1 Introduction

Yellow rust disease caused by Puccinia striiformis f. sp. tritici pathogen leads to significant losses in wheat production worldwide because of its high devastating property [37, 54]. It is a global disease and has been determined in more than 60 countries [54]. Turkey also suffers from the wheat yellow rust epidemic [13]. It caused crop damage estimated at ten million US dollars in Central Anatolia in 2010 [6]. Due to disease, yield losses might reach up to 100% when hypersensitive cultivars are used and when suitable weather conditions occur [11]. Moreover, yellow rust limits not only the yield but also the quality of wheat. As a result, it is critical to monitor the development of yellow rust in natural conditions using reliable methods for minimizing economic losses. The disease causes physical changes on wheat leaves such as leaf color, leaf texture, and leaf morphology [26], the severity of the disease is evaluated through these physical deformations. In other words, texture features provide enough evidence about wheat yellow rust [40]. However, the process of determining the severity level of yellow rust is carried out according to certain standards through visual inspection in field conditions by specialists. This process is subjective, labor-intensive, and time-consuming. Failure to perform accurate severity level evaluation may pose problems in later growth stages, especially in precision agriculture in which detecting infection type is critical in taking necessary precautions against the disease.

Texture features analysis over two-dimensional images is one of the most popular approaches in digital image processing problems [2]. It focuses on partitioning images into regions of interest and, later, classifying those regions. It utilizes existing information on the spatial arrangement of colors or intensities in an image. It extracts information about texture features, structural features of surfaces, and their relationship with each other [17]. Texture feature extraction methods have been used for about half a century [18] in various computer vision applications, including object detection and image classification [55]. Over time, the number of features has increased to better adapt to the various tasks. The methods consist of various techniques such as Gray Level Co-occurrence Matrix (GLCM), Local Binary Patterns (LBP), Gabor Filter, Segmented Fractal Texture Analysis (SFTA) each based on different calculations. GLCM is a popular measurement among the other texture feature extraction methods designed to perform statistical calculations over different image textures and forms [17]. It is often used especially in the fields of texture analysis and machine learning. GLCM performs calculations using statistical distributions of combinations of different intensity values for an image [48]. However, capturing information only from grayscale images limits GLCM features’ effectiveness with color images. Thus, recently, color spaces have attracted the attention of researchers as they reveal hidden patterns in color channels [15, 20].

On the other hand, recent developments in both information technologies and image analysis offer new methods to extract data-dependent features to develop more sophisticated systems for image-based problems [36, 47, 50]. Convolutional Neural Network (CNN) is in the leading position among these methods due to its ability to have both automatic feature extraction and classification. It also has the advantage of spatial information over other types of deep networks. CNN is a deep learning (DL) method that is especially popular in classifying images, text, sounds, and videos. One of CNN’s strongest sides is that it automatically extracts data-dependent appropriate features from images in the dataset. Researchers are still on the hunt to improve the effectiveness and generalizability of deep learning models [35, 51].

Herein, various models have been proposed to determine the severity of yellow rust disease in wheat. The models were trained on a dataset that included not only CNN-based auto-generated data representation features, but also traditional texture features. The study focuses on the hypothesis of “diversity in the feature set is crucial in developing state-of-art learning models trained on image datasets even though dataset-representative image-based features are available”. Pre-trained DenseNet-201 was employed to extract CNN-based features. DenseNet-201 was fine-tuned by the so-called Yellow-Rust-19 dataset [19] to obtain task-specific features. DenseNet-201’s Global Average Pooling (GAP) layer was utilized for extracting 1920 features. Also, using both the Gray Level Co-occurrence Matrix (GLCM) and the Color-Gray Level Co-occurrence Matrix (CGLCM), a total of 88 texture features were obtained, 22 features corresponding to each grayscale and 22 features corresponding to each color channel for each leaf image in Yellow-Rust-19 dataset [19]. Moreover, three different color spaces such as Red-Green-Blue (RGB), Hue-Saturation-Value (HSV), and Lightness-Red/Green-Blue/Yellow (L*a*b) were utilized for obtaining color textural features.

Experimental results confirm that models trained using combined features sound and are promising. Experimental results also confirm that the models developed using combined features perform better than the models developed by using one of the CNN-based features and textural features. The support Vector Machine (SVM) method and K-Nearest Neighbors (KNN) were employed as the classifiers. Figure 1 shows the overall workflow.

Fig. 1
figure 1

The workflow of the study

Full size image

The remaining part of this paper is organized as follows. In the next section, some studies’ reviews related to the methodologies used in the study are presented. Section 3 explains the methodological framework. This section gives information about the final data set and feature extraction methods. The combined textural analysis results based on different color spaces and CNN results are presented in Section 4. In the last section, the results are summarized and recommendations for further studies are given.

2 Related work

Precision farming is heavily dependent on information technology, with data and connectivity at its core. With artificial intelligence, analytics, and connected sensors, it can further increase yields, detect diseases and build sustainability and resilience across crop cultivation. Kundu et al. [25] developed a CNN-based model called Custom-Net that collects the imagery and parametric data from the pearl millet farmland and predicts the blast and rust diseases in pearl millet. Hayit et al. [19] introduced a multi-label deep convolutional neural networks-based model to classify the severity of yellow rust disease on wheat. The model inputs the wheat leaf image and classifies it by the percentage of rust. However, Abayomi-Alli et al. [1] aimed to improve the accuracy of deep learning models on low-quality test images using data augmentation. They generated synthetic images by histogram transformation technique to recognize important colour-based features, which are less sensitive to the deficiencies of low-quality images.

On the other hand, there exists recent studies on the diagnosis of plant diseases through rich, informative, and data-representative features. Almadhor et al. [3] developed an artificial intelligence (AI)-based model to detect and classify the most common guava plant diseases by utilizing a combination of color histograms and Local Binary Patterns (LBP) textural features. However, GLCM is a widely used statistical method for texture analysis, but it can only capture information from grayscale images [18, 20], which limits its effectiveness with color images. Recently, color spaces have also attracted researchers’ attention due to their ability to reveal color data hidden in pixel values. Thus, many efforts have been made to extend the effectiveness of GLMC by combining or integrating the color in it. For example, Hossain and Parekh [20] studied the possibilities of including color information from color texture images to improve texture analysis. The results proved that color GLCM combination provided better accuracies as compared to standard GLCM. In another study, Gui et al. [15] demonstrated the texture feature extraction based on Color Co-occurrence Matrix is superior over the GLCM-based texture feature extraction methods. HSI and RGB color spaces were used for texture feature extraction. More on combining color and grey-level co-occurrence matrix features, see [23, & references therein].

Nowadays, with the emergence of deep learning, especially CNNs, there has been a shift to the use of both hand-crafted and learning-based features. Zhou et al. [58] employed hand-crafted features like color histogram and LBP for image retrieval as well as the deep features extracted from the various CNNs such as VGG16, ResNet, and AlexNet. Zhang et al. [56] proposed a classification model for medical images, which the model was trained by utilizing deep and handcrafted visual features. In [53], Wei et al. proposed a hybrid feature descriptor based on DL by integrating the deep and handcrafted features. An SVM-based classification model was used for training. Experimental results show that the proposed method outperforms the state-of-the-art predictors.

Nowadays, with the emergence of deep learning, especially CNNs, there has been a shift to the use of both handcrafted and learning-based features. Zhou et al. used deep features extracted from various CNNs such as VGG16, ResNet and AlexNet as well as handcrafted features such as color histogram and LBP for image acquisition. Zhang et al. [58] proposed a classification model for medical images, in which the model is trained using deep and handcrafted visual features. In [53], Wei et al. proposed a hybrid feature descriptor based on DL by integrating deep and handcrafted features. A SVM-based classification model was used for training. Experimental results show that the proposed method outperforms state-of-the-art predictors. Nanni et al. [32] proposed a model for the classification of virus images by using both hand-crafted features and pre-trained deep neural network features. SVM was used for features training. Various pattern analysis methods were used as handcrafted features extraction methods. The experiment results showed that the combined method with hand-crafted and deep features improve the accuracy. Luz et al. [27] presented an urban sound classification method based on deep and handcrafted features combination. CNN model was used for extraction deep features. Random Forest and SVM were used as classifiers and also two separate data sets were used. For both classifiers, the proposed method outperformed handcrafted features. Zhang et al. [57] used a combination of handcrafted and deep CNN features extracted from X-Ray images for classify between healthy, widespread pneumonia, and Covid-19 patients. The performance of the proposed approach was compared with a standard CNN and the SVM trained with texture features. It was reported that combining the features improves the accuracy of the classification compared to the independent application of handcrafted and deep features. Naz et al. [33] proposed a hybrid method based on the texture features and deep features for automated detection of gastrointestinal diseases by using wireless capsule endoscopy images. Texture features were extracted by using SFTA and LBP texture analysis methods. Also, deep features were extracted by using VGG-16 and Inception-V3 pre-trained CNN models. Extracted deep features were combined serially along with the obtained texture features to obtain an ensemble deep feature vector. It was reported that the results of proposed approach are promising.

In addition, state-of-the-art studies utilized the DenseNet [21] model for similar tasks that this study is concerned with [12, 31, 44, 49]. Each of these studies used leaf images of different plants as a dataset and successfully detected diseases. While some of these studies use the basic DenseNet model, others modified the DenseNet to their specific task.

The studies summarized above indicate that a combination of textural and deep features improves the performance of learning-based models, which motivated us to explore a learning-based classification performed on a combination of textural and deep features to determine the severity level of yellow rust disease in wheat.

3 Materials and method

3.1 Wheat yellow rust infection types

Yellow rust disease causes physical changes in wheat leaves such as leaf color, leaf texture, and leaf morphology [26], which provide sufficient evidence of wheat yellow rust disease [40] to determine the type of yellow rust infection. Relative to the resistance of wheat to the disease, there are six different infection types according to the modifed Cobb scale [38]. Infection types no disease (Immune), Resistant (R), Moderately Resistant (MR), Moderately Resistant—Moderately Susceptible (MR-MS), Moderately Susceptible (MS), and Susceptible (S) in Table 1 presented by Johnston and Browder [22] were used for disease assessment. However, the process of determining the severity level of yellow rust is carried out according to certain standards through visual inspection by specialists.

Table 1 Infection types of wheat leaf rust used in disease assessment
Full size table

3.2 Dataset

The classification has been performed on the dataset called Yellow-Rust-19, which was initially created by Hayıt et al. [19] and composed of wheat leaf images. In order to obtain images, first, wheat leaves were collected from the cultivation area belonging to the Republic of Türkiye Ministry of Agriculture and Forestry Directorate of Field Crops Central Research Institute located in Ankara, Türkiye. After that each leaf was examined to determine the severity level of yellow rust by two specialists. Then, they were photographed to yield a raw dataset. After that, each image in the raw dataset went through some image pre-processing. Finally, Yellow-Rust-19 was created by the labeling and data augmentation process, for details see [19]. Sample leaf images from Yellow-Rust-19 are represented in Fig. 2 and Table 2 shows the statistics for the dataset.

Fig. 2
figure 2

Sample leaf image of each class [19]

Full size image
Table 2 Final dataset statistics
Full size table

3.3 Color spaces

Texture, the spatial arrangement of pixels, has been one of the most popular features in image processing tasks [39]. Although grayscale textures provide enough information to solve many tasks, many researchers have started to consider color information in recent years [4, 7, 14, 15, 20]. And, in these studies, it has been observed that structural feature extraction through color texture has led to a considerable increase in sensitivity. Recall that color space is an abstract mathematical model for representing colors numerically, that is, the color space establishes a mapping between numeric values and specific colors. However, there are different color spaces for various purposes in image analysis. In this study, three different color spaces such as Red-Green-Blue (RGB), Hue-Saturation-Value (HSV), and Lightness-Red/Green-Blue/Yellow (L*a*b) were utilized for obtaining color textural features.

3.4 Feature extraction

3.4.1 Gray Level Co-occurrence Matrix (GLCM) and Color-Gray Level Co-occurrence Matrix (CGLCM) feature extraction

Visual texture (in short texture) is a feature of repeating patterns of local variations in image intensity utilized for determining objects or regions of interest in an image. It provides information about the spatial arrangement of colors or intensities in the image and is described by the spatial distribution of intensity levels in a neighborhood. The statistical texture, on the other hand, is a quantitative texture measure of the arrangement of intensities in a region. Gray Level Co-occurrence Matrix (GLCM) contains information about the positions of pixels having similar gray level values and, mathematically defined by,

$$ G_{(dx,dy)}(i,j)=\sum\limits_{x=1}^{n}\sum\limits_{y=1}^{n}\begin{cases} 1, & \text{if } I(x,y)=i \text{ and } I(x+dx,y+dy)=j \\ 0, & \text{otherwise } \end{cases} $$
(1)

where i and j represent the pixel values, x and y represent the spatial positions in the digital image I, dx and dy represent the displacement values and define the spatial relation for which this matrix is calculated, and I(x,y) represents the pixel value at the spatial position (x,y). In words, G(dx,dy)(i,j) counts all pairs of pixels having gray values i and j occur within (dx,dy) neighbors of I(x,y). We observe that G(dx,dy) is an n-by-n matrix. However, the normalized GLCM is defined by

$$ P(i,j)=\frac{G_{(dx,dy)}(i,j)}{{\sum}_{i=1}^{n}{\sum}_{j=1}^{n}G_{(dx,dy)}(i,j)} $$
(2)

where the entries are between 0 and 1.

Haralick et al. [18] proposed a set of 14 textural features that can be inferred from the co-occurrence matrix given at (2) in 1979. The set contains information about image textural characteristics such as homogeneity, linearity, and contrast. Later on, additional eight textural features were introduced [8, 17, 43]. Some of these features have intuitive and familiar definitions, for example, Energy is the second moment of P(i,j). The mathematical definitions of features are listed at Table 3.

Table 3 Texture features computed from GLCM P(i,j)
Full size table

On the other hand, for colored images, relatively little is known about the relationship between color information and GLCM textural features. It is known that GLCM textural features tend to be globally adaptable but not locally optimized. Shearer [42] suggested the use of color texture analysis to overcome the shortcomings of traditional gray-level texture analysis. Within the scope of this study, all textural features listed in Table 3 were obtained from each color image, i.e. 22 features for each channel. Moreover, each image was converted to grayscale, and the similar feature set was obtained for each grayscaled image. Thus, in total, 88 features were contributed to the dataset by each image.

Furthermore, each leaf was represented in three different color spaces such as RGB, HSV, and L*a*b, making three different textural datasets corresponding to each image color space. Here is a good place to point out that each image was resized to 224 × 224 before the textural feature processing was applied, as the DenseNet network converts the input images to 224 × 224 before any other processing. Finally, each feature was normalized using the min-max normalization method to reduce the classification time. Figure 3 depicts textural features for an image in RGB color space.

Fig. 3
figure 3

A Sample representation for CGLCM_RGB

Full size image

As a result of all these processes, three different feature matrices, corresponding to each color space, with 15.000 rows, corresponding to each image, and 88 columns, corresponding to each textural feature, were obtained for the dataset Yellow-Rust-19.

3.4.2 DenseNet-201 feature extraction

DenseNet (Dense Convolutional Network) [21] is a convolutional neural network with dense connection, fewer parameters, but high accuracy. Huang et. al. [21] introduced DenseNet because of the observation that convolutional networks could be substantially deeper, more accurate, and efficient to train if they contained shorter connections between layers close to the input and those close to the output. DenseNet concatenates the outputs of the previous layer and the future layer, meaning each layer is directly connected to every other layer in a feed forward structure within each dense block. DenseNet is one of the new technologies in deep learning for visual object recognition and has different versions such as DenseNet-121, DenseNet-160, DenseNet-201 and DenseNet-264. DenseNet accepts 224-by-224-by-3 input images. DenseNet diminishes the vanishing-gradient problem, consolidates feature propagation, promotes feature reuse, and considerably reduces the number of parameters. Herein, DenseNet-201 was preferred for deep features extraction.

DenseNet-201 is a convolutional neural network that is 201 layers deep. DenseNet-201 consists of four dense blocks with varying number of layers [6,12,48,32]. In each dense block, each layer puts some additional features on top of the existing feature maps. Inside the dense blocks, the feature maps’ size remains the same. DenseNet-201 was trained on the ImageNet database [10]. Like other convolutional neural networks, DenseNet-201 network creates a hierarchical representation of the input images. Deeper layers contain higher-level features created using lower-level features of previous layers. Figure 4 shows a visualization of a sample leaf image from the first Convolution Layer (CL) and the Batch Normalization Layer (BNL) in DenseNet-201 architecture. Features extracted at the Global Average Pooling (GAP) layer at the end of the network were utilized in the classification phase. The GAP provided 1920 features in total. See Table 4 for architecture details.

Fig. 4
figure 4

Visualization of a sample leaf image from the first Convolution Layer (CL) and the Batch Normalization Layer (BNL) in DenseNet-201 architecture

Full size image
Table 4 DenseNet-201 architecture
Full size table

And, Table 5 presents the hyperparameters utilized in training the model.

Table 5 DenseNet-201 hyperparameters
Full size table

3.5 Classification

This study aimed to investigate the feasibility of using only textural features of colored images compared to the combination of textural features of colored images and DenseNet-201 deep features to determine the severity level of yellow rust disease in wheat. The dataset was labeled, meaning each data was attached with the severity level of yellow rust disease. There are six severity levels such that Immune, R, MR, MR-MS, MS, and S. Thus, it is a multi-class classification task. Many multi-class classification methods exist in the literature, however, Multiclass Support Vector Machines (SVM) and K-Nearest Neighbors (KNN) were utilized in the classification phase. Indeed, during experimental studies, other classifiers such as Artificial Neural networks, Decision trees were also tested, but SVM and KNN performed better than others, thus, the current study presents results by SVM and KNN.

3.5.1 Support vector machines

SVM is a widespread technique in texture classification tasks [29], was employed in many texture classification studies [5, 24, 46].

SVM seeks to find an optimal boundary between the possible labels. It is based on maximizing the minimum distance from the separating hyperplane to the nearest label. While the fundamental SVM only supports binary classification, its extensions are designed to handle the multiclass classification situation as well. To distinguish K classes, the classifier can use multiple SVMs, each SVM predicting membership in one of the K classes. The idea is to map data points to the hyperplane to achieve mutual linear separation between every pair of classes. There are two approaches such that One-to-One approach and One-to-Rest. The current study utilized One-to-One approach, which puts a hyperplane to separate every two classes, neglecting the data points of the other classes, thus, it uses \(\frac {K(K-1)}{2}\) SVMs. In this case, the classifier used 15 SVMs. The parameters for the multiclass SVM classifer is outlined at Table 6.

Table 6 Multiclass SVM hyperparameters
Full size table

3.5.2 K-Nearest neighbors

KNN, a widely used and non-parametric classification method, relies on visible pattern similarities and distance metrics to generate precise predictions, in short, it is a similarity measure. It is frequently utilized in texture classification problems [16] since patterns close to each other in the feature space most likely belong to the same pattern class. KNN estimates the likelihood that a data point is a member of one group or another based on which group the data points closest in distance belong to. There are different distance metrics such as Euclidean distance, Manhattan distance, Minkowski distance, etc. employed with KNN, however, Euclidean distance is the most accepted metric and, for the data points x = (x1,x2,⋯ ,xn) and y = (y1,y2,⋯ ,yn), it is defined as:

$$ \textrm{Euclidean distance}(\textbf{x},\textbf{y})=\sqrt{\sum\limits_{i=1}^{n}{x_{i}^{2}}-{y_{i}^{2}}}. $$
(3)

On the other hand, The k value in KNN classifier represents the number of nearest neighbors to be included in groups during the voting process. The selection of k is crucial and depends upon the data, however, generally, while large k values reduce the effect of the noise on the classification, they may cause boundaries between classes less distinct.

The current study uses Euclidean distance to measure the distance between two data points and takes k as 10. Moreover, it has used inverse distance weighting to assign higher weights to the closest training samples.

4 Results

The study investigates the feasibility of using only colored images’ textural features compared to the combination of colored images’ textural features and DenseNet-201 deep features to determine the severity level of yellow rust disease in wheat, among one of six severity levels such that Immune, R, MR, MR-MS, MS, and S. Initial dataset was Yellow-Rust-19, which contains 15000 wheat leaf images with varying size. Initially, each image was converted to the dimension of 224 × 224. Then, the dataset was randomly split into two parts using the stratified selection strategy that guarantees an even distribution of images from each class to allocate 90% training samples for model development and a 10% test sample for the model’s generalization evaluation. Table 7 shows the training set and test set numbers.

Table 7 Training and test sets statistics
Full size table

The implementation, as well as training of the proposed transfer model, was carried out using both Python and MATLAB\({\!}^{\circledR }\) programming environments.

For each leaf image, 66 textural colored features were obtained corresponding to each color space as RGB, HSV, and L*a*b, making three different colored textural features datasets. Moreover, for each image 22 gray level textural features were also obtained from each grayscale leaf image. Furthermore, for each image 1920 DenseNet-201 deep features provided by the GAP layer were obtained.

Different models on different datasets with different classifiers were developed. A total of twelve models, six different multiclass models for each SVM and KNN classifier, were developed and evaluated, six of them, three for each classifier, were trained using only texture features, and six of them, again three for each classifier, were trained using both textural and deep features. Table 8 shows the datasets’ names and corresponding color spaces.

Table 8 Datasets, color spaces
Full size table

Among models trained using texture features, the SVM classifier with an accuracy of 73.1% performed best via CGLCM_RGB feature set, however, the model using SVM classifier trained on the feature set CNN-CGLCM_HSV had the best accuracy with 92.4%. Table 9 presents the methods, corresponding datasets, and their accuracies. Observe from Table 9 that models trained using both textural and deep features outperformed models trained using textural features alone, moreover, the models trained by the dataset CNN-CGLCM_HSV for both SVM and KNN classifiers performed better.

Table 9 The accuracy rates (%)
Full size table

On the other hand, Fig. 5 shows the confusion matrices for the models trained on the dataset CNN-CGLCM_HSV, with the classifier SVM on the left and KNN on the right. The confusion matrix reveals how a classifiers performed in each class. 0 represents “Immune”, number 1 represents “R”, 2 represents “MR”, 3 represents “MR-MS”, 4 represents “MS” and 5 represents “S” class. A close look at confusion matrices reveals that prediction to MS class was poor by both models.

Fig. 5
figure 5

The Confusion matrix for the models trained on the dataset CNN-CGLCM_HSV, on the left by the classifier SVM and on the right by KNN

Full size image

The test set confusion matrix across all runs was recorded, and the following three metrics were also calculated: recall, precision and F1-score are demonstrated in Table 10. Among the textural features, observe that the models trained on CGLCM_RGB dataset perform better than others. Indeed, models trained on combined features considerably outperform models trained on textural features alone. The model trained on CNN-CGLCM_HSV dataset, with the SVM classifier achieves the best results in all performance criteria.

Table 10 Performance comparison of the models
Full size table

Table 11 presents the class performance values of the best model trained on the CNN-CGLCM_HSV dataset and using the SVM classifier. A closer look at the table reveals that the model performs poorly in distinguishing MR-MS and MS classes compared to the others.

Table 11 Classwise performance comparison for the CNN-CGLCM_HSV dataset
Full size table

5 Discussion

From the experiment results, we observed that the combined features have better accuracy rates than traditional textural analysis methods. The dataset which represents healthy wheat leaves and diseased wheat leaves with different types of infection (Yellow-Rust-19) was used in all methods. The best performance for combined features was calculated as 92.4% for HSV color space. In addition, it was observed that average accuracy rate increase for all color spaces. The same SVM parameters were used for all classifications for fair comparison.

The popular texture feature extraction techniques are stated in the literature [29]. The GLCM is a robust method for statistical texture analysis [28, 34, 43] and is an effective texture identifier. It is also proven that it has better accuracy value and computational time than other texture analysis methods [9]. Therefore, the authors preferred the GLCM for texture feature extraction. In addition, it is stated that GLCM can be improved to be applied on different color spaces [29, 41, 45]. This paper also supports this view.

Although, there is a performance improvement in all color spaces, confusion matrices showed there were differences in the performance obtained for each class. This deficiency must be taken into account since the statistics of the confusion matrices were similar in all color spaces. However, deep learning methods can be helpful for increasing the performance of textural features and creating a more stable system. Although performance tests prove the performance of combined methods, better classification methods should be tested to use the features in real conditions.

Transfer learning (TL) as a deep learning approach, allows an existing pre-trained model on a large-scale dataset for one task on a domain to be used as a starting point for another task on a similar domain with relatively less labeled data. TL-based models need less training time, compared to models trained from scratch [30].

Wang and Zhang [52], reported that DenseNet surpasses other pre-trained models based on the comparison results of popular pre-trained models. Also, in the same study, Wang and Zhang have proved that DenseNet-201 is more successful than other DenseNet versions. Therefore, it was preferred DenseNet-201 for deep feature extraction.

This paper, discusses the feasibility of combination of texture features and deep features to determine the severity level of yellow rust in wheat. Although textural and deep features were combined in this study, the expected high classification performance could not be achieved in the MS class, which contains similar samples to MR-MS and S classes. On the other hand, even though, the Yellow-Rust-19 dataset was labeled by two experts via consensus to create a gold standard (reference standard), it is still possible that mislabeling happens and the performance of the proposed method might be degraded.

Finally, the results of this paper and related works which research combined features are presented in Table 12. As a result, the contribution of deep features to classification performance is evident in other studies, as in this study.

Table 12 Performance comparison for the combined features (CF) against the standard texture features (STF)
Full size table

6 Conclusion

Throughout the study, a hybrid method based on combining deep and textural features was proposed in a six-way multi-label classification task for the Yellow-Rust-19 dataset. Deep features were obtained by the pre-trained network DenseNet. Different models on different datasets with different classifiers were developed. Test accuracy, as well as other performance metrics on the test set, such as precision, recall, F1 score, indicate that combined methods are promising (Table 10). Although the textural features alone are not sufficient, it has been proved that it improves sensitivity when combined with deep features. In future studies, the results can be compared by testing different pre-trained networks other than DenseNet. In addition, other texture analysis methods other than CGLCM used in this study can be tested and their contribution to success can be discussed.