The cultivation of broccoli (Brassica oleracea var. italica) has become increasingly attractive and profitable. It is highly regarded for its nutritional value and also its antioxidant and anticancer properties (Owis, 2015). In Spain, the production of broccoli has increased exponentially in the last decades, and it is expected to continue rising in future. In 2018, broccoli crop yields reached up to 561,000 tons in Spain, and most of them were exported to European countries (latest available data reported by Ministerio de Agricultura, Pesca y Alimentación, www.mapa.gob.es).

Pests and plant diseases are a great challenge in modern agriculture and the main cause of production and economic losses in agriculture worldwide (Savary et al., 2012). Current practices and social activities, such as intensified monoculture in large areas, the use of genetically uniform plant varieties, and international trading of agricultural commodities, contribute largely to the widespread of plant disease epidemics and rapid pathogen evolution (Zhan et al., 2015). Like other Brassica crops, broccoli plants are susceptible to infection by fungi (Alternaria leaf spot, anthracnose, blackleg, or mildews), some viruses (virus mosaic), and bacteria (black rot, soft rots, bacterial leaf spots). Among the bacterial pathogens, Xanthomonas campestris is one of the most important in brassicas (Mansfield et al., 2012; Ekman et al., 2014; Dep. Primary Industries and Regional Development, Government Western Australia 2018, www.agric.wa.gov.au/broccoli/diseases-vegetable-brassicas). The most notable pathovar of X. campestris is campestris (Xcc), which is the causal agent of black rot of crucifers and affects all cultivated brassicas. Indeed, and according to the report elaborated by EIP-AGRI Focus Group for the European Commission, Xcc is a threat to the production of broccoli, cauliflower, and cabbage throughout Europe (2016, Integrated Pest Management for Brassica, https://ec.europa.eu/eip/agriculture/en/publications/eip-agri-focus-group-ipm-brassica-final-report). Moreover, Xcc can be subdivided into nine races on the basis of the responses they induce on different cultivars. Among the nine races described for Xcc, races 1 and 4 are considered the most virulent and spread worldwide (Fargier and Manceau, 2007; Tortosa et al., 2018).

Precision agriculture demands the development of imaging sensor-based methods of detection and diagnosis of plant stress, including diseases. Several optical sensors are currently implemented to monitor crop fields (Aasen et al., 2019; Gerhards et al., 2019; Maes and Steppe, 2019; Pérez-Bueno et al., 2019a; Kashyap and Kumar, 2021; Pineda et al., 2021). Their applicability at lab scale and in high-throughput platforms by proximal sensing, and in the field by remote sensing, makes them particularly useful. However, the data provided by imaging sensors are large and complex and, consequently, difficult to interpret. Hence, improving our ability to extract useful information from these vast datasets requires the use of machine learning algorithms (Sperschneider, 2020). Machine learning is a subset of artificial intelligence (AI), consisting of algorithms that are able to learn patterns from a database of known samples and, based on those patterns, identify or categorize new samples. In agriculture, these algorithms can assist in the monitoring and decision-making processes of crop management (Chlingaryan et al., 2018; Golhani et al., 2018; Liakos et al., 2018; Gao et al., 2020). Thus, the implementation of imaging sensors and AI is a pivotal tool for crop management based on digital agriculture (Talaviya et al., 2020; Jung et al., 2021). However, alteration in growth conditions due to climate change imposes an additional challenge to plant disease detection methods based on AI. The expected rises in CO₂ concentration and temperature associated to climate change would have an impact on agriculture, affecting plants and pathogen physiology (Trivedi et al., 2022) and their geographical distribution (Aidoo et al., 2021). For that reason, potential future climate should be considered as an experimental variable to develop more robust detection methods.

Thermography and multi/hyperspectral reflectance imaging are the most common sensors applied in agriculture (Zarco-Tejada et al., 2018; Maes and Steppe, 2019; Pérez-Bueno et al., 2019b). On the one hand, canopy to air differential temperature (T_C-T_A) is an indirect measurement of the vegetation transpiration rate (Scarth et al., 1948; Milthorpe and Spencer, 1957; Fuchs and Tanner, 1966) and is widely used in proximal and remote sensing for stress detection, as recently reviewed by Pineda et al. (2021). On the other hand, the high spectral resolution of hyperspectral reflectance imaging allows the creation of a growing collection of vegetation indices (VIs). These VIs are transformations of two or more spectral bands which allow reliable temporal and spatial inter-comparisons of vegetation attributes. Thus, VIs are quite simple and effective parameters to quantitatively and qualitatively evaluate vegetation traits such as vigor, fitness, and pigment composition, among other applications (Huete et al., 2002).

In fact, many VIs can be found in the literature. One of the most widely used is the normalized difference vegetation index (NDVI), which is related to vitality of canopies (Tucker, 1979; Pettorelli, 2013). Other VIs correlate with a wide range of plant physiological traits. For example, the photochemical reflectance index (PRI) correlates with photosynthetic activity (Gamon et al., 1992); the carotenoid reflectance index (CRI) (Gitelson et al., 2002) and the anthocyanin reflectance index (ARI) (Gitelson et al., 2001) are related to pigment contents; and the water balance index (WBI) is connected to water content in leaves (Peñuelas et al., 1993). Indeed, recent works have implemented VIs to the study of plants infected by Xanthomonas spp. Abdulridha et al. (2019) used a collection of VIs (ARI and NDVI among them) implemented on classifying algorithms to successfully identify tangerine plants infected by X. citri pv. citri. Similarly, NDVI values correlated well with the extension of the lesions caused by X. campestris pv. oryzae on rice leaves (Zhang et al., 2022). Moreover, several works have compared the association between climate change and the interannual variability registered on NDVI in several locations around the world (Kalisa et al., 2019; Bagherzadeh et al., 2020; Zhao et al., 2021). Nonetheless, new reflectance parameters could be defined for a given purpose to maximize differences when standard VIs are not sensitive enough (Miao et al., 2007; Mahlein et al., 2013; Zhang et al., 2017; El-Hendawy et al., 2019; Jia et al., 2019; Yuan et al., 2019).

The main aim of this work was to develop an efficient method for the detection of Xcc infection in broccoli plants based on thermal and hyperspectral reflectance imaging on individual leaves. For this purpose, six parameters were recorded: leaf temperature (particularly T_C-T_A) and five already known VIs. Moreover, three novel VIs specifically designed for detecting the Xcc infection were extracted from the reflectance spectra of healthy and diseased broccoli leaves. They were named diseased broccoli indices (DBI₁-DBI₃). This set of nine parameters were implemented on a selection of algorithms widely used on precision agriculture for their success in classifying infected plants: the multilayer perceptron-based artificial neural network (MLP), the support vector machine (SVM), and the k-nearest neighbor (kNN). Finally, the suitability of the trained models was evaluated by comparing their performances in correctly classifying new samples as healthy or diseased leaves under a range of climate conditions, including intermediate and extreme climate change scenarios, as well as current climate conditions. Furthermore, the relevance of every input parameter for the detection of Xcc infection in broccoli plants was evaluated.

Imaging techniques appear to be essential for precision agriculture due to their fast time-spatial response to biotic and abiotic stress in a non-destructive manner (Barón et al., 2016; Mahlein, 2016). In the last years, thermal and (multi- or hyper-) reflectance imaging sensors have been broadly used for monitoring stress in crop fields. Furthermore, sustainable agriculture is increasingly relying on AI (such as classifying algorithms) coupled with computer vision, to solve farming issues and to promote the automation of decision-making process (Tian et al., 2020; Nabwire et al., 2021). However, these methods require basic research to define informative parameters that efficiently report the health state and fitness of a particular crop. This work aims to define optimal VIs and classifiers for the detection of Xcc-infected broccoli leaves. Furthermore, the robustness of those models was analyzed under climate conditions mimicking those expected for Region of Murcia in years 2081–2100.

According to Fargier and Manceau (2007), when a cultivar of a brassica is infected by Xcc, a collection of polymorphisms is obtained depending on the race inoculated. In the case of broccoli plants used for this study (B. oleracea var. italic cv. calabrese natalino), the symptoms developed under CCC triggered by races 1 and 4 were similar. However, Xcc race 1 showed a slower timecourse than race 4 under the same ambient conditions. The process of undergoing climate change could affect considerably plant biochemistry and therefore plant defense responses. For this reason, it is relevant to include future climate conditions in disease detection studies. Indeed, each disease may respond differently to these variations, and thus, climate change would cause neutral, positive, or negative effects on plant responses to diseases (Trebicki et al., 2017; Velásquez et al., 2018; Cheng et al., 2019). According to the results reported here, symptoms caused by Xcc on susceptible broccoli plants would not be altered on an intermediate climate change scenario like RCP 4.5. These results are in accordance with previous works, as extensively reviewed by Gullino et al. (2018). However, the RCP 8.5 imposed a stress condition limiting or slowing down plant growth. It will be of particular interest to gain knowledge about the impact of combined high temperature and high CO₂ on photosynthetic processes of the broccoli plant which, as a C₃ species, is well adapted to mild temperatures. Nevertheless, further research would be needed to fully understand the impact of climate change on broccoli plant physiology (particularly on photosynthesis), as well as on the physiology of Xcc races, and/or their interaction with host plants.

In literature, a number of classic VIs derived from multispectral (and hyperspectral) imaging can be found. This VIs can be used to detect, classify, and quantify specific diseases with varying degrees of success (Lowe et al., 2017). However, high-resolution spectra recorded with hyperspectral imaging sensors allow the selection of an optimized set of wavelengths to maximize differences between healthy and infected plants. Those wavelengths can be used to create novel parameters specific for a given host–pathogen system or stress factor. This approach has already been demonstrated to be suitable to detect diseased plants when combined with AI algorithms. Thus, Mahlein et al. (2013) reported specific spectral disease indices for the detection of sugar beet plants infected with Cercospora leaf spot, sugar beet rust, or powdery mildew. Those indices improved disease detection and identification when implemented on classifiers. Similarly, Yuan et al. (2019) proposed a novel method for detecting anthracnose in tea plants based on hyperspectral imaging that included two new disease indices in the classificatory models. Moreover, the analysis of reflectance spectral data of healthy and diseased wheat ears allowed the creation of a novel index that demonstrated a stronger ability to determine the severity of the Fusarium head blight compared with other sixteen existing spectral indices (Zhang et al., 2020).

In this work, three novel VIs have been developed to successfully distinguish between healthy and Xcc-infected broccoli plants (Table 4). Leaf reflectance is a complex phenomenon dependent on biochemical and biophysical properties of the canopy leaves, which in turn are affected by growth conditions and diseases. Thus, the visible reflectance range (400–700 nm) is mostly influenced by the leaf pigment content; the reflectance in the near-infrared range (700–1100 nm) depends on water content and leaf structure, or internal scattering processes; and the short-wave infrared (1,100–2,500 nm) is influenced by the composition of leaf chemicals and water, as reviewed by Mahlein (2016). Since chlorophylls are the main pigments influencing reflectance spectrum at 400 and 690 nm, both DBI₁ and DBI₂ indices could be indicative of the severity of chlorosis. Moreover, DBI₃ could be also related to the contents on chlorophylls and carotenes (Carter and Knapp, 2001).

DBI₁, DBI₂, and DBI₃, together with thermal (T_C-T_A) and other common hyperspectral reflectance parameters (NDVI, PRI, ARI, CRI, and WBI), were implemented in three different supervised classifiers (MLP, SVM, and kNN) for each experimental condition. Since the learning process of each algorithm differs from each other, so will the quality of its predictions on the new samples. In this sense, it is a common procedure to compare the performance of several algorithms when sorting new samples (the validation datasets) after training on the same dataset. Metrics such as specificity (true negative rate), sensitivity (true positive rate), accuracy (percentage of right guesses), or F-measure (harmonic mean of precision and sensitivity) evaluate the performance of classification, that is, the estimation of the true risk of error of the output prediction of a machine learning algorithm (Shalev-Shwartz and Ben-David, 2014; Liakos et al., 2018). The MLPs were the most effective classifier, with the highest overall accuracy and F-measure under the three growing conditions assayed. The performance of the MLP models was comparable to that reported by other authors for disease detection classifiers. Indeed, an increasing number of studies apply classifiers to spectral data (including or not thermal parameters) to identify infected plants at conditions resembling CCC. For example, Sankaran et al. (2013) reported an accuracy of 87% when classifying citrus trees infected by Candidatus Liberibacter spp, a bacteria causing Huanglongbing disease. Zarco-Tejada et al. (2018) obtained accuracies of disease detection exceeding 80% when classifying Xylella fastidiosa-infected olive trees. This pathogen, alike Xcc, is a xylem bacterium. Abdulridha et al. (2020a,b) identified tomato plants infected with tomato yellow leaf curl virus, Xanthomonas perforans, or Corynespora cassiicola (a fungus) with 94–100% accuracy depending on the pathogen. Pérez-Bueno et al. (2019b) detected avocado trees suffering white root rot (caused by the fungus Rosellinia necatrix) with accuracies up to 82.5%. Nguyen et al. (2021) achieved accuracies ranging from 82 to 96.75% when identifying vines affected by the Grapevine vein-clearing virus. Similarly to the results here reported, Yuan et al. (2019) designed novel hyperspectral reflectance indices which help to identify Gloeosporium theae-sinesis Miyake-infected tea plants with 94 and 98% accuracies at pixel and leaf levels, respectively.

The performance of the models was affected differentially by growing conditions, depending on the classifier. In both RCPs, the accuracy of the classifiers decreased in all cases. However, the accuracy of models based on MLP only decreased from 88.1% at CCC to 76.9 and 83.3% for RCP 4.5 and RCP 8.5, respectively. This advantage of MLPs against SVM and kNN models could be related to the fact that MLPs are less affected by noise factors (compared to other algorithms), which in turn reduces significantly the influence of the unknown variability. Therefore, MLPs are usually more robust models that often outperform other classifying algorithms in solving a variety of classification problems (Basheer and Hajmeer, 2000; Bala and Kumar, 2017). To our best knowledge, this is the first time that machine learning classifiers have been applied to hyperspectral and thermal data taken under climate conditions mimicking those projected for the future in order to classify healthy and infected plants.

DBI₁, DBI₂, and DBI₃ proved to be important features for plant classification according to a surrogate RF used for testing the global importance of variables. Overall, DBI₁ and DBI₂ were the most informative parameters of the set for CCC. However, the global variable importance varied for each parameter depending on the climate conditions. Under RCP 4.5 conditions, T_C-T_A and DBI₃ were the most informative parameters. In contrast, DBI₂ was the parameter with the highest global variable importance at RCP 8.5. This effect may be due to the impact of growing conditions on the symptomatology (degree of chlorosis and/or necrosis) of the infections, as discussed above.

The parameters DBI₁, DBI₂, and DBI₃ here presented are good reporters for Xcc infection in broccoli leaves. Furthermore, the model based on MLP and the set of parameters DBI₁, DBI₂, and DBI₃ along with common VIs (ARI, CRI, NDVI, PRI, and WBI) and T_C-T_A would be an effective procedure for the identification of Xcc-broccoli infected plants. In addition, this model proved to be robust regardless of the climate conditions.

The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding authors.

MP, MLP-B, and MB conceived and designed the experiments and wrote the manuscript. MP and MP-B conducted the experiments. MP analyzed the data, interpreted the results, and mounted images. MB contributed to the materials, equipment, and analysis tools. All the authors reviewed it and approved the final version.

This work was supported by grant RTI2018-094652-B-I00 funded by MCIN/AEI/10.13039/501100011033 and by “ERDF: A way of making Europe”. The free open access publication was partially funded by Consejo Superior de Investigaciones Científicas (CSIC) through the Unidad de Recursos de Información Científica para la Investigación (URICI).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.