It can be noted in Figure 4 that the selected features obtained an accuracy of over 60% for the two genotypes and all classifiers.

Table 10 summarizes the overall accuracy classification for the Bean Turipana dataset, and Figure 5 shows the graphical comparison. For the Mungo genotype, the higher accuracy was 96.3%, and the F1 Score of 98.1% was obtained using the SVM classifier and the region bands in Table 4 combined with the best VI; also, for this representation space, the NN obtained a 96.1% overall accuracy and a 98.0% F1 Score. The best result for the Caupí genotype was obtained using NN and the full spectra (98.7% overall accuracy and 99.4% F1 Score); however, the SVM classifier and the region bands combined with the best VI achieved a 97.4% accuracy and a 98.7% F1 Score. Once again, NDVI presents a lower performance. Figure 5 shows that the select bands, features, and region bands obtained consistent results when SVM or NN was used.

Classification of the Avocado La Selva and Maize La Selva datasets was more challenging since these include more than two classes. For Avocado La Selva, Table 11 presents the overall accuracies for the three classifiers that quantify the stress level. In the case of genotype ANNS88, the best result was 76.5% accuracy (86.8% F1 Score) using NN and the select bands; for genotype ANGUI52, the best result was 83.5% obtained by NN using the complete spectra. Again, the worst performance was obtained using either NDVI or another VI. Figure 6 shows that, for this case, the NN classifier obtained the best performance when selected bands or the full spectra were used. Considering the experimental setup, stress detection was also performed using only the extreme treatments (T1 and T4), presented in Table 12 and Figure 6. Overall accuracies and F1 Score increase because only two classes are considered. For this case, the best result for genotype ANNS88 was obtained using the full spectra (80.1% accuracy and 89.0% F1 Score). Still, a close result was achieved using NN and the relevant regions combined with the best VI (78.1% accuracy and 87.8% F1 Score). Similarly, for genotype ANGUI51, the best result was obtained using full spectra and NN (90.8% accuracy and 95.1% F1 Score), but, using the selected features, we achieved an 82.0% accuracy and an 89.9% F1 Score also using NN. Figure 7 shows how stress detection is better than quantification, which is expected, given that more classes exist. Again, the NN obtains the best results for avocado.

Finally, Table 13 presents the results for the 10 genotypes of Maize La Selva. The best results are obtained by SVM or NN, depending on the genotype. The NN classifier obtains the best result for G1, G2, G8, and G9. The SVM classifier obtains the best result for G3, G7, and G10; both obtain the same result for G4, G5, and G6. Figure 8 shows two of the best results obtained from G3 and G10, showing the excellent performance of both SVM and NN; as in the previous results, RF does not present significant results, nor does NDVI.

4. Discussion

The results obtained with the four crops suggest that working directly with the full spectra is better. Despite the several configurations tested, the higher overall accuracies were obtained using the spectra libraries without processing. This result is expected. However, this entails high data acquisition costs and a higher computational processing load. As pointed out in [24], there is a general tendency to perform studies based on spectral data where the choice of equipment is limited by its affordability. The higher the spectral resolution required, the more expensive the equipment [30]. Thus, using this equipment in real crop monitoring applications will be more challenging. This creates the need to develop methodologies to establish wavelengths or regions of the spectrum relevant to the problem. In addition, using the complete spectrum implies a larger computational complexity. For instance, each spectrum in this study had 1485 spectral bands (once the samples were cut out between 450 to 900 nm), representing a high dimensional space for machine learning.

This study identified regions of interest along the electromagnetic spectrum in the visible to near-infrared range. As presented in Section 3.1 and Section 3.2, the regions that provide the most relevant information for stress detection are related to features such as chlorophyll absorption, the red edge, and mainly the near-infrared region. As shown in previous work [33,34], the quantification of nitrogen and its effects on different crops can be performed using near-infrared or higher bands. However, these bands also imply high-cost equipment, which increases as it moves away from the visible region [35]. Therefore, the results obtained in the spectral characterization are relevant since they can be used to develop cost-effective monitoring systems rather than high-resolution spectrometers, focusing the acquisitions on bands or regions according to the type of crop [36]. The consistency in the results of the spectral characteristics obtained both analytically from the signatures (Table 4) and automatically through machine learning methods (Table 5 and Table 6) provides evidence of the potential of spectral data for monitoring different crops, even when they are under various conditions [37]. For instance, the band ranges obtained for maize under nitrogen deficiency were consistent with the findings of [38], who established that the 706 to 721 nm and 760 to 1142 nm ranges are more sensitive to nitrogen content at advanced stages of plant growth. In the case of beans, the results show consistency with the work of [39], demonstrating a high correlation between the amount of potassium and bands in the visible range, such as 410, 550, and 690 nm.

The results obtained in this work lead to the analysis and reconsideration of the extensive use of vegetation indices such as proposed by [14] with NDVI. This is because the information provided by the entire spectrum or specific representative bands and regions can be used with better results. This work demonstrates the effectiveness of these approaches compared to VI. A few indices were selected for the feature selection (Table 7 and Table 8). Similarly, it is observed that NDVI, according to the automatic feature selection methods, does not provide sufficient discriminant information for abiotic stress levels. Indices such as CCCI, CIG, and NDRE were selected more frequently across crops and genotypes than NDVI. Previous works have shown correlations with R² greater than 0.5 among VI and nitrogen content in maize [38,40] or phosphorous content in beans [39]. However, it should be noted that our approach assumes the problem of stress detection, i.e., seeking from the spectral data to determine whether or not there is stress (categorical or classification problem). Hence, metrics such as precision and F1 Score evidenced poor VI performance. In addition, previous works have demonstrated the dependence of the scales of VI on the plant status but also on the sensing instrument (e.g., MultiSpec, spectrometers, satellite images) [14,41]. Thus, generalizing VI for various crops, genotypes, and stresses becomes more difficult. However, as demonstrated in this study, finding band ranges to monitor various crops is possible.

The results demonstrated the ability of SVM and NN classifiers to detect and quantify stress. Good results were not obtained using RF with the proposed methodology. For stress detection, in the case of Maize Turipana and Bean Turipana, the best results from the proposed method were obtained using SVM. The best result was obtained from NN for Avocado Rionegro (using only two treatments). On the other hand, NN achieved the best results for stress quantification (Avocado Rionegro and Maize Rionegro). Unlike parametric methods, machine learning-based approaches have a significant potential for stress detection and quantification. However, challenges associated with these approaches include the limited availability of data. The accuracy of the models is contingent on the amount of data for training and, in specific applications related to agriculture and crop monitoring, obtaining a substantial amount of data is challenging due to the extensive manual effort required, mainly outdoors. On the other hand, it is evident from the results obtained in this study that the final accuracy of the models depends not only on the type of crop but also on the genotypes. Particular attention should be directed toward crops with a rootstock in their structure, such as avocado, where these can notably impact the genetic [42] and physiological behavior [43,44,45,46] of a common scion or canopy. Therefore, it is challenging to generalize crop monitoring processes using spectrometry, as previously recognized by [37].

5. Conclusions

This study demonstrated the capabilities of supervised machine learning algorithms for detecting and quantifying abiotic stress in three crops: maize, common beans, and avocado. Three classification algorithms were compared: SVM, RF, and NN. The results with RF were not satisfactory. On the other hand, SVM and NN allowed the detection or quantification of stresses in the different crops and genotypes studied. Although the best results were obtained with the full spectrum, it was shown that obtaining a set of relevant spectral bands to evidence the effects of abiotic stress is possible. The problem of quantifying the stress level was more complex, as lower accuracies were obtained compared to detection in the avocado experiment. To improve these results, increasing the amount of training data is necessary. However, their acquisition is quite complex because it is a manual process, and the high variability in plants and environmental conditions leads to the discarding of signatures in the preprocessing.

In addition, a characterization of the spectral signatures of the crops under study was presented, identifying regions of the spectrum relevant to the study of the different types of stress. Spectral features were identified in the visible, red edge, and near-infrared regions, obtaining consistent results compared with an unsupervised band selection approach. In the different crops, the near-infrared region was one of the most relevant for stress detection.

Future work includes using the relevant regions obtained to develop low-cost devices and seek to implement this technology in productive units. On the other hand, the methodology used in this study will serve as a basis for studying different types of stresses (biotic and abiotic) in crops.