Processing math: 100%
ARTICLE   Open Access    

Identification and content analysis of volatile components in 100 cultivars of Chinese herbaceous peony

  • # Authors contributed equally: Aixin Wang, Yasang Luo

More Information
  • Herbaceous peony (Paeonia lactiflora Pall.) is a well-known and traditional flower in China, occupying a significant position in Chinese traditional culture. The floral scent of the herbaceous peony however remains relatively understudied. The objective of this study was to investigate the floral composition of herbaceous peony by collecting and identifying floral volatiles from 100 cultivars, including P. lactiflora 'Hangbaishao', P. lactiflora 'Hongrongqiu', P. lactiflora 'Biandihong', P. lactiflora 'Zijin Daipao', P. lactiflora 'Zixia Yingxue', and P. lactiflora 'Fenchi Dicui'. The volatile compounds were collected using the dynamic headspace technique and identified through gas chromatography-mass spectrometry (GC-MS). The results demonstrated qualitative and quantitative variations in the floral fragrances emitted by the 100 cultivars, with a total of 16 volatiles belonging to six categories (six alkanes, three alcohols and esters, two terpenes, as well as one each of ether and phenol) being identified. However, it is notable that not all volatile categories were emitted by every cultivar. Moreover, while some compounds were present in all 100 herbaceous peony cultivars, others were exclusive to specific cultivars. The screening revealed that ten of the 16 identified flower volatile compounds exhibited unique floral components. It is noteworthy that benzene,1,4-dimethoxy-, was identified as the most prominent compound in several cultivars, including P. lactiflora 'Taohua Huancai', P. lactiflora 'Xishifen', P. lactiflora 'Dabanhong', P. lactiflora 'Fumantang', and P. lactiflora 'Zhushapan'. Furthermore, the clustering classification results demonstrated that benzene,1,4-dimethoxy-, exhibited the highest variable importance in projection (VIP) value of 3.153, as determined by partial least squares discriminant analysis (PLS-DA).
  • Cameras are ubiquitous and deployed in public areas, producing large quantities of video data. To get a better understanding of the dynamic traffic systems, vehicle re-identification (i.e., vehicle reID or v-reID) can be performed. Each camera in the road network captures an image of every passing vehicle and stores them in a gallery. The cameras can have overlapping or non-overlapping views. Given a query image, i.e., an image taken from a vehicle, v-reID aims to re-identify the vehicle by finding its occurrences from the gallery. Like any deep learning model, a lot of training data is required for v-reID models. Two of the widely used and publicly available datasets are VeRi-776[1] and VehicleID[2]. Although they are commonly used, to the best of our knowledge, there is no existing research that specifically comments on the content and limitations of such datasets.

    Artificial intelligence (AI) ethics refers to a framework of ethical principles and methods aimed at guiding the responsible creation and use of AI technologies[3]. There is no prior research on ethics regarding v-reID. We assume that the reason behind is that no humans are directly involved in the re-identification process. However, what about humans being indirectly involved? One major issue is that some faces of drivers and passengers in the dataset are visible. Because existing reID models are black boxes, it is unknown what happens behind the scenes. In addition, we notice other observations in the datasets that may decrease the performance of v-reID models. This paper uses CycleGAN, an image-to-image translator, to tackle these limitations.

    While the datasets in this paper have been extremely valuable for the v-reID research, they are not without their limitations. The primary purpose of this paper is to highlight the limitations within commonly used training datasets for v-reID and propose solutions to address these concerns. By identifying and addressing these limitations, we aim to promote safer research practices and raise awareness about the ethical implications of AI models involving humans. In this way, we can promote safer research, educate our readers, and build a more human-centered AI. This paper proposes the use of existing image-to-image translation models, such as CycleGAN, to mitigate the identified limitations. By employing CycleGAN, we aim to train it using two datasets: one containing the limitations present in current datasets and another without these limitations. Through this approach, we can generate images absent of the identified limitations. In general, our main contribution is threefold: (1) we discuss for the first time about the ethics involved in vehicle re-identification; (2) different limitations of existing and widely used datasets that can impede the performance of reID models are discussed, and (3) a method is developed to generate data that is anonymous and respects the privacy of drivers and passengers.

    To ensure that AI is reliable throughout its lifecycle, the Australian Government has proposed eight voluntary AI ethics principles[4]. If V-reID were to be deployed in public and used in multi-camera vehicle tracking and traffic surveillance[5], the following considerations, among others, would need to be prioritised:

    • Model: the deep learning models used for v-reID should be transparent for users and those that are impacted. Users should comprehend the hows and whys of the models, while the general public should receive a more accessible education on the AI systems (principle: transparency and explainability).

    • Data: the data collection and usage should guarantee the anonymity of drivers and passengers, i.e., their privacy should be protected and respected (principle: privacy protection and security). Furthermore, the v-reID model should enable diversity and should not perform any sort of unjustified surveillance (principle: human-centered values).

    To the best of our knowledge, limited studies have been done on re-identification in consideration of ethical issues. Dietlmeier et al.[6] anonymized datasets by blurring faces on person reID benchmarks and demonstrated that in doing so did not compromise the performance of person reID. To introduce privacy and security in person reID, Ahmad et al.[7] solved person reID using event cameras. The latter captures dynamic scenes by responding to brightness changes only without providing any RGB image content. Richardwebster et al.[8] used saliency maps to help differentiate between individuals that are visually similar. This helps to better understand the person reID model's decisions and helps to reduce false matches in high-stake reID, such as autonomous driving, criminal justice, and healthcare[9]. There is a need to set boundaries or regulations to ensure a safer practice of v-reID research.

    Methods using Deep Neural Networks (DNNs), such as Convolutional Neural Networks (CNNs)[1016], Recurrent Neural Network (RNNs)[17] and Transformers[1822], have been extensively used in v-reID. Although v-reID has resulted in improvements in performance, these deep learning models still remain black boxes. This means that the internal inference processes are either unknown or non-interpretable to us[23].

    On one hand, some research primarily aims to improve the performance of v-reID models using esoteric algorithms. However, the outcomes are not useful for real-world applications (or called open-world re-identification[24]). Some works have tackled the issue of rendering re-identification more explainable in the context of persons[25] and vehicles[19]. Chen et al.[25] proposed an Attribute-guided Metric Distillation (AMD) method that learns an interpreter that uses semantic attributes to explain the results of person reID methods. The interpreter is capable of quantifying the contributions of attributes so that users can know what attributes differentiate two people. Moreover, it can visualize attention maps of attributes to show what the most significant attributes are. Our previous work[19] aimed to render vehicle reID research more digestible. A step-by-step guide was proposed on how to train a reID model, as existing research papers are more complex.

    Conversely, people cannot entirely trust the results produced by these black-box models. A small alteration of an image that is undetectable for the human eye, can lead to a DNN being confused, thinking it is something completely different[26]. Moreover, even though v-reID does not aim to re-identify people directly, facial recognition can happen as a by-product if the training data involves images of humans. This means that data should guarantee the anonymity of people or that deep learning models should not perform any unknown behind-the-scenes facial recognition.

    With the growing introduction of publicly available vehicle large-scale datasets, v-reID has increased in popularity. Some publicly available v-reID datasets include: Comprehensive Cars (CompCars)[27], PASCAL VOC[28], PKU-VD1 and PKU-VD2[29], Vehicle-1M[30], CityFlow[31], VERI-Wild[32], PKU VehicleID[2], and VeRi-776[1]. In this paper, the spotlight is on the VehicleID and VeRi-776 datasets, due to their relevance as widely recognized benchmarks in the field of v-reID. Meanwhile, VehicleID is a large-scale dataset with controlled views, VeRi-776 provides real-world diversity.

    The VehicleID[2] dataset contains 221,763 images of 26,267 vehicle images captured by multiple real-world surveillance cameras in a small city in China, offering one of the largest available collections for v-reID. This dataset focuses primarily on front and rear viewpoints. VeRi-776[1] is a vehicle re-identification dataset which contains 49,357 images of 776 vehicles from 20 cameras. The dataset is collected in the real traffic scenario, also in China. In contrast to VehicleID, VeRi-776 presents a more complex and realistic dataset by including varying environmental conditions, such as different times of day and diverse camera angles. Samples of both datasets are shown in Fig. 1.

    Figure 1.  Sample images of different vehicle models from (a) the VehicleID and (b) VeRi-776 datasets.

    VehicleID and VeRi-776 are compared in terms of four categories: resolution, privacy, noise, and views.

    The images captured by the different cameras are in higher resolution for VehicleID compared to VeRi-776. This is important for v-reID as every local feature is taken into consideration when performing the re-identification task. The smallest detail, such as the size, shape, or color of the windshield sticker is crucial.

    Unfortunately, the high resolution has its flaws. In the case of VehicleID, the faces of drivers and passengers are visible. This is not privacy-compliant. Fig. 2 shows examples of VehicleID data where faces are clearly visible when zooming in. This is an issue if data were leaked and distributed, and the faces of people, including children, could be accessed. Furthermore, the AI could learn the faces instead of the vehicle features by performing facial recognition indirectly. In the worst scenario, the AI could be performing unjustified surveillance without our awareness.

    Figure 2.  The pros and cons of the VehicleID dataset: (a) visible faces, (b) noisy red time stamp, (c) only two directions. (The images were captured in high quality).

    Some images in VehicleID have red writing on them, see samples in Fig. 2. The writing indicates the timestamp and trip-related information. This adds additional noise to the reID model which can therefore impact its performance either by thinking that the red writings are essential to the reID process, or by being confused and deducing that it has nothing to do with vehicles anymore.

    Finally, VehicleID is captured by an unknown number of cameras from two views (e.g. front and back), while the images in VeRi-776 with the same identity have eight different appearances captured from 20 different camera views. A reID dataset should preferably contain vehicles from various perspectives.

    Image-to-image translation[33] involves learning a mapping function between two domains X and Y, to generate an image from one domain to another domain. Examples include converting images from greyscale to color, from synthetic objects to real objects, or photos to paintings. In v-reID, image-to-image translation becomes useful in two cases: (1) when training data is scarce and needs to be augmented, or (2) implementing a model that can generalize, i.e., a model that is trained on a source dataset X and then applied on a target detest Y, where X and Y are of different domains. This is also called domain adaptation (DA)[34]. The differences in domains can be due to illumination, view, or environment changes, or different camera networks, types, settings and resolutions.

    Let X be the source domain and Y the target domain. The goal is to construct a mapping function from X to Y, i.e., given an image xX, we want to function a mapping function that transforms x into yY. Traditionally, image-to-image translation techniques needed paired data. This means that for the same examples in X, the same examples in Y with the required modifications are provided, e.g. a dog from a certain position and a drawing of the same dog in the same position. An example of paired image-to-image translation is Pix2Pix[33]. While paired image-to-image translation produces great results, the requirement for training data is very limited and expensive to prepare, or even impossible depending on the field, e.g., v-reID. As a consequence, Zhu et al.[35] introduced CycleGAN which is an approach to unpaired image-to-image translation built on the Generative Adversarial Network (GAN)[36].

    Image-to-image translation techniques have been widely used in vehicle reID[3740]. Given an input view of a vehicle, Zhou & Shao[37] generated cross-view vehicle images. Wang et al.[38] and Luo et al.[40] augmented their data using image-to-image translation techniques, such as SPGAN[41] and CycleGAN[35], while Zhou et al.[39] employed a GAN-Siamese network to transform images from day-time domain to night-time domain, and vice versa. However, to the best of our knowledge, no prior work has tackled the safety issue of the existing datasets using image-to-image translation thus far.

    Existing works transfer image styles to augment their data or to adapt to the different domains. The above-mentioned shortcomings in terms of noise, views, and privacy for VehicleID can limit the performance of reID models. To this end, we propose a method to circumvent these issues related to the dataset. This is the very first attempt to tackle this topic.

    CycleGAN is an image-to-image translation technique that does not require paired examples. This means that the training data does not need to be related, i.e., CycleGAN requires a source set {xi}Ni=1 and a target set {yj}Mj=1 with no information on which xiX matches to which yjY. This flexibility makes CycleGAN particularly well-suited for our task that works with v-reID datasets that do not have a direct one-to-one relationship. Additionally, CycleGAN has been widely adopted in research due to its versatility in applications such as style transfer, domain adaptation, and data augmentation. Its open-source code is well-documented and easy to understand, making it accessible for researchers and developers to implement and customise.

    CycleGAN is constructed using GANs, and GAN is an unsupervised technique that is built on two models: a generator model and a discriminator model. As the name suggests, the generator generates outputs from the domain, and the discriminator receives synthetic data from the generator and the real dataset. The discriminator then determines whether its input is real or fake (generated). Both models are trained until the generator has learned to fool the discriminator and the discriminator is not able to distinguish real from generated data.

    CycleGAN involves training two GANs, i.e., training two generators and two discriminators simultaneously. Let X and Y be two domains with training samples {xi}Ni=1 (xiX) and {yj}Mj=1 (yjY). The two generators are G: X → Y and F: Y → X. Generator G tries to generate images G(x) that look similar to images from domain Y. Likewise, generator F aims to generate images F(y) that look most similar to images from domain X. The two discriminators are DX and DY , where DY aims to distinguish between images {x} in X and the generated images {F(y)}, and DX aims to discern between {y} and {G(x)}.

    The overall objective function involved in training consists of two adversarial losses[36] and a cycle consistency loss[42]. The adversarial losses are applied to both mapping functions G and F, and match the distribution of the generated images to the data distribution in the target domain:

    LGAN(G,DY,X,Y)=Eypdata(y)[log(DY(y))]+Expdata(x)[log(1DY(G(x)))] (1)
    LGAN(F,DX,Y,X)=Expdata(x)[log(DX(x))]+Eypdata(y)[log(1DX(G(y)))] (2)

    where, x~pdata(x) and y~pdata(y) denote the data distributions of X and Y, respectively. Generator G tries to minimize Eqn (1), while adversary DY aims to maximize it, i.e., minGmaxDYLGAN(G,DY,X,Y). Similarly, generator F tries to minimize Eqn (2), while the adversary DX aims to maximize it, i.e., minFmaxDXLGAN(F,DX,Y,X).

    The cycle consistency loss aims to reduce the space of possible mapping functions, and therefore to encourage G and F to be bijections. Zhu et al.[35] illustrated this property with an example in translation: if a sentence is translated from English to French, and the output is translated back to English from French, the result should be the original sentence[43]. It indicates that both mapping functions should be cycle-consistent, i.e., x → G(x) → F(G(x)) ≈ x (forward cycle consistency) and y → F(y) → G(F(y)) ≈ y (backward cycle consistency). The cycle consistency loss is then formulated as Eqn (3):

    Lcyc(G,F)=Expdata(x)[F(G(x))x1+Eypdata(y)G(F(y))y1] (3)

    The full objective function is:

    L(G,F,DX,DY)=LGAN(G,DY,X,Y)+LGAN(F,DX,Y,X)+λLcyc(G,F) (4)

    where, λ controls the relative importance of the two objectives. Finally, the aim is to solve for Eqn (4):

    G,F=argminG,FmaxDX,DYL(G,F,DX,DY) (5)

    Details of CycleGAN are available in Zhu et al.[35].

    In the context of our work and for the sake of visualization, Fig. 3 shows how CycleGAN can be employed to translate images from the synthetic domain to the real domain, and vice-versa. The two chosen datasets are unpaired, i.e., the images are captured at different locations and times. This means that we do not have the exact same correspondence in both datasets. CycleGAN consists of GAN1, which transfers photos from the real domain to the synthetic domain, and GAN2, transferring images from the synthetic domain to generate images from the real domain, whilst solving Eqn (5). We use CycleGAN to transfer images from VehicleID to VeRi-776 to tackle the issues related to privacy, noise, and views mentioned previously.

    Figure 3.  CycleGAN model architecture transferring between VeRi-776 (real domain) and VehicleX (synthetic domain).

    CycleGAN is trained on VeRi-776 and VehicleID, and employed to transfer images from VehicleID to VeRi-776 style. This section presents the implementation details and the results generated from experiments VCGAN-A and VCGAN-B.

    Two experiments have been run using the default settings of CycleGAN. Both were trained for 100 epochs with an initial learning rate of 0.0002, followed by 200 epochs with a linear learning rate decay. The default optimiser is Adam[44]. The input images were scaled to 256 × 256 pixels and each experiment was trained on three NVIDIA A100 80GB GPUs.

    The difference between the two experiments lies in the training size, as shown in Table 1. VCGAN-A is trained using the same number of data in both datasets, while VCGAN-B is trained using the entirety of both datasets. While a testing set is not required for the training of CycleGAN, it is beneficial for evaluating the model's performance on unseen data. In the next paragraphs, examples from the VehicleID test set will be used to assess the performance of the trained models. In the following paragraphs, VCGAN-A and VCGAN-B will also be referred to as 'A' and 'B' respectively for simplicity.

    Table 1.  Training data sizes for VCGAN-A and VCGAN-B.
    VehicleID VeRi-776
    VCGAN-A 37,778 37,778
    VCGAN-B 113,346 37,778
     | Show Table
    DownLoad: CSV

    The properties of datasets mentioned previously and how CycleGAN addresses them are described in this subsection, including four categories: resolution, privacy, noise, and views.

    The resolution is crucial in the v-reID process. The higher the visibility of the details, the more discriminant features can be extracted and can therefore aid the v-reID decision. As we train CycleGAN on transforming images from VehicleID to VeRi-776, which has a decreased resolution compared to VehicleID, it is only logical that the images output by CycleGAN also present a lower resolution. This could be problematic depending on what features get lost. If we do not want to perform any license plate or facial recognition, losing features related to the license plates, faces, or camera descriptions, becomes an advantage. However, if the lost features include logos or windshield stickers, then this can influence the performance of the v-reID model.

    Privacy is an important factor that needs to be addressed. Even though the reID model is a black box, we need to ensure that no facial recognition is unintentionally conducted. As shown in Fig. 4 (left), we notice that in experiments A and B, the faces are blurred out, such that they are not visible anymore. While the outputs of A are satisfactory, the transformed images of B appear much smoother (look at the windshield, where there are fewer 'white strokes'). CycleGAN can erase the faces from the windshields.

    Figure 4.  Left: faces are blurred out, such that the privacy of the driver is ensured; Right: the timestamps in red are (partially) erased.

    The model might incorporate the red timestamps in its re-identification decision. After all, these timestamps are more noise than anything else. One could circumvent this by cropping the image, such that the timestamps are not visible anymore. While this would be possible for some samples, e.g., Fig. 4a (right), this wouldn't work for others, e.g., Fig. 4bd (right), as a chunk of the vehicle would then be cut out. The results look promising when transforming the images using the trained CycleGAN. The outputs in Fig. 4 (right) for A and B show that the red writing is completely (Fig. 4ac right), or partially (Fig. 4d right) erased.

    Finally, the cars in VehicleID are captured from two views (back and front), while VeRi-776 has more. This characteristic is visible in Fig. 5, where given a vehicle input from the front, CycleGAN generates vehicles from different views: side or back.

    Figure 5.  Generated views by CycleGAN.

    CycleGAN managed to erase or blur out the faces of people as well as to remove the noisy red timestamps from the images. This is beneficial as this renders the dataset anonymous as well as less distractive. On the other hand, the resulting images are of lower resolution, and potential discriminative features get lost.

    Besides our observations, we summarize additional outcomes that can benefit the research. These outcomes were unexpected and are worthy of mention. It is worth mentioning that this is not an exhaustive list, and there might be other elements we failed to notice.

    One of the unexpected outcomes is that the background, such as the lane or the pedestrian crossing markings, were removed, as shown in Fig. 6 (left). This can aid the v-reID model in focusing only on the vehicle rather than the background, as the latter is useless for reID.

    Figure 6.  Left: background removal; Right: cropped vehicles still look like vehicles.

    Surprisingly, CycleGAN handled cropped vehicles well. As shown in Fig. 6 (right), it can be observed that even though the vehicles were partially captured due to their position or camera glitch, CycleGAN managed to transform the vehicle only.

    CycleGAN saw certain objects in specific types of images that we did not perceive with our eyes, as shown in Fig. 7. Some vehicles in red that were captured from a specific frontal view were transformed into a taxi (Fig. 7a), highlighting one potential bias of the dataset. Fig. 8a shows four images taken of different taxis. We don't know the ratio of taxis and red vehicles. However, given the confusion by CycleGAN, the ratio must be on the higher end. VeRi-776 doesn't have many images captured during the nighttime. Hence, CycleGAN misinterprets images that are taken at night. Depending on the image, two scenarios result: a bush or an abstract image of the vehicle. On the one hand, when there are some variations of yellow or green on the bottom, CycleGAN detects bushes (Fig. 7b). As we notice, both VCGAN-A and VCGAN-B transformed the rear bumper into a bush. This transformation is due to the amount of certain camera views in VeRi-776 where the vehicles are hidden behind bushes, as shown in Fig. 8b, where we notice the bushes on the bottom of the images. On the other hand, when these yellow patches are absent, CycleGAN struggles and produces an output that does not make much sense (Fig. 7c). Finally, yellow street markings are transformed by CycleGAN into fences. Referring to Fig. 8a, we can deduce that this behavior is due to some camera views that are positioned such that these fences are visible. Interestingly enough, CycleGAN added something we did not expect: a license plate, see Fig. 4 (right (c)), Fig. 5, and Fig. 6. This was unexpected, but could mean that there is a large amount of data in VeRi-776 where the license plates have not been blackened out completely. Upon checking, we can verify this observation. Fig. 8b clearly shows that some vehicles in VeRi-776 do not have their license plates blackened out.

    Figure 7.  I spy with my convoluted eyes: (a) a taxi, (b) a bush, (c) an abstract self portrait, (d) a fence.
    Figure 8.  VeRi-776: samples of (a) taxis and yellow fence, and (b) vehicles captured behind bushes and non-blackened out license plates.

    CycleGAN transformed yellow-green patches into bushes, yellow markings into fences, added license plates, and transformed red vehicles into taxis. When the images are taken at night, CycleGAN produces outputs that do not make much sense. Rather than being useful for v-reID, these observations show that the source dataset has its drawbacks, notably a lack of images taken during the night, occluded vehicles by bushes, and most importantly, license plates that were not removed or hidden. This also means that we could go the opposite way, and generate vehicle images without license plates or images in a night setting.

    Multiple potential solutions can be explored to address these unexpected outcomes that could negatively impact the quality of the generated dataset. One potential solution is to ensure that the datasets are more diverse; however, this approach goes back to data collection, which can be more challenging and time-consuming. Another potential solution is to apply preprocessing steps to remove or obscure license plates before training, which would prevent the model from adding them during transformations. Finally, domain adaptation techniques, where transformations are more closely aligned with the target domain could be employed to better control the outputs generated by CycleGAN.

    In this subsection, we compare VCGAN-A and VCGAN-B quantitatively and qualitatively.

    There is a trade-off between smoothness and details. Images generated by VCGAN-A fits its purpose, however VCGAN-B does a better job overall. Based on all the previous figures, we deduce that VCGAN-B does a cleaner job in the sense that the outputs turn out to be smoother and resemble less to synthetic images than of VCGAN-A. Furthermore, referring to Fig. 9, VCGAN-B can produce 'whole' vehicles, while VCGAN-A outputs a deformed vehicle.

    Figure 9.  Comparing the two experiments: the outcomes of B are qualitatively more reliable than A.

    To compare both experiments numerically, we plot the generator, discriminator, and forward cycle consistency losses (Fig. 10). While the discriminator and cycle-consistency losses are decreasing by each epoch, the generator loss increases from epoch 120 on. This indicates that the training of the generator and discriminator is imbalanced, with the discriminator becoming too strong. Compared to VCGAN-A, the discriminator loss and forward cycle consistency loss are lower for VCGAN-B. However, the generator loss shows the opposite behavior.

    Figure 10.  Generator loss, discriminator loss and forward cycle consistency loss curves per epoch for VCGAN-A and VCGAN-B.

    This paper is the first-of-its-kind to analyze the ethical aspects and limitations of v-reID models as well as widely used vehicle re-identification datasets. It is found that there is a lack of explainable v-reID works and that existing datasets are not privacy compliant or contain biases that can confuse v-reID models. Due to the black-box properties of deep neural networks, researchers do not understand the internal procedures of how v-reID models perform their reasoning. Hence, we need to ensure the following: training data should not contain: (1) any face of drivers and passengers, to respect people's privacy and to ensure that no indirect facial recognition is performed, and (2) any artifact in images that do not originate from the source scene, to avoid confusion of the models. It is found that VehicleID did not respect these characteristics since some samples included faces that were visible and recognizable, and others carried red writing.

    To this end, CycleGAN was proposed to transfer images from VehicleID to VeRi-776. With the proposed method, we managed to generate samples of vehicle images where faces were blurred out and the red timestamps were removed. The experiments produced images that did not resemble vehicles anymore or added objects that were not present in the source image (e.g., bushes, fences, or license plates). These limitations, however, point out further data bias of VehicleID and VeRi-776 datasets, such as the lack of variety in images or non-hidden license plates that should be further investigated.

    Using GAN-based blurring offers advantages over simple resolution reduction by preserving image quality and targeting specific areas. Unlike resolution reduction, which can degrade the entire image, GAN-based methods maintain the overall details of the image. Additionally, GANs can precisely target the face area, leaving other parts of the image, such as the vehicle itself, untouched, which is not possible with simple resolution reduction.

    With the introduction of generative models such as Stable Diffusion1 or Midjourney AI2, vehicle re-identification data generation could also be improved, producing better results than the proposed method. However, our proposed approach is more manageable and smaller in terms of parameters and resources that are needed. Furthermore, it is also more affordable compared to existing generative models as no external server is needed for training. It should be noted that both datasets are built from images collected in China. This makes these datasets a bit biased towards China in terms of people, car models, backgrounds, and direction of traffic. Additionally, this work focuses mainly on two datasets: VehicleID and VeRi-776. While we have not specifically examined other v-ReID datasets for similar problems, these limitations likely exist in them as well. High-resolution images could potentially expose faces, raising privacy issues if not carefully managed.

    In future work, additional datasets will be inspected to identify any further limitations or challenges they may present, and more datasets from different countries and areas will be used to further validate the proposed method regarding efficiency and effectiveness. Furthermore, it is necessary to develop proper regulations and laws to ensure the legal use of v-ReID technologies[45].

  • The authors confirm contribution to the paper as follows: study conception and design: Qian Y; data collection: Qian Y; analysis and interpretation of results: Qian Y; draft manuscript preparation: Qian Y, Barthélemy J, Du B, Shen J. All authors reviewed the results and approved the final version of the manuscript.

  • The data used in this study are from the following resources available in the public domain: VeRi-776 (https://vehiclereid.github.io/VeRi/) and VehicleID (https://www.pkuml.org/resources/pku-vehicleid.html).

  • The authors declare that they have no conflict of interest. Bo Du and Jun Shen are the Editorial Board members of Digital Transportation and Safety who were blinded from reviewing or making decisions on the manuscript. The article was subject to the journal's standard procedures, with peer-review handled independently of these Editorial Board members and their research groups.

  • [1]

    Sun J, Chen T, Wu Y, Tao J. 2021. Identification and functional verification of PLFT gene associated with flowering in herbaceous peony based on transcriptome analysis. Ornamental Plant Research 1(1):7

    doi: 10.48130/OPR-2021-0007

    CrossRef   Google Scholar

    [2]

    Zhang D, Xie A, Yang X, Shi Y, Yang L, et al. 2022. Study of 15 varieties of herbaceous peony pollen submicroscopic morphology and phylogenetic relationships. Horticulturae 8(12):1161

    doi: 10.3390/horticulturae8121161

    CrossRef   Google Scholar

    [3]

    Wang X, Shi X, Zhang R, Zhang K, Shao L, et al. 2022. Impact of summer heat stress inducing physiological and biochemical responses in herbaceous peony cultivars (P. lactiflora Pall.) from different latitudes. Industrial Crops and Products 184:115000

    doi: 10.1016/j.indcrop.2022.115000

    CrossRef   Google Scholar

    [4]

    Song C, Wang Q, Teixeira da Silva JA, Yu X. 2018. Identification of floral fragrances and analysis of fragrance patterns in herbaceous peony cultivars. Journal of the American Society for Horticultural Science 143(4):248−58

    doi: 10.21273/JASHS04420-18

    CrossRef   Google Scholar

    [5]

    Sun Y, Wang W, Zhao L, Zheng C, Ma F. 2019. Changes in volatile organic compounds and differential expression of aroma-related genes during flowering of Rosa rugosa 'Shanxian'. Horticulture, Environment, and Biotechnology 60:741−51

    doi: 10.1007/s13580-019-00166-0

    CrossRef   Google Scholar

    [6]

    Du F, Wang T, Fan JM, Liu ZZ, Zong JX, et al. 2019. Volatile composition and classification of Lilium flower aroma types and identification, polymorphisms, and alternative splicing of their monoterpene synthase genes. Horticulture Research 6:110

    doi: 10.1038/s41438-019-0192-9

    CrossRef   Google Scholar

    [7]

    Jin F, Xu J, Liu XR, Regenstein JM, Wang FJ. 2019. Roasted tree peony (Paeonia ostii) seed oil: benzoic acid levels and physicochemical characteristics. International Journal of Food Properties 22(1):499−510

    doi: 10.1080/10942912.2019.1588902

    CrossRef   Google Scholar

    [8]

    El-Hawary SS, El-Tantawi ME, Kirollos FN, Hammam WE. 2018. Chemical composition, in vitro cytotoxic and antimicrobial activities of volatile constituents from Pyrus communis L. and Malus domestica Borkh. fruits cultivated in Egypt. Journal of Essential Oil Bearing Plants 21(6):1642−51

    doi: 10.1080/0972060X.2018.1553637

    CrossRef   Google Scholar

    [9]

    Hu J, Huang W, Zhang F, Luo X, Chen Y, et al. 2020. Variability of volatile compounds in the medicinal plant Dendrobium officinale from different regions. Molecules 25(21):5046

    doi: 10.3390/molecules25215046

    CrossRef   Google Scholar

    [10]

    Liu G, Fu J, Wang L, Fang M, Zhang W, et al. 2023. Diverse O-methyltransferases catalyze the biosynthesis of floral benzenoids that repel aphids from the flowers of waterlily Nymphaea prolifera. Horticulture Research 10(12):uhad237

    doi: 10.1093/hr/uhad237

    CrossRef   Google Scholar

    [11]

    Pan Y, Quan W, Li C, Hao J, Gao Y. 2021. Analysis of allelochemicals in the leaves of four alpine rhododendrons by gas chromatography-mass spectrometry. BioResources 16(2):3096−102

    doi: 10.15376/biores.16.2.3096-3102

    CrossRef   Google Scholar

    [12]

    Bera P, Mukherjee C, Mitra A. 2017. Enzymatic production and emission of floral scent volatiles in Jasminum sambac. Plant Science 256:25−38

    doi: 10.1016/j.plantsci.2016.11.013

    CrossRef   Google Scholar

    [13]

    Yang L, Aobulikasimu·Nuerbiye, Cheng P, Wang JH, Li H. 2017. Analysis of floral volatile components and antioxidant activity of different varieties of Chrysanthemum morifolium. Molecules 22:1790

    doi: 10.3390/molecules22101790

    CrossRef   Google Scholar

    [14]

    Xin H, Wu B, Zhang H, Wang C, Li J, et al. 2013. Characterization of volatile compounds in flowers from four groups of sweet osmanthus (Osmanthus fragrans) cultivars. Canadian Journal of Plant Science 93(5):923−31

    doi: 10.4141/cjps2012-333

    CrossRef   Google Scholar

    [15]

    Pereira AG, Cassani L, Liu C, Li N, Chamorro F, et al. 2023. Camellia japonica flowers as a source of nutritional and bioactive compounds. Foods 12(15):2825

    doi: 10.3390/foods12152825

    CrossRef   Google Scholar

    [16]

    Yang S, Meng Z, Li Y, Chen R, Yang Y, et al. 2021. Evaluation of physiological characteristics, soluble sugars, organic acids and volatile compounds in 'Orin' apples (Malus domestica) at different ripening stages. Molecules 26(4):807

    doi: 10.3390/molecules26040807

    CrossRef   Google Scholar

    [17]

    Xiong H, Yang Y, Guo W, Yuan J, Yang W, et al. 2024. Study on quality difference between Belamcanda chinensis (L.) DC and Iris tectorum Maxim. based on chemical chromatogram analysis, biological activity evaluation and in vivo distribution rule. Journal of Ethnopharmacology 319:117091

    doi: 10.1016/j.jep.2023.117091

    CrossRef   Google Scholar

    [18]

    Wang Z, Zhao X, Tang X, Yuan Y, Xiang M, et al. 2023. Analysis of fragrance compounds in flowers of Chrysanthemum genus. Ornamental Plant Research 3:12

    doi: 10.48130/OPR-2023-0012

    CrossRef   Google Scholar

    [19]

    Li Z, Zhang X, Li K, Wang P, Li C, et al. 2022. Integrative analysis of transcriptomic and volatile compound profiles sheds new insights into the terpenoid biosynthesis in tree peony. Industrial Crops and Products 188:115672

    doi: 10.1016/j.indcrop.2022.115672

    CrossRef   Google Scholar

    [20]

    Knudsen JT, Gershenzon J. 2020. The chemical diversity of floral scent. In Biology of Plant Volatiles, 2nd edition, eds Pichersky E, Dudareva N. Boca Raton: CRC Press. pp. 57−78. doi: 10.1201/9780429455612-5

    [21]

    Stamm JD. 2023. The language of flowers in the time of COVID: finding solace in zen, nature and ikebana. US: John Hunt Publishing. pp. 152−53.

    [22]

    Nadeem MA, Dubey R, Singh A, Pandey R, Bundela KS. 2017. Processing and quality evaluation of menthol mint oil. International Journal Of Mathematics And Statistics Invention 5(2):68−70

    Google Scholar

    [23]

    Zhang Y, Li C, Wang S, Yuan M, Li B, et al. 2021. Transcriptome and volatile compounds profiling analyses provide insights into the molecular mechanism underlying the floral fragrance of tree peony. Industrial Crops and Products 162:113286

    doi: 10.1016/j.indcrop.2021.113286

    CrossRef   Google Scholar

    [24]

    Qiao Z, Hu H, Shi S, Yuan X, Yan B, et al. 2021. An update on the function, biosynthesis and regulation of floral volatile terpenoids. Horticulturae 7(11):451

    doi: 10.3390/horticulturae7110451

    CrossRef   Google Scholar

    [25]

    Li R, Song C, Niu T, Wei Z, Guo L, et al. 2023. The emitted pattern analysis of flower volatiles and cloning of PsGDS gene in tree peony cultivar 'High Noon'. Acta Horticulturae Sinica 50(2):331−44

    doi: 10.16420/j.issn.0513-353x.2021-0870

    CrossRef   Google Scholar

    [26]

    Zhao Q, Gu L, Li Y, Zhi H, Luo J, et al. 2023. Volatile composition and classification of Paeonia lactiflora flower aroma types and identification of the fragrance-related genes. International Journal of Molecular Sciences 24(11):9410

    doi: 10.3390/ijms24119410

    CrossRef   Google Scholar

    [27]

    Li S, Zhang L, Sun M, Lv M, Yang Y, et al. 2023. Biogenesis of flavor-related linalool is diverged and genetically conserved in tree peony (Paeonia × suffruticosa). Horticulture Research 10(2):uhac253

    doi: 10.1093/hr/uhac253

    CrossRef   Google Scholar

    [28]

    Wang S, Luo Y, Niu T, Prijic Z, Markovic T, et al. 2024. Comparative analysis of the volatile components of six herbaceous peony cultivars under ground-planted and vase-inserted conditions. Scientia Horticulturae 334:113320

    doi: 10.1016/j.scienta.2024.113320

    CrossRef   Google Scholar

    [29]

    Wu Y, Li L, Yuan W, Hu J , Lv Z. 2021. Application of GC × GC coupled with TOF–MS for the trace analysis of chemical components and exploration the characteristic aroma profile of essential oils obtained from two tree peony species (Paeonia rockii and Paeonia ostii). European Food Research and Technology 247:2591−608

    doi: 10.1007/s00217-021-03823-w

    CrossRef   Google Scholar

    [30]

    Li X, Wu J, Wang H, Zhang K, Song F. 2022. Evaluation and comparison of pear flower aroma characteristics of seven cultivars. Horticulturae 8(5):352

    doi: 10.3390/horticulturae8050352

    CrossRef   Google Scholar

    [31]

    Yang YH, Zhao J, Du ZZ. 2022. Unravelling the key aroma compounds in the characteristic fragrance of Dendrobium officinale flowers for potential industrial application. Phytochemistry 200:113223

    doi: 10.1016/j.phytochem.2022.113223

    CrossRef   Google Scholar

    [32]

    Kimani SK, Wang S, Xie J, Bao T, Shan X, et al. 2024. Integration of RNA-Seq and metabolite analysis reveals the key floral scent biosynthetic genes in herbaceous peony. Horticulturae 10(6):617

    doi: 10.3390/horticulturae10060617

    CrossRef   Google Scholar

    [33]

    Pott DM, Osorio S, Vallarino JG. 2019. From central to specialized metabolism: an overview of some secondary compounds derived from the primary metabolism for their role in conferring nutritional and organoleptic characteristics to fruit. Frontiers in Plant Science 10:835

    doi: 10.3389/fpls.2019.00835

    CrossRef   Google Scholar

    [34]

    Hirata H, Ohnishi T, Watanabe N. 2016. Biosynthesis of floral scent 2-phenylethanol in rose flowers. Bioscience, Biotechnology, and Biochemistry 80(10):1865−73

    doi: 10.1080/09168451.2016.1191333

    CrossRef   Google Scholar

    [35]

    Yetisen M, Guclu G, Kelebek H, Selli S. 2022. Elucidation of key aroma enhancement in cloudy lemon juices by the addition of peel oil using GC–MS-Olfactometry. International Journal of Food Science & Technology 57(8):5280−88

    doi: 10.1111/ijfs.15857

    CrossRef   Google Scholar

    [36]

    Abbas F, Zhou Y, O'Neill Rothenberg D, Alam I, Ke Y, et al. 2023. Aroma components in horticultural crops: chemical diversity and usage of metabolic engineering for industrial applications. Plants 12(9):1748

    doi: 10.3390/plants12091748

    CrossRef   Google Scholar

    [37]

    Farré-Armengol G, Filella I, Llusià J, Peñuelas J. 2017. β-Ocimene, a key floral and foliar volatile involved in multiple interactions between plants and other organisms. Molecules 22(7):1148

    doi: 10.3390/molecules22071148

    CrossRef   Google Scholar

    [38]

    Niu TF, Xue X, Guo LL, Yu M, Zhang CJ, et al. 2023. Effects of exogenous methyl jasmonate on volatile components and content of Paeonia suffruticosa 'Luoyanghong' in greenhouse. Scientia Silvae Sinicae 59(5):53−60

    doi: 10.11707/j.1001-7488.LYKX20210977

    CrossRef   Google Scholar

    [39]

    Saad AM, Mohamed AS, Ramadan MF. 2021. Storage and heat processing affect flavors of cucumber juice enriched with plant extracts. International Journal of Vegetable Science 27(3):277−87

    doi: 10.1080/19315260.2020.1779895

    CrossRef   Google Scholar

    [40]

    Wang T, Xie A, Zhang D, Liu Z, Li X, et al. 2021. Analysis of the volatile components in flowers of Paeonia lactiflora Pall. var. Trichocarpa. American Journal of Plant Sciences 12(01):146−62

    doi: 10.4236/ajps.2021.121009

    CrossRef   Google Scholar

    [41]

    Hosseini H, Zahedi B, Jowkar A, Kermani MJ, Karami A. 2021. Investigation of floral scent and essential oil of Rosa iberica petals. Journal of Ornamental Plants 11(2):89−97

    Google Scholar

    [42]

    Zhao G, Ding LL, Hadiatullah H, Li S, Wang X, et al. 2020. Characterization of the typical fragrant compounds in traditional Chinese-type soy sauce. Food Chemistry 312:126054

    doi: 10.1016/j.foodchem.2019.126054

    CrossRef   Google Scholar

    [43]

    Chinyere I, Julius IU. 2020. Determination of aroma components in Vitex doniana fruit syrup following hydrodistillation extraction. Journal of American Science 16(9):84−93

    doi: 10.7537/marsjas160920.07

    CrossRef   Google Scholar

    [44]

    Wang X, Xiong H, Wang S, Zhang Y, Song Z, et al. 2023. Physicochemical analysis, sensorial evaluation, astringent component identification and aroma-active compounds of herbaceous peony (Paeonia lactiflora Pall.) black tea. Industrial Crops and Products 193:116159

    doi: 10.1016/j.indcrop.2022.116159

    CrossRef   Google Scholar

    [45]

    Xiao Y, Huang Y, Chen Y, Xiao L, Zhang X, et al. 2022. Discrimination and characterization of the volatile profiles of five Fu brick teas from different manufacturing regions by using HS–SPME/GC–MS and HS–GC–IMS. Current Research in Food Science 5:1788−807

    doi: 10.1016/j.crfs.2022.09.024

    CrossRef   Google Scholar

    [46]

    Orodu VE. 2021. Migration and degradation by volatile compounds from scent leaf (Ocimum gratissimum) on polyethylene terephthalate packaged water. Chemistry and Physics of Polymers 1(1):1−11

    Google Scholar

    [47]

    Hoepflinger MC, Barman M, Dötterl S, Tenhaken R. 2024. A novel O-methyltransferase Cp4MP-OMT catalyses the final step in the biosynthesis of the volatile 1,4-dimethoxybenzene in pumpkin (Cucurbita pepo) flowers. BMC Plant Biology 24:294

    doi: 10.1186/s12870-024-04955-3

    CrossRef   Google Scholar

    [48]

    Zhang HX, Hu ZH, Leng PS, Wang WH, Xu F, et al. 2013. Qualitative and quantitative analysis of floral volatile components from different varieties of Lilium spp. Scientia Agricultura Sinica 46:790−99

    doi: 10.3864/j.issn.0578-1752.2013.04.013

    CrossRef   Google Scholar

    [49]

    Luo X, Yuan M, Li B, Li C, Zhang Y, et al. 2020. Variation of floral volatiles and fragrance reveals the phylogenetic relationship among nine wild tree peony species. Flavour and Fragrance Journal 35(2):227−41

    doi: 10.1002/ffj.3558

    CrossRef   Google Scholar

    [50]

    Mostafa S, Wang Y, Zeng W, Jin B. 2022. Floral scents and fruit aromas: functions, compositions, biosynthesis, and regulation. Frontiers in Plant Science 13:860157

    doi: 10.3389/fpls.2022.860157

    CrossRef   Google Scholar

    [51]

    Zhao Q, Zhang M, Gu L, Yang Z, Li Y, et al. 2024. Transcriptome and volatile compounds analyses of floral development provide insight into floral scent formation in Paeonia lactiflora 'Wu Hua Long Yu'. Frontiers in Plant Science 15:1303156

    doi: 10.3389/fpls.2024.1303156

    CrossRef   Google Scholar

    [52]

    Ma H, Zhang C, Niu T, Chen M, Guo L, et al. 2023. Identification of floral volatile components and expression analysis of controlling gene in Paeonia ostii 'Fengdan' under different cultivation conditions. Plants 12(13):2453

    doi: 10.3390/plants12132453

    CrossRef   Google Scholar

    [53]

    Middleton R, Tunstad SA, Knapp A, Winters S, McCallum S, et al. 2024. Self-assembled, disordered structural color from fruit wax bloom. Science Advances 10(6):eadk4219

    doi: 10.1126/sciadv.adk4219

    CrossRef   Google Scholar

    [54]

    Su X. 2010. Alcohols compound: the basic material of aromatic chemicals. Value Engineering 29(3):39

    doi: 10.14018/j.cnki.cn13-1085/n.2010.03.102

    CrossRef   Google Scholar

    [55]

    Englezos V, Torchio F, Cravero F, Marengo F, Giacosa S, et al. 2016. Aroma profile and composition of Barbera wines obtained by mixed fermentations of Starmerella bacillaris (synonym Candida zemplinina) and Saccharomyces cerevisiae. LWT 73:567−75

    doi: 10.1016/j.lwt.2016.06.063

    CrossRef   Google Scholar

    [56]

    Wang W, Ge J, Zhang Y, Zhang J. 2024. The male's scent triggered a neural response in females despite ambiguous behavioral response in Asian house rats. Integrative Zoology 19(4):694−709

    doi: 10.1111/1749-4877.12768

    CrossRef   Google Scholar

    [57]

    Song C, Ma H, Li R, Zhao G, Niu T, et al. 2024. Analysis of the emitted pattern of floral volatiles and cloning and functional analysis of the PsuLIS gene in tree peony cultivar 'High Noon'. Scientia Horticulturae 326:112750

    doi: 10.1016/j.scienta.2023.112750

    CrossRef   Google Scholar

  • Cite this article

    Wang A, Luo Y, Niu T, Gao K, Wang S, et al. 2024. Identification and content analysis of volatile components in 100 cultivars of Chinese herbaceous peony. Ornamental Plant Research 4: e032 doi: 10.48130/opr-0024-0029
    Wang A, Luo Y, Niu T, Gao K, Wang S, et al. 2024. Identification and content analysis of volatile components in 100 cultivars of Chinese herbaceous peony. Ornamental Plant Research 4: e032 doi: 10.48130/opr-0024-0029

Figures(5)  /  Tables(4)

Article Metrics

Article views(1142) PDF downloads(264)

ARTICLE   Open Access    

Identification and content analysis of volatile components in 100 cultivars of Chinese herbaceous peony

Ornamental Plant Research  4 Article number: e032  (2024)  |  Cite this article

Abstract: Herbaceous peony (Paeonia lactiflora Pall.) is a well-known and traditional flower in China, occupying a significant position in Chinese traditional culture. The floral scent of the herbaceous peony however remains relatively understudied. The objective of this study was to investigate the floral composition of herbaceous peony by collecting and identifying floral volatiles from 100 cultivars, including P. lactiflora 'Hangbaishao', P. lactiflora 'Hongrongqiu', P. lactiflora 'Biandihong', P. lactiflora 'Zijin Daipao', P. lactiflora 'Zixia Yingxue', and P. lactiflora 'Fenchi Dicui'. The volatile compounds were collected using the dynamic headspace technique and identified through gas chromatography-mass spectrometry (GC-MS). The results demonstrated qualitative and quantitative variations in the floral fragrances emitted by the 100 cultivars, with a total of 16 volatiles belonging to six categories (six alkanes, three alcohols and esters, two terpenes, as well as one each of ether and phenol) being identified. However, it is notable that not all volatile categories were emitted by every cultivar. Moreover, while some compounds were present in all 100 herbaceous peony cultivars, others were exclusive to specific cultivars. The screening revealed that ten of the 16 identified flower volatile compounds exhibited unique floral components. It is noteworthy that benzene,1,4-dimethoxy-, was identified as the most prominent compound in several cultivars, including P. lactiflora 'Taohua Huancai', P. lactiflora 'Xishifen', P. lactiflora 'Dabanhong', P. lactiflora 'Fumantang', and P. lactiflora 'Zhushapan'. Furthermore, the clustering classification results demonstrated that benzene,1,4-dimethoxy-, exhibited the highest variable importance in projection (VIP) value of 3.153, as determined by partial least squares discriminant analysis (PLS-DA).

    • The herbaceous peony, a well-known traditional flower in China[1], is characterized by its large and aesthetically pleasing flowers. The herbaceous peony is a member of the family Paeoniaceae[2], displaying notable adaptability, and significant ornamental value[3]. Studies on aromatic ornamental plants involve an examination of aromatic components, and genetic mechanisms[4], including Rosa rugosa Thunb.[5], Lilium brownii var. viridulum Baker[6], and Paeonia suffruticosa Andr.[7], Pyrus communis L.[8], Dendrobium officinale [9], Nymphaea tetragona[10], Rhododendron simsii[11], Jasminum sambac[12], studies were conducted on Chrysanthemum morifolium[13], Osmanthus fragrans[14], Camellia japonica[15], Malus[16], and Iris tectorum Maxim.[17] Historically, research has focused on factors such as flower shape, color, blooming season, and resilience, with less attention given to the floral scent[18].

      The floral scent has been identified as a significant ornamental attribute of herbaceous peony[3,19], and is also a prominent feature in numerous plant species[20]. It is frequently described as the 'essence of flowers'[21] and is derived from a range of volatile compounds that are synthesized within the plant and subsequently released into the atmosphere[22]. To date, over 1,700 volatile compounds have been identified in a variety of plants, with a multitude of applications in the manufacture of perfumes, cosmetics, culinary seasonings, and pharmaceuticals[23,24]. The composition and concentration of these volatile compounds exhibit considerable variation across different species, genus, and cultivars. Nevertheless, there is a paucity of research dedicated to the analysis of fragrance constituents and their respective concentrations in herbaceous peony and tree peony[25]. Song et al.[4] identified a total of 130 volatile compounds across 30 cultivars of herbaceous peony, encompassing 72 aromatic constituents. The 24 cultivars exhibiting heightened fragrance were categorized into five distinct aroma profiles: woody scent, fruity scent, lily scent, rose scent, and an orange blossom scent. Zhao et al.[26] conducted a study in which 68 volatile compounds and 26 significant aroma constituents were identified from a sample of 87 herbaceous peony cultivars. The researchers determined that herbaceous peony contain characteristic aromatic substances, including linalool (resembling lily of the valley), geraniol (exhibiting a pleasant geranium-like scent), citronellol (evoking a fresh and light rose and leaf fragrance), and phenylethyl alcohol (noted for its distinctive rose aroma), based on the content and odor threshold of these main aroma components. In a separate study, Li et al.[27] identified 128 volatile compounds from 24 tree peony cultivars, with the predominant classes being terpenes, alcohols, and esters. The distribution pattern of these primary fragrance constituents led to the categorization of 24 tree peony cultivars into four types: grass scent (ocimene), woody scent (longifolene), lily of the valley scent (linalool), and fruity scent (2-ethyl hexanol). It has been demonstrated that the distinctive fragrances of different plant species are the result of the presence of specific volatile compounds in varying quantities and ratios. Furthermore, the quantity of fragrance emitted by flowers is contingent upon their developmental stage[28].

      Floral substances derived from plants are classified as secondary metabolites, which are released by flowering plants and predominantly comprise a range of volatile compounds characterized by relatively low molecular weights. In a comprehensive analysis of the aromatic compounds present in P. rockii and P. ostii 'Fengdan', Wu et al.[29] employed two-dimensional gas chromatography coupled with time-of-flight mass spectrometry (GC × GC-TOF/MS). The results indicated that the aroma profile of P. rockii was primarily characterized by the presence of alcohols, alkanes, and acids, while the aroma profile of P. ostii 'Fengdan' was predominantly defined by aldehydes, alcohols, and terpenes. In a separate investigation, Li et al.[30] sought to identify and analyze the volatile compounds present in the flowers of seven pear cultivars (Anli, Bayuesu, Golden, Brown Peel, KorlaXiangli, Lyubaoshi, and Xizilü). The findings indicated that certain aldehydes constitute significant characteristics of these cultivars and are recognized as essential active odorants, which emit pronounced citrus and floral fragrances. Yang et al.[31] successfully identified and characterized 34 volatile compounds in the Dendrobium officinale flowers. Of these, 18 compounds were identified as principal odorants, including 1-octen-3-ol, hexanal, nonanal, phenylacetaldehyde, linalool, 4-oxoisophorone, theaspirane, and methyl salicylate. Furthermore, Kimani et al.[32] identified geraniol, β-caryophyllene, 2-phenylethanol, citronellol, and 1,8-cineole as the primary aromatic constituents in 24 cultivars of herbaceous peony, including P. lactiflora 'LianTai' and P. lactiflora 'Hongyan Feishuang'. Aromatic compounds are recognized as the primary chemical constituents of aromatic plants, playing a crucial role in the synthesis of secondary metabolites[33], and fragrance development. These compounds exhibit a diverse range of forms. For example, phenethyl alcohol is found in rose, mint contains menthol, and lemon includes citric acid[22,34,35].

      This study employed a combination of dynamic headspace sampling technology[36] and GC-MS to analyze the volatile components and concentrations in 100 international herbaceous peony cultivars during the half-opening stage. The objective was to elucidate the aromatic profile of the herbaceous peony. The findings of this study establish a fundamental framework for further investigation and exploitation of the fragrances of herbaceous peony flowers and provide a valuable resource for enhancing the economic value of herbaceous peony.

    • The experimental materials used in this study were obtained from the Luoyang Academy of Agriculture and Forestry Sciences (Luoyang City, Henan Province, China) between 20 April and 8 May 2022. The majority of the materials were collected between 10 and 12 am. The subjects of the experiment were herbaceous peony plants sourced from the herbaceous peony resource garden affiliated with the Henan University of Science and Technology. As outlined in Table 1 and Fig. 1, herbaceous peony cultivars demonstrating consistent growth patterns and flowering stages were identified, and the methodology entailed the repetition of each sample on three occasions.

      Table 1.  Names and numbers of 100 herbaceous peony cultivars.

      100 herbaceous peony cultivars
      'Hangbaishao' 'Hongrongqiu' 'Biandihong' 'Zijin Daipao' 'Zixia Yingxue' 'Fenchi Dicui' 'Xishifen' 'Yinlong Hanzhu' 'Yinxian Xiuhongpao' 'Jindaiwei'
      'Luhong' 'Xueyuan Hongxing' 'Mozijin' 'Yahong' 'Wulong Tanhai' 'Hongyan Zhengshuang' 'Xingguang Chanlan' 'Yanlihong' 'Hongling Chijin' 'Fenzhuangyuan'
      'Taohua Huancai' 'Zhongshenghong' 'Ziling' 'Luxihong' 'Zifurong' 'Hongling Chijin' 'Huguang Shise' 'Hongyuqiu' 'Yanzhi Dianyu' 'Lantian Piaoxiang'
      'Zhushapan' 'Hongyun Yingri' 'Yanzi Xiangyang' 'Yanzhihong' 'Zaoyuanhong' 'ChilongCaifeng' 'Chaoshihong' 'Qingwen' 'Shaifugui' 'Ziyanshuang'
      'Gaoganfen' 'Qundiehui' 'Meirenmian' 'Meiju' 'Dafugui' 'Zhifeng Zhaoyang' 'Xueyuan Hongxing' 'Dahongpao' 'Zixiuqiu' 'Canglong'
      'Gaoganhong' 'Hongyan Feishuang' 'Dabanhong' 'Zifengyu' 'Hongpan Jinqiu' 'Hushui Dangxia' 'Yinlong Huihai' 'Baihuazi' 'Taohuafen' 'Wawamian'
      'Fenpanjinxing' 'Heixiuqiu' 'Shuanghonglou' 'Changshouhong' 'Hongyan Lushuang' 'Tuopan Jinhua' 'Hongling Chijin' 'Linglongyu' 'Jinzan Ciyu' 'Xiangyang Qihua'
      'Jinbian Hongge' 'Duoyezi' 'Fenzilou' 'Furong Jinhua' 'Fenkui' 'Guifei Chacui' 'Huolian Jingang' 'Hongguanfang' 'Fenmian Taohua' 'Taoranzui'
      'Zhaoyanghong' 'Hongfengyu' 'Fumantang' 'Shaonvfen' 'Danfeng' 'Liantaizi' 'Meiguihong' 'Fenfurong' 'Fenling Hongzhu' 'Fenqiu'
      'Fencuiqiu' 'FengChao Chuyu' 'Lanju' 'Jinsanhong' 'Zhaoyuanfen' 'Hongfeng' 'Qiaoling' 'Tuanye Jinqiu' 'Guohong' 'Tongquechun'
      The numbers 1−100 are listed from top to bottom, left to right respectively.

      Figure 1. 

      Morphological characteristics of 100 herbaceous peony cultivars at the half-opening stage. The order of the above pictures is relative to the order of cultivars in Table 1.

    • The Gas Chromatography-Mass Spectrometry System (GC8890-MS5977B) from Agilent Technologies, USA, and the Atmospheric Sampler QC-1S from the Beijing Institute of Labor Protection were utilized in the study.

    • The reagents used included Tenax TA as the adsorbent, ethyl caprate, dichloromethane, pentane, n-hexane of chromatography grade, ethyl decanoate, ethyl acetate, and a standard solution of n-alkane mixture (ranging from C8 to C40) obtained from Sigma-Aldrich, USA.

    • The dynamic headspace adsorption technique employed in this study was a sampling bag (355 mm × 508 mm, Reynolds, USA) hermetically sealed at one end with an activated carbon filter tube. The bag was meticulously wrapped around a live peony flower to minimize contact and prevent damage to the bag. The bag's opposite end was connected to a Tenax TA adsorption tube (6 mm outer diameter, 100 mm length, filled with adsorbent) and an atmospheric sampler via tasteless silicone tubing. The flow rate of the atmospheric sampler was set at 400 mL·min−1 and the sampling duration was 3 h. Following the sampling period, the adsorption tube was sealed with cling film and aluminum foil, then placed in a self-sealing bag and stored in an ultra-low temperature cooler for transport to the laboratory. The sample was then eluted with n-hexane during sample processing, and the eluate was transferred to a brown sample bottle for subsequent analysis.

      The following conditions were observed in the gas chromatography (GC) analysis: the chromatographic column employed is a flexible quartz capillary column, with a length of 30 mm, an internal diameter of 0.25 mm, and a pore size of 0.25 μm. The flow rate of the column is set at 1.2 mL·min−1. The temperature of the column is maintained according to a specific protocol. It is initially set at 70 °C and held for 1 min, then increased to 136 °C at a rate of 6 °C·min−1, followed by further increases to 138 °C at a rate of 1 °C·min−1, then to 142 °C at a rate of 2 °C·min−1, and finally to 143 °C at a rate of 0.5 °C∙min−1. The temperature is increased by 5 °C·min−1 and subsequently to 160 °C at a rate of 2 °C·min−1, before reaching 250 °C at a rate of 10 °C·min−1. The injector temperature is set at 250 °C, with a carrier gas of high-purity helium at a flow rate of 1 mL·min−1. The injection mode is a split injection, with a split ratio of 9:1, and the injection volume is 2 μL.

      The following conditions were employed for the mass spectrometry (MS) analysis: The electron impact (EI) source is operated at 70 eV, with the interface temperature set to 250 °C and the ion source temperature maintained at 230 °C. The quadrupole temperature is controlled at 150 °C, and the scan range is from 25 to 400 amu.

    • Before the analysis of the sample using gas chromatography, the 500 mg∙L−1 n-alkane mixed standard solution should be diluted with n-hexane at a ratio of 1:50, in accordance with the specified conditions for the chromatography. It is essential to record the retention time for each n-alkane and to compare the resulting Retention Index (RI) values with those documented in the literature to facilitate the identification of the compounds in question. The following formula is used to calculate the RI:

      RI=100×n+100×(txtn)/(tn+1tn)

      The location of the aforementioned item is as follows: The retention index (RI) represents the retention time of the volatile substances under examination. The number of carbon atoms in the straight-chain alkane preceding the analyte is represented by n. The retention time of the analyte is represented by tx. The retention time of the straight-chain alkane before the analyte is represented by tn. The retention time of the straight-chain alkane following the analyte is represented by tn+1. The retention time of the analyte falls between the retention times of tn and tn+1. Qualitative analysis of volatile components is conducted through consultation with the NIST 17 spectral library, with cross-referencing of RI values, literature sources, and other pertinent resources, including books.

    • An internal standard solution, comprising 69.32 mg∙L−1 of ethyl decanoate in ethyl acetate, is employed. A volume of 0.4 μL of the internal standard solution is added to each 80 μL sample. Subsequently, quantitative calculations are performed in accordance with the following formula:

      Contentofeacharomasubstance(μgg1)=PeakareaofeacharomasubstancePeakareaoftheinternalstandard×Concentrationoftheinternalstandard(mgL1)×Volumeoftheinternalstandard(μL)Volumeofthesample(g)×f

      where, f is the correction factor of each component to the internal standard, f = 1.

    • The analysis of variance can be conducted using the statistical software package SPSS, while graph plotting can be accomplished with the Origin 2022 software. The software Metaboanalyst and the Microbiome Analysis Platform are capable of performing data normalization, partial least squares discriminant analysis (PLS-DA), and cluster analysis.

    • An analysis was conducted to determine the main volatile compounds present in 100 herbaceous peony cultivars during the half-opening stage. This was achieved through the utilization of database retrieval and manual identification methods. The results are outlined in Table 2. A total of 16 volatile components were identified and classified into six distinct groups. The data revealed that alkanes constituted six types, representing 37.5% of the total volatile components. This was followed by four types of esters at 25%, three types of alcohols at 18.75%, and one type each of terpenes, ethers, and phenols, each accounting for 6.25% of the total volatile components. The results of this analysis indicate that the predominant categories of volatile compounds found in herbaceous peony cultivars are alkanes, esters, and alcohols.

      Table 2.  The volatile components of 100 herbaceous peony cultivars.

      Compound number RT (min) CAS number Compounds Compound classification Chemical formula Retention index
      Calculated value Reference value
      1 3.273 111-84-2 Nonane Alkanes C9H20 900 900
      2 4.805 124-18-5 Decane Alkanes C10H22 1,000 1,000
      3 5.727 13877-91-3 (Z)-β-ocimene Terpenes C10H16 1,038 1,037
      4 7.063 60-12-8 Phenylethyl alcohol Alcohols C8H10O 1,115 1,109
      5 8.133 150-78-7 Benzene,1,4-dimethoxy- Ethers C8H10O2 1,165 1,168
      6 9.502 106-22-9 Citronellol Alcohols C10H20O 1,228 1,228
      7 10.084 106-25-2 Nerol Alcohols C10H18O 1,220 1,219
      8 10.187 103-45-7 Methyl cinnamate Esters C10H12O2 1,260 1,258
      9 12.976 103-26-4 2-Propenoic acid,3-phenyl-,methyl ester Esters C10H10O2 1,389 1,380
      10 14.88 131-11-3 Dimethyl phthalate Esters C10H10O4 1,456 1,466
      11 16.174 629-62-9 Pentadecane Alkanes C15H32 1,500 1,500
      12 16.664 128-37-0 Butylated hydroxytoluene Phenols C15H24O 1,513 1,513
      13 19.877 544-76-3 Hexadecane Alkanes C16H34 1,600 1,601
      14 24.137 629-78-7 Heptadecane Alkanes C17H36 1,699 1,700
      15 31.517 84-74-2 Dibutyl phthalate Esters C16H22O4 1,964 1,907
      16 33.398 646-31-1 Tetracosane Alkanes C24H50 2,400 2,400
    • As illustrated in Fig. 2, alkane compounds were undetected in 30 cultivars, including P. lactiflora 'Hushui Dangxia', P. lactiflora 'Tuopan Jinhua', P. lactiflora 'Qiaoling', P. lactiflora 'Yinlong Hanzhu', and P. lactiflora 'Yanlihong'. Among the 100 herbaceous peony cultivars, the highest concentration of alkane compounds was observed in P. lactiflora 'Heizijin' (10.66 ± 2.01 μg·g−1), with the range of alkane compounds concentration spanning from 0.00 to 10.66 μg·g−1.

      Figure 2. 

      Comparative heat map depicting the release of six types of volatile compounds from various herbaceous peony cultivars.

    • As shown in Fig. 2, ester compounds were discernible in all 44 cultivars of herbaceous peony at the half-opening stage. However, the content of ester compounds was generally not notably high in most cultivars. The highest ester compounds content was observed in P. lactiflora 'Changshouhong' (9.15 ± 0.03 μg·g−1), followed by P. lactiflora 'Zaoyuanhong' (3.55 ± 0.40 μg·g−1), P. lactiflora 'Hongyun Yingri' (3.37 ± 0.11 μg·g−1), and P. lactiflora 'Saifugui' (3.25 ± 0.67 μg·g−1). The ester compounds content among these three cultivars was found to be similar, with a range of 0.00 to 9.15 μg·g−1.

    • As depicted in Fig. 2, the majority of the 100 cultivars of herbaceous peony at the half-opening stage exhibited the presence of alcohol compounds. Only 23 cultivars, including P. lactiflora 'Taohua Huancai', P. lactiflora 'Zhushapan', and P. lactiflora 'Gaoganhong' exhibited no detection. The highest alcohol compounds content was observed in P. lactiflora 'Hongfeng' (22.98 ± 3.86 μg·g−1), which was significantly higher than that of other herbaceous peony cultivars. Subsequently, P. lactiflora 'Wandai Shengse' (16.23 ± 2.28 μg·g−1) exhibited the second-highest alcohol compounds content, with a range of 0.00 to 22.98 μg·g−1.

    • As illustrated in Fig. 2, only 19 of the herbaceous peony cultivars exhibited detectable levels of terpene compounds, with significant differences in content (p < 0.05). The highest content was observed in P. lactiflora 'Hongfengyu' (8.19 ± 1.02 μg·g−1), followed by P. lactiflora 'Wandai Shengse' (4.93 ± 0.09 μg·g−1), P. lactiflora 'Jinzan Ciyu' (2.92 ± 1.75 μg·g−1), P. lactiflora 'Dabanhong' (0.07 ± 0.13 μg·g−1), P. lactiflora 'Jinbian Hongge' (0.14 ± 0.23 μg·g−1), and P. lactiflora 'Mozi Hanjin' (0.16 ± 0.28 μg·g−1), among others. The range of terpene compounds content was found to vary from 0.00 to 8.19 μg·g−1.

    • The analysis of 50 herbaceous peony cultivars revealed the presence of ether compounds in all samples, with notable variations in their content (p < 0.05). The highest content of ether compounds was observed in P. lactiflora 'Dabanhong' (22.84 ± 2.15 μg·g−1), followed by P. lactiflora 'Taohua Yingcai' (19.53 ± 2.44 μg·g−1). The lowest levels were observed in P. lactiflora 'Danfeng' (0.06 ± 0.11μg·g−1), P. lactiflora 'Ziling' (0.15 ± 0.26 μg·g−1), and P. lactiflora 'Huolian Jingang' (0.12 ± 0.21 μg·g−1). The range of ether compounds content was observed to vary from 0.00 to 22.84 μg·g−1.

    • The analysis revealed that only five herbaceous peony cultivars exhibited discernible levels of phenol compounds, namely P. lactiflora 'Jinbian Hongge' (0.15 ± 0.05 μg·g−1), P. lactiflora 'Zhaoyanghong' (0.34 ± 0.02 μg·g−1). The remaining cultivars exhibited lower levels of phenol compounds, with the lowest concentration observed in P. lactiflora 'Hongrongqiu' (0.17 ± 0.03 μg·g−1), followed by P. lactiflora 'Xueyuan Honghua' (0.01 ± 0.02 μg·g−1), and P. lactiflora 'Ziling' (0.27 ± 0.05 μg·g−1). The five cultivars exhibited notably lower levels of phenol compounds, with values consistently below 1 μg·g−1. The remaining cultivars were found to be devoid of phenol compounds.

    • The analysis of the 16 volatile compounds detected revealed that, aside from alkanes such as nonane, the remaining 10 compounds from five classes all exhibited characteristic aromas, as detailed in Table 3. These aromatic compounds were present in the majority of samples, with concentrations exceeding 0.01 μg·g−1. Of particular note is the detection of benzene,1,4-dimethoxy-, in the majority of samples, with relatively high concentrations observed (Fig. 3).

      Table 3.  Characteristics of aroma compounds.

      No. Compound name Odor characteristics
      1 (Z)-β-ocimene The scent of grass and flowers is accompanied by the aroma of orange blossom oil[37]
      2 Phenylethyl alcohol Sweet rose-like fragrance[38]
      3 Benzene,1,4-dimethoxy- The fragrance of cloves[39]
      4 Citronellol Has a sweet rose aroma[40]
      5 Nerol There is a sweet rose fragrance[41]
      6 Acetic acid, 2-phenylethyl ester There is a reminiscent of honey-like floral fragrance[42]
      7 Methyl cinnamate Sweet smelling fragrance[43]
      8 Dimethyl phthalate The substance emits a delicate fragrance[44]
      9 Butylated hydroxytoluene The presence of a carbonic acid taste can
      influence the aroma of wine[45]
      10 Dibutyl phthalate The substance emits a delicate fragrance[46]

      Figure 3. 

      Content of characteristic aroma compounds in herbaceous peony cultivars.

    • A data matrix of dimensions 100 × 10 was constructed, representing the content of 10 aromatic compounds in 100 herbaceous peony cultivars as variables. A cluster heatmap was generated using the microbiome analysis platform, as illustrated in Fig. 4. In light of the clustering results and a comprehensive consideration of the major aromatic components, the 100 herbaceous peony cultivars are ultimately classified into two groups (Table 4). The first group of herbaceous peony cultivars is distinguished by a marked prevalence of benzene,1,4-dimethoxy-, with markedly elevated levels in comparison to other cultivars. This gives rise to a pronounced clove scent, indicative of a clove floral type. This initial classification is based on the presence of specific compounds and is therefore applicable to only five cultivars. The cultivars in question are P. lactiflora 'Taohua Huancai', P. lactiflora 'Xishifen', P. lactiflora 'Dabanhong', P. lactiflora 'Fumantang', and P. lactiflora 'Zhushapan'. The second group generally exhibits lower levels of aromatic compounds, resulting in milder scents that may be characterized as a light floral type. The second group comprises 95 cultivars, including representative cultivars such as P. lactiflora 'Meiju', P. lactiflora 'Shaonvfen', P. lactiflora 'Fenmian Taohua', P. lactiflora 'Fenling Hongzhu', and P. lactiflora 'Guohuo', among others.

      Figure 4. 

      Heat map showing the clustering analysis of 100 herbaceous peony cultivars. A-Benzene,1,4-dimethoxy-, B-Citronellol, C-Nerol, D-Acetic acid, 2-phenylethyl ester, E-Methyl cinnamate, F-Dimethyl phthalate, G-(Z)-β-ocimene, H-Phenylethyl alcohol, I-Butylated hydroxytoluene, J-Dibutyl phthalate. The numbers 1−100 correspond to the cultivar names listed in Table 1.

      Table 4.  Cluster analysis of characteristic aroma components in different herbaceous peony cultivars.

      Groups Herbaceous peony cultivars
      1 'Taohua Huancai', 'Xishifen', 'Dabanhong', 'Fumantang', and 'Zhushapan'
      2 'Liantaizi', 'Hushui Dangxia', 'Shaifugui', 'Hongfeng', 'Wandai Shengse', 'Zhaoyuanfen', 'Wawamian', 'Lanju', 'Shuanghonglou', 'Fenling Hongzhu', 'Guohuo', 'Fenmian Taohua', 'Yinlong Tanhai', 'Chaoshihong', 'Shaonvfen', 'Meiju', 'Huolian Jingang', 'Meiguihong', 'Chilong Huancai', 'Yinlong Hanzhu', 'Yanlihong', 'Zhaoyanghong', 'Yinxian Xiuhongpao', 'Fenchi Dicui', 'Xueyuan Hongxing', 'Fenfurong', 'Linglongyu', 'Xiangyang Qihua', 'Hongrongqiu', 'Huguang Shise', 'Yanzhihong', 'Duoyezi', 'Mozijin', 'Guifei Chacui', 'Ziling', 'Zixia Yingxue', 'Zixiuqiu', 'Jinzan Ciyu', 'Meirenmian', 'Zifengyu', 'Jinshanhong', 'Hongyan Lushuang', 'Hongguanfang', 'Jindaiwei', 'Canglong', 'Tuopan Jinhua', 'Huolian Chijin', 'Fengchao Chuyu', 'Hongyuqiu', 'Xueyuan Hongxing', 'Qiaoling', 'Dahongpao', 'Qundiehui', 'Tuanye Jinqiu', 'Dafugui', 'Taoranzui', 'Yanzhi Dianyu', 'Tongquechun', 'Ziyanshuang', 'Gaoganfen', 'Fenpan Jinxing', 'Fenkui', 'Lantian Piaoxiang', 'Zifeng Zhaoyang', 'Xingguang Canlan', 'Hongyan Feishuang', 'Biahuazi', 'Taohuafen', 'Danfeng', 'Hongfengyu', 'Fenzilou', 'Yanzi Xiangyang', 'Zaoyuanhong', 'Luhong', 'Yahong', 'Luxihong', 'Furong Jinhua', 'Jinbian Hongge', 'Wulong Tanhai', 'Zhongshenghua', 'Zifurong', 'Hongyan Zhengshuang', 'Gaoganhong', 'Heixiuqiu', 'Hongling Chijin', 'Hongyun Yingri', 'Changshouhong', 'Fencuiqiu', 'Qingwen', 'Hongpan Jinqiu', 'Zijin Daipao', 'Biandihong', 'Fenqiu', 'Hangbaishao' and 'Fenzhuangyuan'
    • Following the clustering of 100 cultivars into two groups, a partial least squares discriminant analysis (PLS-DA) was conducted on the content of 10 aroma compounds in the 100 cultivars using Metaboanalyst software. The results of the analysis are presented in Fig. 5. The PLS model for aroma compounds demonstrated satisfactory reliability, as evidenced by R2 and Q2 values of 0.702 and 0.598, respectively. Moreover, the PLS-DA results demonstrated variations in the profile of aroma compounds between the two groups of cultivars (Fig. 5a). The application of a VIP criterion greater than 1 identified a differentiating component (Fig. 5b). The VIP values in the PLS-DA model provided further insight into the contribution of each component to the model, with components having a value of VIP > 1 being considered significant. For instance, benzene,1,4-dimethoxy-, exhibited a VIP value of 3.153 and was identified as a principal component accountable for the discrepancies among herbaceous peony cultivars (Fig. 5b), corroborating the findings of the clustering analysis. It can therefore be posited that benzene,1,4-dimethoxy- is a characteristic aroma component of these herbaceous peony cultivars.

      Figure 5. 

      PLS-DA scores of 100 herbaceous peony cultivars under two cluster groups.

      Variables A and B represent the first and second categories, respectively. The specific variables include A-Benzene,1,4-dimethoxy-, B-Citronellol, C-Nerol, D-Acetic acid, 2-phenylethyl ester, E-Methyl cinnamate, F-Dimethyl phthalate, G-(Z)-β-ocimene, H-Phenylethyl alcohol, I-Butylated hydroxytoluene, J-Dibutyl phthalate.

    • The present study comprises a comprehensive identification and analysis of the volatile constituents present in 100 herbaceous peony cultivars during the half-opening stage. The findings indicated that alkanes, alcohols, and ethers were the most prevalent volatile compounds, with benzene,1,4-dimethoxy- was identified as the distinctive aromatic components.

      One such molecule is benzene,1,4-dimethoxy-, a methoxylated aromatic volatile compound that is known to elicit physiological and behavioral responses in a diverse range of insect pollinators. It serves as a principal floral volatile in a number of plant species belonging to diverse genera, including Salix, Lithophragma, Nelumbo, Catasetum, Allium, and Fragaria[47]. Wang et al.[40] identified the common floral component, benzene,1,4-dimethoxy-, in all eight herbaceous peony cultivars. Furthermore, Kimani et al.[32] identified 95 volatile organic compounds in 24 herbaceous peony cultivars, including benzene,1,4-dimethoxy-, which is a phenolic methyl ether containing a benzene skeleton but not derived from aromatic amino acids. Rather, it is a member of a particular chemical class that is responsible for the olfactory characteristics of specific plant varieties. The types and contents of volatile components of herbaceous peonies may be associated with the sampling method, sampling location and time. Additionally, the types and contents of volatile compounds in plants may be influenced by different planting environmental conditions[48].

      In recent years, there has been a growing emphasis on the natural floral volatiles present in herbaceous peonies, with the fragrance components demonstrating a diverse range of applications in the fields of healthcare, perfumes, and cosmetics[49]. Floral scent represents a significant component of plant volatiles, which are primarily composed of terpenes, aromatic hydrocarbons, fatty acids, and their derivatives, as well as sulfur and nitrogen-containing compounds[27,50]. These compounds are taxonomically categorized into three primary classes, contingent upon their biogenic origins, namely fatty acid derivatives, phenylpropanoids/benzenoids, and terpenoids[51]. In the present study, the volatile components of the 100 cultivars of herbaceous peony were predominantly identified as alkanes, esters, and alcohols. The most abundant type of compound was identified as alkanes. The available evidence suggests that straight-chain alkanes represent the primary constituents of plant leaf wax[52]. These waxes are not exclusive to leaves but may also be found on other plant organs, including flowers and fruit surfaces[53]. This indicates that wax layers may cover the surfaces of the majority of herbaceous peony cultivars. Alcohols play a significant role in the fragrance industry, serving as essential raw materials for synthetic fragrances and as an indispensable component in perfumery[54]. The presence of abundant ether compounds results in the production of pleasant floral and fruity aromas, while simultaneously enhancing the richness, typicality, and complexity of plant fragrances[55].

      However, due to the constraints of the existing literature, some volatile components, such as specific alkanes, have not yet been conclusively identified as fragrance components. Further research is required to ascertain whether these components contribute to the fragrance of herbaceous species. Alkane compounds have relatively high thresholds[56] and make minimal contributions to the overall scent[28]. Accordingly, the analysis of fragrance compounds excludes the contributions made by nonane, decane, pentadecane, hexadecane, heptadecane, and tetracosane.

    • This study employed dynamic headspace bag adsorption of live plant materials and gas chromatography-mass spectrometry (GC-MS) analysis techniques to identify a total of 16 volatile components in 100 herbaceous peony cultivars at the half-opening stage[57]. The components were primarily categorized into six major groups: alkanes, esters, alcohols, terpenes, ethers, and phenols. The predominant volatile compounds were alkanes, alcohols, and ethers, while benzene,1,4-dimethoxy- was identified as the main aromatic component. Significant variations in the total content of the main aromatic components were observed among the different herbaceous peony cultivars at the half-opening stage. In particular, P. lactiflora 'Taohua Huancai', P. lactiflora 'Xishifen', P. lactiflora 'Dabanhong', P. lactiflora 'Fumantang', and P. lactiflora 'Zhushapan' exhibited the highest content of aromatic components, resulting in a more intense floral fragrance. The intensity and characteristics of the aroma exhibited notable variation among different herbaceous peony cultivars, attributable to differences in the quantity and composition of the aromatic components. This is a crucial indicator for evaluating the quality of herbaceous peony. This study provides a theoretical foundation for understanding the formation and regulation mechanisms of herbaceous peony aroma characteristics, while also offering technical support for accelerating industrial development and utilization of herbaceous peony aromas.

      • This research was funded by the Science and Technology Innovation Talents in Universities of Henan Province (Grant No. 22HASTIT036) and the Project of Henan Province Traditional Chinese Medicine Industry Technology System (Grant No. 2024-24).

      • The authors confirm contribution to the paper as follows: study conceptualization, reviewing, editing and funding acquisition: Guo L; material preparation: Wang A, Luo Y, Niu T, Zhao X, Gao K; data curation: Wang A, Luo Y, Niu T, Wang S; draft manuscript preparation: Wang A, Luo Y; manuscript reviewing and editing: Hou X. All authors reviewed the results and approved the final version of the manuscript.

      • All data generated or analyzed during this study are included in this published article.

      • The authors declare that they have no conflict of interest.

      • # Authors contributed equally: Aixin Wang, Yasang Luo

      • Copyright: © 2024 by the author(s). Published by Maximum Academic Press, Fayetteville, GA. This article is an open access article distributed under Creative Commons Attribution License (CC BY 4.0), visit https://creativecommons.org/licenses/by/4.0/.
    Figure (5)  Table (4) References (57)
  • About this article
    Cite this article
    Wang A, Luo Y, Niu T, Gao K, Wang S, et al. 2024. Identification and content analysis of volatile components in 100 cultivars of Chinese herbaceous peony. Ornamental Plant Research 4: e032 doi: 10.48130/opr-0024-0029
    Wang A, Luo Y, Niu T, Gao K, Wang S, et al. 2024. Identification and content analysis of volatile components in 100 cultivars of Chinese herbaceous peony. Ornamental Plant Research 4: e032 doi: 10.48130/opr-0024-0029

Catalog

  • About this article

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return