香京julia种子在线播放

A written informed consent was obtained from the individual for the publication of any potentially identifiable images or data included in this article.

Two experiments were conducted under controlled conditions in a growth room at the Asociación de Productores de Oleaginosas y Trigo (ANAPO) research facility in Santa Cruz de la Sierra, Bolivia. Wheat cultivars were planted in pots of 15 cm diameter, filled with vermicast:silt (3:1 [v/v]), and grown at 18−25°C, 14 h light/10 h dark photoperiod, and 50–60% relative humidity. Plants were fertilized, and insecticides were sprayed when needed. Plants were arranged in a randomized complete block design with wheat cultivars having various levels of resistance to MoT, two inoculation levels (inoculated and non-inoculated), and four replicates. Wheat cultivars with a range of sensitivity to the wheat blast were used for the experiments. Experiment one included Bobwhite and South American spring cultivars Atlax, BR-18, Motacú, Urubó, AN-120, Sossego, and San Pablo and for experiment two the cultivars included BR-18, San Pablo, Bobwhite, and Atlax (Baldelomar et al., 2015; Fernández-Campos et al., 2020).

Plants were inoculated at the growth-stage Feekes 10.5, when the spike had completely emerged, with MoT isolate 008-C (Figure 1A), according to a modified inoculation protocol previously published (Cruz et al., 2016a). A conidial suspension was adjusted to 20,000 spores/ml, and each spike received 1 ml of the spore suspension. Immediately after the spikes were sprayed with the MoT inoculum, plants were moved to a dew chamber (Figure 1B) to induce MoT infection (i.e., 24–26°C, 95–98% RH, and 14 h light photoperiod). Forty-eight hours after inoculation, plants were removed from the dew chamber and left under controlled environment room conditions [(24–26°C and relative humidity of 50–60%), until day 19 after inoculation; Figure 1B].

Following phytopathometry terminology, we used the term “estimate” for visual disease estimations made by humans and the term “measurement” for estimations made by image analysis (Bock et al., 2010; Gongora-Canul et al., 2020). Visual estimate of DS was obtained by observing the disease area covered in the spike and assigned a corresponding severity value from 0 to 100%. In this study, image analysis disease measurements were achieved by manually measuring spike disease area (pixels) using RGB color threshold segmentation with the image analysis software Fiji ImageJ v.1.52a (Schindelin et al., 2012; Sibiya and Sumbwanyambe, 2019). First, the measurement of the total spike area was obtained, then the diseased area was measured. Finally, the percentage of diseased severity (DS) of the individual spike was calculated (Equation 1), where A_Diseased is the proportion of the area of spike that is diseased divided by the total area of the spike A_Total (See Video 1 in Supplementary Material).

Visual estimations of wheat spike blast symptoms were taken seven times after inoculation in each experiment. In experiment one, visual estimations and images were collected 4, 6, 9, 12, 14, 16, and 19 days after inoculation (DAI) and in experiment two, 0, 5, 7, 10, 12, 14, and 19 DAI. Each spike side (four sides total) was visually estimated for DS by Rater 1 (a plant pathologist with experience on wheat blast, rice blast, and other diseases). Simultaneously, an image from each spike side was captured perpendicular to the spike with a distance of 50 cm approximately with a DSLR EOS 6D Canon camera (Canon Inc., Tokyo, Japan) (Figure 1D) using a photography studio set up with umbrellas, lights, and screens (Neewer 2.6 m × 3 m/8.5 ft × 10 ft Background Support System and 800 W 5,500 K Umbrellas Softbox Continuous Lighting Kit for Photo Studio Product) that helped create a uniform light and smooth environment (Figure 1C).

The total spike disease estimations of Rater 1 paired with the corresponding image were converted to a three-category scale according to the amount of severity that served to fed training and testing dataset of CNN model. The category selection was based on wheat blast results from published work conducted over the last decade (Baldelomar et al., 2015; Cruz et al., 2016b; Vales et al., 2018; Cruppe et al., 2020; Fernández-Campos et al., 2020). Category 1 (healthy spikes) was used as a baseline (i.e., negative control or fully immune). Category 2 showed 0.1–20% severity (resistant and moderately resistant/low levels of symptoms) corresponding to the selected putative population for successive trials under variable conditions (controlled environment or field). Category 3 showed 20.1–100% severity (moderately susceptible and susceptible/intermediate and high levels of symptoms) corresponding to the plant population that will not be selected for successive trials because of the high potential to be or to become susceptible to the disease studied (Figure 2).

Rater 1 played a critical role in estimating DS and classifying into categories of all the images belonging to Dataset 1 and Dataset 2 (Datasets are described in the section, generation of data sets according to wheat spike physiological changes). Therefore, an inter-rater analysis was needed to determine the reliability of visual estimations of Rater 1. Inter-rater agreement assesses the degree of agreement between two and more raters who obtain independent ratings about the characteristics of a set of subjects. Subjects of interest include people, things, or events that are rated (Madden et al., 2007; Everitt and Anders, 2010; Bock et al., 2020).

To determine the agreement of disease estimations of Rater 1, we performed an inter-rater analysis including a second-rater, and ImageJ was used as an image analysis software baseline. Rater 2, is a plant pathologist and expert in the wheat blast. ImageJ is an image analysis software used to measure plant diseases from images.

We used the power analysis Wilcoxon signed-rank test to determine the sample size for the inter-rater agreement studies of the two training datasets. The test consisted of DS estimations or measurements of 31 and 29 images from the CNN training Dataset 1 and training Dataset 2, respectively. From now on, the 31 images selected from Dataset 1 will be called sample Dataset 1 and the 29 images from Dataset 2 will be referred to as sample Dataset 2. Rater 2, who is an experienced researcher with more than 4 years of working with the wheat blast disease, visually estimated DS from the sample Dataset 1 and Dataset 2. Additionally, disease measurements were obtained from the sample Dataset 1 and the sample Dataset 2 using ImageJ software as indicated above. Ultimately, the DS results of visual disease estimations of human raters and ImageJ measurements were compared. The estimated and measured DS values from both samples were analyzed for inter-rater agreement in two scenarios, one with a scale of 0–100% DS (continuous data), and the other with the images divided into three categories of DS (ordinal data). We, therefore, computed Lin's Concordance Coefficient, Fleiss kappa, and weighted kappa statistics.

The Lin's concordance coefficient (ρc or CCC) is used to estimate the accuracy₁ between two raters using continuous data. From the analysis, we obtained the estimation of accuracy₁, precision₁, and bias of the disease estimations and disease measurements between the two raters (Lin, 1989; Madden et al., 2007; Bock et al., 2010). For accuracy₁ (ρc) and precision₁ (r), values range from 0 to 1; values close to 1 indicate high accuracy₁ and precision₁. Bias (Cb) ranges from 0 to 1, and values close to 1 indicate less bias (Nita et al., 2003). Lin's concordance analysis was performed by using PROG REG ALL procedure on SAS v.9.4 (Cary, NC), based on the macro developed by Lawrence Lin and verified by Min Yang (Lin et al., 2002).

To determine the degree of association between the estimation of categorical information provided by the two raters (inter-rater agreement), the weighted kappa statistics were computed (Chmura, 1992; Graham and Jackson, 1993; Nelson and Edwards, 2015). The Fleiss kappa coefficient was used to compare the agreement of categorical information among all raters, (i.e., Rater 1, Rater 2, and ImageJ) (Fleiss et al., 2003). The values of both the weighted kappa and Fleiss kappa coefficients range from 0 to 1. Values from 0.5 to 1 indicate that the agreement is better than what is expected by chance (Nelson and Edwards, 2015; Tang et al., 2015; Mitani et al., 2017; Gamer et al., 2019). The Fleiss kappa statistics and weighted kappa were computed with the irr package of the R software (Team, 2017).

Wheat was inoculated at the growth-stage Feekes 10.5 (spike completely emerged) of the host plant. Approximately every 2 days after the inoculation, the spike images were collected to capture the changes developed. Indirectly, progressive physiological changes in spikes were recorded, as maturing begins at wheat growth-stage Feekes10.5.4 (kernels watery ripe) and continues through the growth-stage Feekes 11.4 (mature kernels) (Large, 1954; Wise et al., 2011). During this period, the kernel hardened, and the green spike lose its color (maturing), which mimic the typically bleached spikes caused by wheat spike blast symptoms.

Two datasets were generated considering the (color) physiological changes that can lead to confusion when training the CNN model. Dataset 1, included maturing and non-matured wheat spikes; and Dataset 2 included only non-matured spikes (data available at: https://purr.purdue.edu/publications/3772/1). The proposed CNN model was trained using the two datasets. Each dataset was randomly separated into the training and testing datasets. The CNN model automatically extracted the features of each image in the training dataset to learn a good classifier, whereas the testing dataset was used to evaluate the performance of the trained CNN model. In general, an unseen dataset was applied to evaluate the CNN model to ensure that the model was not under-fitting or over-fitting. In this research, 80% of the images were categorized and used as the training set, and the remaining 20% as the testing set. Table 1 lists the original distribution of the number of images in Dataset 1 and Dataset 2. Although Category 3 covers a large variability, it does not mean the number of the data in Category 3 is larger than the other two categories. The number of images in each category was extremely imbalanced and using them indiscriminately could have resulted in a biased model. Fortunately, there are several viable methods to cope with the disproportionate training data in each category.

Data augmentation is a common technique providing a viable solution to data shortage issues by adding copies of original images with modification or noise (Boulent et al., 2019). Data augmentation was used in this study to balance the number of images in each category. In this study, images were randomly flipped horizontally and vertically in order to increase the number of images in Categories 2 and 3. Thus, for Dataset 1, training data were triplicated in Category 2 and quadrupled in Category 3 (Table 1). For Dataset 2, training data were quadrupled in Category 2 and quintupled in Category 3 (Table 1).

In recent years, the feasibility of using artificial intelligence, in particular deep learning, has been expanded into a variety of applications (Atha and Jahanshahi, 2018; Chen and Jahanshahi, 2018; Kumar et al., 2018; Wu and Jahanshahi, 2019). Deep learning is a subset of machine learning that enables computers to automatically extract features from a huge amount of data and learn to classify data.

In this study, wheat spike blast symptoms were automatically detected and classified into three severity categories using a pre-trained CNN model. This model may be more efficient than classifying images visually. To obtain a general and reliable CNN model, the network needed to be trained using a large labeled training dataset. The performance of the CNN model is highly dependent on the number and quality of the training data. However, it was hard to collect a wheat blast dataset having a million images in a short time. The performance CNN model can easily lead to under- or over-fitting due to the lack of a large dataset for training. To address this issue, transfer learning was used as a practical solution where a network was trained using a typically different larger dataset such as ImageNet. A major advantage of using transfer learning is that it can adapt the parameters trained from an abundant number of images. Transfer learning starts with a pre-trained model, e.g., VGG16 model, and replaces the fully-connected (FC) layers of the model with new FC layers. A network trained on the ImageNet dataset was used to initialize the network parameters, and the whole network was fine-tuned since the nature of our dataset was very different from the ImageNet dataset. In this study, an FC layer that consisted of three nodes, representing three categories, were appended to the end of the network. A residual neural network architecture (ResNet101), a CNN model with 101 layers with recurrent connection trained on ImageNet data (He et al., 2015, 2016), was selected as the pre-trained model. Furthermore, as shown in Table 1, it was extremely difficult to obtain a large number of images in each category. An unbalanced dataset can result in a biased CNN model. To address this issue in the dataset, the loss function, which was used to optimize the parameter in a neural network, was transformed into a weighted loss function (Equation 2) by assigning individual weights to each category. Equation (2) defines the cross-entropy loss function in the CNN model, where ω_category is the assigned weight to each of the categories, the first term in Equation (2) is a negative log-likelihood loss, and the second term in Equation (2) is log-softmax. Four cases of study were tested with an individual weight set to the loss functions assigned to different categories. In the experiments, “cases” refer to specific combinations of weight loss functions for each of the three DS categories (Table 2). Case 1 was the non-weight set [1, 1, 1], with all categories sharing the same class weight. Case 2 used [1, 10, 1] class weights in the loss function, meaning that the highest weight was for Category 2, which includes plants at early disease stages and low levels of disease symptoms. Case 3 used [2, 5, 1] class weights in the loss function, meaning that the higher weight was assigned to Categories 1 (no symptoms) and 2 (early stages and low levels of disease symptoms). Case 4 had class weights [2, 1, 1] in the loss function, assigning a higher weight to category 1 (no symptoms) (Table 2).

The network was trained for 15 epochs using a stochastic gradient descent optimizer (Bottou, 2010), a learning rate of 0.0001 was used, and the batch size was 16. Additionally, 5-folds cross-validation was applied to the training process. The training took place on a Linux server with Ubuntu 14.04. The server included two Intel Xeon E5-2620 v4 CPUs, 256-GB DDR4 memories, and four NVIDIA Titan X Pascal GPUs. Pytorch (Paszke et al., 2017) was used to implement the CNN.

The performance of the CNN model was evaluated via the classified results of the testing dataset. A 3 × 3 confusion matrix was used to describe the prediction result of the model. Each row of the confusion matrix represented the ground truth of the data, and each matrix column corresponded to a predicted category by the CNN model. Thus, the diagonal elements of the matrix, called true positive (TP), were the number of wheat images correctly classified into the ground truth. The false positive (FP) for each Category was the sum of all errors in that column. For example, the FP of Category 1 was the number of Category 2 and Category 3 severities that were incorrectly classified as Category 1. Based on the confusion matrix, additional evaluation metrics were calculated.

Accuracy₂ was defined as the total number of TP among three categories divided by the total number of the predictions. Precision₂ was defined as the total number of the TP instances divided by the total number of predicted positive examples, which was the summation of TP and FP instances in the binary classification task (Equation 3). Similarly, the precision₂ of the multi-classes task illustrates the number of instances that were correctly predicted given all the predicted labels for a given category. Recall was defined as the TP instance divided by all the positive samples (TP and FN) (Equation 4). F1 score is a single metric that encompasses both precision₂ and recall (Equation 5). Accuracy₂, precision₂, recall, and F1 score metrics ranged from 0 to 1, where higher values indicate the high predictive ability of the model.

The final wheat spike blast severity was at day 19 after inoculation when cultivar Atlax reached 100% average DS, followed by Bobwhite (99.7%), San Pablo (32.9%), BR-18 (8.7%), Motacú (3.7%), AN-120 (3.31%), Urubó (1.9%), and Sossego (0.83%). Wheat spike blast symptoms developed on all tested cultivars, with reactions to MoT infection consistent with previous reports, except for cultivar San Pablo that showed moderate susceptibility (Baldelomar et al., 2015; Cruz et al., 2016b; Cruppe et al., 2020; Fernández-Campos et al., 2020; Gongora-Canul et al., 2020). Cultivar Atlax exhibited the highest DS of all the cultivars and had a high level of susceptibility to wheat spike blast.

The Lin's concordance correlation analysis showed a high accuracy₁ (ρc = 0.89–0.91), high precision₁ (r = 0.91–0.94), and less bias (Cb = 0.95–0.99) in the sample Dataset 2 than in the sample Dataset 1 (ρc = 0.77–0.85, precision₁ r = 0.80–0.87, and bias Cb = 0.93–0.98) (Figure 3). In the sample Dataset 1, the highest accuracy₁ was between Rater 1 and Rater 2 (ρc = 0.85) and between Rater 1 and ImageJ (ρc = 0.85). In the sample Dataset 2, the highest accuracy₁ value was between Rater 1 and ImageJ (ρc = 0.92), followed by between Rater 1 and Rater 2 (ρc = 0.91). In both sample datasets, strong accuracy₁, high precision₁, and low bias involved Rater 1, providing evidence that ratings of disease based on continuous data were done correctly for further classification of the images into categories for model training.

The weighted kappa statistics (κ), used to quantify inter-rater agreement, were higher in the sample Dataset 1 than in the sample Dataset 2, with κ = 0.72–0.88 (p < 0.01) and κ = 0.78–0.85 (p < 0.01), respectively (Table 3). In the sample Dataset 1, the highest agreement occurred between Rater 1 and ImageJ (κ = 0.88), and in the sample Dataset 2, the highest agreement was between Rater 1 and Rater 2 (κ = 0.85). In both sample datasets, the substantial agreement involved the ground truth (Rater 1), providing evidence that ratings were done correctly for further classification of the images into categories for model training.

The Fleiss kappa coefficient (Fκ), which compared the association of ordinal categorical information of two or more raters, showed an Fκ = 0.771 (n = 31, z = 9.26, p < 0.001) for the sample Dataset 1 and 0.697 (n = 29, z = 8.1, p < 0.001) for the sample Dataset 2, indicating substantial agreement among the human raters and ImageJ in both datasets. However, the sample Dataset 1 possessed a higher Fleiss kappa coefficient index than the sample Dataset 2, both presented substantial agreement between the rates and ImageJ. Yet, the evidence supported the fact that the three raters correctly estimated the amount of the disease from the same image.

To train the proposed CNN model, two different datasets were used. As mentioned above in the section Generation of Datasets According to Wheat Spike Physiological Changes, testing reliability of Rater 1, Dataset 1 included matured and non-matured wheat spikes and Dataset 2 included only non-matured spikes (Table 1). Four cases applied different weight set of loss functions in both Datasets (Table 2, Supplementary Figures 1, 2). The performance of the CNN model was evaluated via the classified result of the testing data.

The testing accuracy₂ of the model trained with Dataset 1 was 90.1% in Case 1, 90.4% in Case 2, 90.0% in Case 3, and 87.7% in Case 4. The testing accuracy₂ of Dataset 2 was 98.4% in Case 1, 93.9% in Case 2, 95.0% in Case 3, and 94.2% in Case 4. Dataset 2 presented higher accuracy₂ values compared to Dataset 1, suggesting that the model was accurate. However, it was not sufficient to claim that the model was reliable based on accuracy₂ alone since the dataset in this study was unbalanced. In addition to accuracy₂, other metrics can help evaluate the performance of the CNN model, such as precision₂, recall, and F₁ score.

Precision₂ indicates the ability to correctly classify an instance in all predicted positive instances. The focus was on the performance of the CNN model in Category 2 as this was the category that breeders and pathologists will concentrate on for breeding purposes. Dataset 1 Case 2 showed the lowest precision₂ (75.4%) among all cases values (Table 4). Moreover, the confusion matrix of Dataset 1 Case 2 showed that the model misclassified 38 images of Category 1 (no symptoms) as Category 2 (early disease stages and low levels of disease symptoms), which was the highest number of wrongly classified images among all the cases (Figure 4B). This suggested that the class weight of Category 2 might be too high since its misclassified images that belonged to other categories as Category 2. Hence, the class weight combination was modified by lowering the weight in Category 2 and increasing the weight in Category 1 as to not overemphasize the impact from Category 2. Precision₂ of Category 2 significantly increased from 75.4% in Case 2 to 84.1% in Case 3, and to 85.0% in Case 4 (Table 4). In Case 2, precision₂ of Category 2 significantly increased from 75.4% in Dataset 1 to 90.2% in Dataset 2 (Table 4). Precision₂ of Category 2 significantly increased from 90.2% in Case 2 to 92.7% in Case 3 and from 90.2% in Case 2 to 94.1% in Case 4 (Table 4).

The recall metric for evaluating the CNN model that indicates the ability to correctly recognize a category was also used. In datasets 1 and 2, the recall of Category 2 was the lowest, illustrating the challenge of the model to classify images of Category 2 (early disease stages and low levels of disease symptoms) (Table 4). The highest recall of Dataset 1 Category 2 was 86.0% in Case 2, and the lowest was 74.2% in Case 1 (Table 4). This was expected given that Case 2 had a higher weight in the loss function of Category 2 compared to Case 1 (non-weighted loss function). In Case 2, Dataset 1, the recall values were similar among the three categories (Table 4). In Dataset 2 Category 2, the lowest recall was 75.0% in Case 1, and the highest recall was 84.2% in cases 2 and 3 (Table 4). The model in these two cases had the highest weight in loss function of Category 2 (early disease stages and low levels of disease symptoms).

F₁ score is a common indicator of the overall performance of the CNN model. In datasets 1 and 2, the F₁ score of Category 2 was the lowest, reaffirming the difficulty of classifying images of Category 2 by the model (Table 4). The lowest F₁ score of Dataset 1 Category 2, was 79.3% in Case 1, while the highest was 82% in both Case 3 and Case 4 (Table 4). In Dataset 2 Category 2, the lowest F₁ score was 82.6% in Case 1, and the highest F₁ score was 88.2% in Case 3 followed by Case 2 with 87.1% (Table 4).

A comparison of outcomes revealed that Category 2 was the most difficult category to classify correctly (Figure 4). This difficulty was attributed to the disease symptoms being barely visible at the early stage of infection, and some wheat spikes in Category 1 were maturing, and their color was similar to that of MoT infected spikes. We observed that the highest number of images exactly classified as Category 2 was obtained with the Case 2 Dataset 1 (Figure 4B). These results suggested that Case 2 was the most appropriate to classify wheat spike blast images in Dataset 1 because it was capable of detecting the infection at an early stage. Even though Case 2 had a slightly lower precision, this is considered the usual trade-off between precision₂ and recall for disease classification purposes. The recall, precision₂, and F₁ score increased after the images of maturing spikes were omitted when training the model with Dataset 2 (Figure 5). The cases 2 and 3 of Dataset 2 presented the highest number of images exactly classified as Category 2 (Figures 5B,C). Cases 2 and 3 were the most appropriate to detect the wheat spike blast in Dataset 2 because the model was capable of detecting the infection in the early stages. Additionally, in all the cases, the model was more stable predicting Category 3, which is relevant because it covers DS from 20.1 to 100%, potentially aiding breeders and pathologists to discern higher levels of susceptibility among cultivars. Although the CNN model misclassified some images of Category 2, it still provided a promising approach to classify the severity of the disease. It demonstrated that the CNN model is potentially a good method for breeders and pathologists.