Introduction

In grapevine research, leaf hair has been used in ampelographic studies as one of the descriptors to identify and define various cultivars1. However, they have more biological features. Depending on the type and density of the hair on the leaves’ lower side, it can play a beneficial role concerning vine health2. Like other agricultural crops, grapevines (Vitis vinifera ssp.) are susceptible to several pests and fungal pathogens, causing extreme yield losses and quality reduction3,4. Some wild Vitis accessions prevent the attack utilizing physico-chemical factors and genetic resistances5,6. To date, several genetic loci have been identified conferring resistance to fungal pathogens7,8,9. Plasmopara viticola ((Berk. and Curt.) Berl. and de Toni), the causative agent of downy mildew, infects all green parts of the plant, with the stomata on the lower leaf surface being the entry site for infection. Recent studies have shown the importance of grapevine abaxial leaf hair as a natural physical barrier against downy mildew infection5,10. Quantitative trait loci (QTL) for leaf hair density identified on four linkage groups of grapevine were also associated with resistance to downy mildew11,12. Furthermore, limited infection with anthracnose (Elsinoë ampelina) on the abaxial vs. adaxial leaf surface was observed due to high leaf hair density on the lower leaf surface that trapped conidia13. A major QTL in grapevine was identified, explaining the relevance of leaf hair density associated with the abundance of a predatory mite (Typhlodromus pyri) acting as a biological control agent14. Despite the high relevance, leaf hair density has so far only been rated visually14,15, and no previous studies have clearly reported or illustrated accurate and efficient means for measuring leaf hair density. Therefore, studies aiming to identify the correlation between leaf hair density and pests or pathogens tend to create biased data, due to high subjectivity in leaf hair phenotyping.

In recent years, advances in machine-learning-based image analysis have been extensively implemented in plant trait phenotyping and growth monitoring studies16,17,18,19. In particular, convolutional neural networks (CNNs) with varying numbers of layers and increased image classification depth have been successfully used in a wide range of biological classification systems20,21,22,23,24. In contrast to CNNs with lower numbers of layers like AlexNet, VGG-13, VGG-16, GoogLeNet, DenseNet169, and PReLU-net, the Residual Network (ResNet)-based CNN architecture has dense layers with fewer parameters, a minor error rate, and skip connections making it an intriguing choice for more difficult tasks25.

Empirical studies and benchmarks have consistently demonstrated that ResNet architectures outperform traditional CNN architectures in various computer vision tasks, including plant disease classification26,27,28,29. The architecture of ResNet enables the reuse of learned features across different layers. This reuse property enhances the network’s ability to generalize to unseen data, contributing to improved classification performance, particularly in scenarios with limited training data25,30,31. Currently, in grapevine research, several CNN-based tools are available demonstrating the power and efficiency of this technology in supporting qualitative and quantitative phenotyping of numerous traits32,33,34,35,36.

The main objective of this work was the development and implementation of an automated phenotyping tool based on ResNet CNN for the objective detection and quantification of grapevine leaf hairs. Therefore, a minimal set of training images and computational resources were used to evaluate the effectiveness of the CNN model in real-world applications. A leaf hair ResNet aims to enhance accuracy, reduce subjectivity, and increase throughput of leaf hair quantification. This ResNet CNN-based approach for leaf hair quantification enables better evaluation and helps to implement this morphological trait in grapevine breeding as a natural physical barrier.

Materials and methods

Plant material

A biparental F1 population of 496 progenies derived by the cross between ‘Morio Muskat’ (‘Silvaner Grün’ × ‘Muscat à Petits Grains Blancs’) and COxGT2 (Vitis coignetiae x ‘Gewürztraminer’), exhibiting quantitative segregation for the trait leaf hair density were utilized for image capturing. In addition, three non-hairy genotypes, ‘Riesling’, ’Regent’ and ‘Cabernet Sauvignon’ and three hairy genotypes, ‘Pinot Meunier’, ‘Tigvoasa’, and a V. thunbergii x V. vinifera hybrid, were also included in the image capturing process. The individuals of the cross population were grown in pots and maintained under greenhouse conditions at Julius Kühn-Institut (JKI), Siebeldingen, Germany. The six additional genotypes were grown under field conditions. The grapevine material used was taken from the own grapevine ex situ repository at JKI, Siebeldingen, in accordance with the relevant regulations and legislation (for material details see Supplementary Table 1; https://www.vivc.de).

Experimental design

Leaf disc preparation and automated image acquisition

A single leaf was sampled from the third or fourth node below the shoot apex from each individual plant. Four leaf discs with 1 cm diameter were excised from each collected leaf using a cork borer. Leaf discs were placed on 1% water agar (Insula Agar Agar Pulver E406, Gustav Essig GmbH & Co. KG, Mannheim, Germany) in 245 mm Square BioAssay Dishes (Corning®, Corning, New York, NY, USA) according to a specific layout template, ensuring the leaf’s adaxial (upper surface) on the agar side (Fig. 1A).

Fig. 1
figure 1

Leaf disc imaging. The layout template for leaf discs (A) combined with Zeiss® Axio Zoom stereomicroscope (B) supported by a motorized table allows high-throughput image acquisition.

Full size image

Images of leaf discs were captured using the Zeiss® Axio Zoom V16 stereomicroscope (Jena, Germany) coupled with a motorized table, enabling movement in the X and Y-axis (Fig. 1B). This fulfills the objective of high throughput with 96 discs/13min. Images were captured at 10.5 fold magnification with a 0.5x magnification lens (PlanApo Z 0.5×/0.125; Free Working Distance 114 mm), contributing to overall system magnification. For sample illumination, a combined light system consisting of a K-LED segmentable ring light, two gooseneck lights at a 45° angle, and a backlight provide a combined exposure time of 18–20 ms/image. JPG was used as the default image format with an image resolution of 6.08 mega pixels (size = 2752 × 2208) and a width/height ratio of 1.25. The ZenBlue version 3.4 (Zeiss, Jena, Germany) program was used to produce a movement template that enabled automated access to the 96 positions. In addition to the motorized table, the Zeiss Axio Zoom (Jena, Germany) V16 has a software autofocus that was utilized to determine the best focal height for each leaf disc prior to imaging. The lowest aperture feasible was used to get the most significant focal depth. For each leaf disc, a single picture was captured and uploaded to the Linux-based server for the model training. A complete detailed and installable workflow can be found with the protocol in GitHub (https://github.com/1708nagarjun/ResNet-CNN-Leaf-hair) (Zeiss Axiozoom users will be able to integrate the template easily in few steps).

OIV-based manual evaluation of leaf hair density

The density of abaxial (lower surface) leaf hair present on the leaf discs was visually evaluated using the established OIV descriptor 08437. It was used to categorize the leaf hair density in five classes: Class 1: None/Very low; class 3: Low 5: Medium 7: High 9: Very high density of leaf hair. Each leaf disc image captured was assigned to the respective class (Fig. 2). All the leaf disc images captured from the F1 individuals of the cross ‘Morio Muskat’ x COxGT2 were classified according to the OIV descriptor 084.

Fig. 2
figure 2

Classification scale for leaf hair density. Exemplary leaf discs of F1 individuals of ‘Morio Muskat’ x COxGT2 assigned to the five classes of the OIV descriptor 084.

Full size image

ResNet CNN model development

Data preprocessing

All the leaf disc images captured from the F1 individuals of the cross ‘Morio Muskat’ x COxGT2 were classified according to the OIV descriptor 084. Then a set of 20 leaf disc images, with four of them representing each of the five OIV 084 classes, was chosen to form the training and testing set for the CNN. Each leaf disc image (Fig. 3A) was sliced into 506 individual image slices with a resolution of 119 × 100 pixels (Fig. 3C) using a Python-based script (chop_images.py), out of which 75% were used as part of the training set, and the remaining 25% were used to form the testing set (Fig. 3B). To enhance the model’s robustness and prevent overfitting, data augmentation was performed by rotating the image slices and incorporating images captured under varying lighting conditions on different days and experiments. All the image slices were manually classified by two experts into the three classes: background (agar), leaf without hair, and leaf with hair using a Python script called image_sorter2_script.py (GitHub repository).

Fig. 3
figure 3

Image preparation for CNN. Start with images of one leaf disc (A). Each image was divided into four equal portions, ¾ parts used for training (with bold borders) and ¼ (dotted border) used for testing the model (B). As input image for the CNN model training served 506 individual image slices (C).

Full size image

Model building

To improve accuracy, two sequential CNNs were developed instead of a single multi-class model. The first CNN classifies regions as “background” or “leaf disc,” while the second classifies the leaf disc as “with hair” or “without hair.” This stepwise approach simplifies the task, enhances accuracy, and reduces complexity. To avoid any subjective errors during manual labelling, the data for testing and validation were separated systematically, ensuring that the background vs. leaf classification and leaf with vs. without hair classifications were handled independently. This approach minimizes the risk of high error rates typically associated with multi-class models that attempt to classify all three categories simultaneously. A ResNet-based CNN architecture was chosen for both the CNNs used in this pipeline (Fig. 4). The first ResNet CNN has the classes “background” and “leaf disc”, from now on referred to as CNN1. It was trained with 10,089 individual image slices as part of the training set and 4,323 individual image slices as part of the validation set. The second ResNet CNN has the classes “leaf with no hair” and “leaf with hair,” from now on referred to as CNN2. It was trained with 4,369 image slices as part of the training set and 1,872 image slices as part of the validation set (Fig. 3B). All datasets and the code to train the CNNs are available in the GitHub repository. The CNNs were trained in Jupyter Notebook, on Google Colaboratory instances running Ubuntu 18.04.5 LTS, a 2-core Xeon CPU at 2.20 GHz, 13.3GB RAM and an Nvidia Tesla K80 GPU with 12GB memory. All CNN calculations were performed on the GPU.

The detailed summary of the CNN layers can be found on the GitHub repository (hair-nohair-model-summary.txt; Fig. 4). For CNN1, each slice in the size of (112 × 112) was loaded with RGB channels. A rescaling layer was used to scale the RGB values lying between 0 and 255 down to 0 and 1, followed by application of two blocks consisting of a 2D convolution layer, a batch normalization layer, followed by an activation layer using ReLU. The convolution layer size for block one was 32 and for block two was 64. After the first two blocks, the residual was set aside for future use. Four more blocks were added with separable 2D convolutions of sizes 128, 256, 512, and 728. Each block consists of an activation layer (ReLU), a separable 2D convolution layer, a batch normalization layer, another activation layer, another separable 2D convolution layer, another batch normalization layer, a 2D max pooling layer, and a final 2D convolution layer with the previous block residual as input. This result is then added back to the current values, which is again set aside for use in the next block. After the four blocks above, another block consisting of a separable 2D convolution layer of size 1024, a batch normalization layer, and an activation layer are added, the result of which is passed through a global 2D average pooling layer, followed by a dropout layer with rate 0.5. Finally, the regular deeply connected dense layer with 1 unit and sigmoid activation followed to get an output value between 0 and 1 where 0 is “background” class and 1 is “leaf disc” class (Fig. 4).

For CNN2, the same model architecture was used as for CNN1, but the dropout rate was changed to 0.2 (Fig. 4). For both the CNNs, the Adam optimizer with a learning rate of 0.001 and loss calculation was applied with the binary_crossentropy function. Both CNNs were trained for 30 epochs each.

Optimizing parameters for the CNNs

Prototyping was used to determine the dropout rate, number of ResNet blocks, and optimizer settings. Multiple variations of the model were trained on the same dataset and compared to the results with the ground truth data derived by the two independent experts who classified each individual slice manually. During training, the models were checkpointed at each epoch, and the epochs with the smallest difference between training and validation accuracy were compared to the ground truth findings. The final image datasets and the Jupyter Notebook containing the code necessary for training of the final version of the CNNs are available on the GitHub repository.

Fig. 4
figure 4

Detailed representation of the two ResNet CNNs architectures. CNN1: Background vs. leaf. CNN2: Leaf without hair vs. leaf with hair.

Full size image

Performance evaluation

Ground truth: model validation by experts

To evaluate the performance accuracy of the model, 20 new leaf disc images (4 images per OIV 084 class) from F1 individuals of the cross ‘Morio Muskat’ x COxGT2 were analyzed using the ResNet-based CNN model. Simultaneously, the two experts categorized the individual image slices (10,120) into the three classes “background”, “leaf with hair” and “leaf without hair”. Additionally, they also visually rated the whole leaf disc images for percentage of leaf area covered with hair. Thus, two experts classified each leaf disc image at the slice level and manual rating (0–100%). The output results of the ResNet CNN (%) were correlated (Pearson correlation) with the ground truth data generated by manual sorting of slices (%), and manual rating of the whole leaf disc image (%) by expert 1 and expert 2, respectively.

Comparison between experts and non-experts

A set of three genotypes with non-hairy leaves (‘Riesling’, ’Regent’, ‘Cabernet Sauvignon’) and three genotypes with hairy leaves (‘Pinot Meunier’, ‘Tigvoasa’, and V. thunbergii x V. vinifera) were used for comparing the variability between two experts and two non-experts. The four evaluators each generated ground truth data by manually sorting individual image slices (3,036) and visually rating the whole leaf disc images (0–100% leaf hair density). Considering the ResNet CNN output (%) as true values and evaluator values as measured, absolute accuracy error (absolute error) was calculated38 for all the individual images.

Audience validation

Manual rating (0–100% leaf hair density) of the leaf disc images of the genotypes with non-hairy leaves (‘Riesling’, ’Regent’, ‘Cabernet Sauvignon’) and the three genotypes with hairy leaves (‘Pinot Meunier’, ‘Tigvoasa’, V. thunbergii x V. vinifera) was carried out by 16 inexperienced evaluators. The leaf disc images were displayed using a projector in a meeting room with all 16 evaluators being present at the same time. A detailed explanation of manual rating was provided by an expert displaying three trial leaf disc images, which varied in leaf hair density (0%, 55%, and 99%, respectively). Absolute accuracy errors were calculated for each individual image between the inexperienced evaluators, to compare and reveal high variability caused due to subjectivity.

Results

Model training and testing results

To improve the phenotyping of the leaf hair density on the lower side of grapevine leaves, automated high-throughput imaging of leaf discs was first established. These images were then used to train and later validate a ResNet model, consisting of two CNNs. The first ResNet CNN “background” vs. “leaf disc” (CNN1) achieved an overall validation accuracy of 98% with a minor value of validation loss after 30 epochs (Fig. 5A). Whereas, the second ResNet CNN “leaf with no hair” vs. “leaf with hair” (CNN2) achieved an overall validation accuracy of 95% and validation loss of 13% after 30 epochs (Fig. 5B). CNN1 has a higher model validation accuracy compared to CNN2 due to the reduced feature complexity of the input slices. Therefore, the CNN2 model was more comprehensive in extracting the features resulting in slightly reduced validation accuracy. However, no overfitting or underfitting of the models was observed and both models achieved satisfying overall validation accuracy.

Fig. 5
figure 5

CNN accuracy. Training and validation accuracy, training and validation loss performances of CNN1 “background” vs. “leaf disc” (A) and CNN2 “leaf without hair” vs. “leaf with hair” (B) after 30 epochs.

Full size image

Ground truth: experts validation

Once the training and testing of the ResNet CNNs was completed, the models were deployed to validate 20 (10,120 image slices) new leaf disc images (4 images per OIV 084 class) (Supplementary Fig. 1). These images were obtained from completely independent experiments conducted on separate days or months, ensuring that each image/image slice for each class originated from a distinct experimental setup. This approach strengthens the robustness and generalizability of the model. Again the two experts independently evaluated the identical set of images by classifying the slices and manually rated the leaf discs for the trait leaf hair (0–100%), hence generating the ground truth data. The independent ground truth datasets generated by the two experts are highly comparable to each other (Supplementary Table 2). The ResNet CNN output data (%) for the 20 images is in close agreement with the ground truth data generated by the two experts. In addition, the ResNet CNN output data (%) appeared to be extremely comparable to the slice sorting data in both experts’ evaluations, demonstrating the accuracy and sensitivity of the model. There are, however, a few inconsistencies such as image 1 in OIV 084 class 9 (Supplementary Fig. 1) being evaluated lower by the ResNet CNN than by expert 1 due to poor image features, uneven lighting, and image quality. The correlation between expert slice classification (%) for both experts and the ResNet CNN output data showed highly significant R values of 0.98 (p < 0.001) and 0.92 (p < 0.001), respectively, and Root Mean Square Error (RMSE) values of 8.20% and 14.18%, respectively (Fig. 6A). Whereas a slightly less significant correlation with R values of 0.86 (p < 0.001) and 0.87 (p < 0.001) and with RMSE of 21.43 and 20.37 were obtained between manual rating (%) of the experts and the ResNet CNN output data (%), respectively (Fig. 6B).

Overall, the correlation in the classification at slice level proved to be more precise than the manual rating correlation of both experts. There was a higher degree of subjectivity between the manual rating (%) in comparison to slice level classification (%) in the results of both experts, indicating the necessity for an objective classification model (Fig. 6). The correlation coefficient varied the most for leaf discs densely covered with leaf hair (70–90%), affecting the overall correlation. Therefore, it is crucial to use objective machine vision to precisely define quantitatively segregating features such as leaf hair. The correlation of ResNet CNN (%) and expert slice classification (%) for both experts (i.e., R = 0.98 and R = 0.92, respectively) clearly demonstrate the model’s accuracy and performance.

Comparison between experts and non-experts

The ResNet CNN model was used to quantify the leaf hair of three genotypes with hairy leaves (‘Pinot Meunier’, ‘Tigvoasa’ and V. thunbergii x V. vinifera) and three genotypes with non-hairy leaves (‘Riesling’, ’Regent’ and ‘Cabernet Sauvignon’) (Supplementary Table 3), followed by absolute accuracy error (absolute error) estimation for the individual leaf disc images. The ResNet CNN output data (%) was considered as actual values and evaluator values, generated by a panel of two experts and two non-experts (Supplementary Table 4), as estimated. For the genotypes assessed, the absolute error variance was found to be highly negligible between the two experts for both the slice (%) and manual (%) ratings (Fig. 7A, A1 and A2).

Fig. 6
figure 6

Validation 1 of experts vs. ResNet CNN. Correlation between two experts slice classification (%) vs. ResNet CNN evaluation (%) (A) and expert’s manual rating (%) vs. ResNet CNN evaluation (%) (B) of leaf disc images.

Full size image
Fig. 7
figure 7

Validation 2 of experts vs. non-experts. Absolute error estimation of experts (A) and non-experts (B) for slice (%) and manual (%) rating, respectively. The red dotted (“0”) line represents the no error line. Filled circles represent evaluator 1 and filled triangles represent evaluator 2.

Full size image

The experts demonstrated greater precision and repeatability in the assessments performed. In the case of both non-experts, the absolute error variance ranged between 15% and 30% for slice classification (%), indicating a strong tendency to overestimate. In the manual rating (%), however, both the non-experts underestimated the leaf hair quantification, with the absolute error variance ranging between -20% to -50% (Fig. 7B, B1). Surprisingly, the three genotypes with hairy leaves were underestimated, showing that manual ratings by non-experts might introduce a significant bias (Fig. 7B, B2). These findings clearly stipulate the need for a technique for quantifying leaf hair that is objective, accurate, and precise.

Audience validation

In the absence of a ResNet CNN classifier, untrained/novice evaluators tend to overestimate the non-hairy genotypes and underestimate the hairy genotypes (Fig. 8, Supplement Table 5). The overall significant bias in terms of absolute error was found to be between 0 and 30% for genotypes with non-hairy leaves, and 5 to 60% for genotypes with hairy leaves. Furthermore, it was obvious that the intermediate genotypes, V. thunbergii x V. vinifera and ‘Tigvoasa’, were challenging to rate (Supplement Table 6).

Fig. 8
figure 8

Validation 3 by audience. Absolute errors (Absolute Error = Measured Value - Actual Value) estimation of manual rating for a panel of 16 novice evaluators. Each dot represents the result of a single evaluation; the red dotted (“0”) line represents the no error line.

Full size image

Leaf disc classification pipeline and command line tool

The whole process of automated leaf hair evaluation was organized into a single command line tool that runs on a UNIX environment. The tool generates an output in two formats, one in the form of a csv files, which aggregate all the image results in the test, the other in the form of jpg images showing slice maps of the original input images for better visualization of the area covered with leaf hairs. The whole pipeline is publicly available on the open source repository GitHub (https://github.com/1708nagarjun/ResNet-CNN-Leaf-hair).

First, four leaf discs are punched out of a leaf, imaged under a microscope and used as input data (refer methods section for details). The images are sliced into 506 slices, which are then classified by CNN1 into the two classes “background” or “leaf disc”. The leaf disc slices are further classified by CNN2 into the classes “leaf disc without hair” and “leaf disc with hair”. The final percentage of hair on the leaf disc is calculated by taking the proportion of slices classified as “leaf disc with hair” relative to the total number of slices classified as “leaf disc” by CNN2, ensuring a robust and accurate quantification of hair density. For each leaf image, the Python-based code (classifier.py) systematically creates 506 slices and saves their coordinates relative to the original image. After classification, the code rearranges the slices into their original order to generate a slice map (Fig. 9), which visually represents the classifications across the entire leaf disc. This process preserves spatial context, ensuring that no data about the distribution of hair density is lost. The slice map also enables visual verification of localized variations, providing a clear overview of the results while maintaining accuracy and scalability. The pipeline runs in a loop over every single image present in the input directory. After classification, the tool generates slice maps that overlay the input images with red boxes outlining the slices containing leaf disc area covered with hairs and blue boxes outlining the slices that contain the background. Finally, the results are summarized and saved in a csv file that contains the name of the image, number of slices with background, number of slices with hairs, number of slices with no hairs, percentage of leaf disc covered with hairs, and percentage of leaf disc not covered by hairs. The generated files, slice of the slice maps and result tables are time-stamped and stored in the output directory (Fig. 9). Detailed instructions about the leaf hair quantification tool are provided in the GitHub repository.

Fig. 9
figure 9

Outline of the ResNet-based leaf hair quantification pipeline. The final output is a table with percentage values and a slice map (.jpg), where the individual slices with leaf hairs are annotated in red. (img name: image name; back: slices identified as background/agar; hair: slices identified as hair contained; nohair: slices identified as no hair contained; perc_hair: area of leaf disc covered with hair (%); perc_nohair: area of leaf disc not covered with hair (%)).

Full size image

Discussion

Quantifying plant functional traits using advanced imaging technology offers breeders a more efficient and progressive approach to assist their decision-making. To uncover the hidden genetics of quantitative traits, the availability of accurate and objective measurement methods is crucial in research. A recent trend in machine learning, such as CNNs, has been widely utilized in plant disease identification39,40,41,42. CNNs combined with digital images are a powerful and cost-effective method for monitoring and maximizing agricultural products43,44. Image analysis has already been demonstrated to be more accurate than visual estimation in a number of pathosystems45,46,47,48. CNNs have also been successfully implemented in the efficient quantification of grapevine diseases such as powdery mildew and downy mildew, which significantly improved the trait analysis11,32,34. Numerous convolutional neural network (CNN) architectures, including VGG (Visual Geometry Group), Inception (GoogLeNet), DenseNet (Densely Connected Convolutional Networks), MobileNet, and ResNet (Residual Networks), have been meticulously developed and persistently advocated for their application in plant disease identification tasks49,50,51. In general, ResNet stands as an excellent choice for plant disease recognition due to its ability to efficiently capture intricate patterns in plant images while mitigating the degradation problem often encountered in deeper networks, thus ensuring high classification accuracy and robust performance in real-world applications. Its state-of-the-art performance, evidenced by numerous empirical studies, underscores ResNet’s superiority in handling the complexities inherent in plant disease classification tasks52,53,54. In this study, modified ResNet CNN was trained to quantify leaf hairiness on grapevine leaf discs using a minimalistic set of RGB images and computational resources. ResNet’s robust residual connections enable accurate feature extraction and generalization, making it well-suited for this task, particularly in laboratory-based studies requiring precise classification. The ResNet CNN tool achieved an overall model accuracy of 95% and validation loss of 13%. It was used in a real-world application and demonstrated its ability to recognize and quantify leaf hair. The evaluation performance of the proposed ResNet CNN model in comparison with the ground truth data generated by two experts showed a significant correlation of R = 0.98 and R = 0.92, respectively (Fig. 6). Furthermore, the comparison of the manual scoring of the experts according to the OIV descriptor 084 with the ResNet CNN results also showed a high level of agreement (Fig. 6). This implies the ability of the ResNet CNN to learn complex features from RGB images and to transfer it with slightly negligible decreased accuracy. In addition, the validation with a complex practical case, images of leaf discs from six varieties with different genetic backgrounds, showed the high performance of the ResNet CNN model. The true values of the ResNet CNN model showed only a minimal deviation from the assessments of the experts (Fig. 7A). In contrast, significant deviations from true values were detected in the non-expert data, resulting in over- and underestimations and affecting the overall phenotyping accuracy (Fig. 7B). This observation was also confirmed by an audience-based validation performed to determine the bias novice evaluators can cause. Varieties with hairless leaves were clearly overestimated, varieties with hairy leaves underestimated, and considerable inconsistent deviations were found between the evaluators (Fig. 8). The data provided clearly suggest that for a quantitative trait like leaf hair, the competence of a trained expert is a prerequisite to get objective phenotypic data. Because the evaluations of non-experts and beginners contribute to a high variance, subsequent analysis such as QTL or GWAS (genome-wide association study) are also strongly affected. This indicates the necessity to implement machine-based vision for objective quantification of leaf hair. In comparison to established classical approaches, the developed ResNet CNN model has outperformed the novice and competent evaluators in leaf hair quantification for different cultivars. While this study focused on developing a robust and accurate tool for quantifying leaf hair density, we recognize the potential biological implications of this trait in processes such as disease resistance, pest deterrence, or other agronomic traits. This model provides a robust foundation for researchers aiming to explore and establish connections between leaf hair density and related traits.

Limitation of work and future perspective

The trained ResNet CNN demonstrated high accuracy and efficiency; however, the first CNN model (CNN1) is specifically designed for an agar background. While it is true that some laboratories might use tinted or colored agar, this is a rare exception. The majority of studies, including ours, use neutral-colored agar, which is the standard in laboratory-based assays. This ensures compatibility with our model and aligns with commonly accepted practices. Any variation in background color or material other than agar may affect the overall accuracy and lead to discrepancies in quantification. In the presented approach, the leaf hairs are quantified on leaf discs and not on whole leaves, but the results are on average consistent with those of whole leaves. Small leaf discs are advantageous, since leaf hairs are small in structure and capturing images of the whole leaf is more difficult. Lower resolution or lower quality images could affect model performance and contribute to bias. Nevertheless, leaf hair on grapevine leaves are uniformly distributed; three to four leaf discs from a single leaf should correspond to the overall percentage of hair coverage. This study offers a simple and efficient approach to quantifying leaf hair density using ResNet-based CNN analysis of leaf disc images. Unlike segmentation methods, which may introduce bias or require microscopic images for densely haired varieties, this model provides a robust and objective solution for reliable phenotyping. Yet, categorical classification approach can be the future work. In addition, segmentation or semantic labelling of different types of leaf hair examined on the leaf disc is a surpassing method of categorical classification.

Conclusion

Accurate and precise phenotyping of quantitative traits such as leaf hair is challenging and laborious. This study provides an automated, valuable tool that will pave the way for an efficient, accurate and reliable quantification of leaf hair. The potential of the ResNet CNN was shown in real-world applications, reducing bias and increasing the reliability of the quantification compared to the results of untrained evaluators. It is highly suggested for quantification of leaf hair in the Vitis genus, including disease resistance, fungicide administration, and insect diversity studies. Thus, the ResNet CNN model is reliable, promising and significantly more advantageous because of its objectified machine vision for evaluating grapevine leaf hair. This tool can be utilized in screening of germplasm repositories and serves as a promising application in generating phenotyping data utilized in QTL and GWAS studies. We are convinced that the ResNet CNN model presented for grapevine leaves can be transferred to other crops as well as other parts of plants after adaption enabling the effective exploration of leaf hair dynamics across different plant species. CNN-based methods in general will become an indispensable tool for plant scientists around the world in the future, clearly improving and accelerating the way we phenotype plant traits.