A Real-Time Common Rust Maize Leaf Disease Severity Identifications and Pesticide Dose Recommendation Using Deep Neural Network

Maize is one of the most widely grown crops in Ethiopia, although common rust maize disease (CRMD) is becoming a serious problem. Conventional CRMD detection and treatment methods are time-consuming, expensive and ineffective. To solve these problems, we propose a real-time deep learning model that provides disease detection and pesticide dosing recommendations. In the model development process, we collect 8000 maize-leaf images experimentally and apply image pre-processing such as image equalization, augmentation, noise removal, and image enhancement to improve the performance of the proposed model. We then jointly apply batch normalization, dropout, and early stopping during training the proposed algorithm to reduce the overfitting problem. Next, we generate an optimal model that recognizes CRMD and classifies it accurately based on its severity levels. To evaluate the proposed techniques, we compare Resnet50 with other state-of-the-art algorithms, such as VGG19, VGG16, and AlexNet algorithms with similar parameters. The performance of ResNet50 is more than 2% better than that of other algorithms in terms of accuracy. The proposed technique provides promising results for the classification of CRMD and pesticide dosage recommendations based on the severity of the disease type. This study also demonstrates the potential of Resnet50 models for improving maize disease management.

Maize is one of Ethiopia's most critical crops, cultivated on over two million hectares of land and serving as a staple food crop, animal feed, and raw material for various industrial products [1]. Agriculture, particularly in Ethiopia, is the backbone of the community. Agriculture in Ethiopia provides 85% of the people's employment and livelihood, 50% of the country's gross domestic product, and 90% of export earnings in Ethiopia. Despite the fact that agriculture is important, the sector's performance in Ethiopia remains weak and exposed to a range of pathogens and abiotic factors. 

Maize crop production faces several biotic challenges, such as common rust diseases, leading to significant yield losses. Common rust maize disease is the most devastating foliar disease in areas where conditions of cool to moderate temperatures and high relative humidity prevail. Eastern Ethiopia's hararghe areas, where the weather conditions are conducive to the development of common rust, are not exempt from the disease's hazards. The temporal development of common rust and its effect on grain yield and yield components were analyzed in this area during the 2013 and 2014 cropping seasons [2]. The results showed that the disease caused a yield loss of up to 60.5%. This leads to food shortages and famine in the area. Thus, monitoring the disease is essential for successful cultivation [3]. Hence, identifying and controlling diseases at an early stage is crucial to minimizing the risk of potential yield loss.

Identifying the common rust diseases and classifying them based on their severity levels using conventional methods requires highly trained experts, which will be impractical for enormous areas. As such, there is a need for a low-cost, quick, and automated approach called machine learning. The technology is useful to identify the disease at different stages. The technology can be applied to analyze the severity of the disease and recommend the dosage of pesticides. However, based on the researchers’ knowledge, there is no work that addresses both the severity of the identification of common rust maize-leaf disease and the pesticide’s dose recommendation system. In this research, we applied five CNN algorithms—Resnet50, VGG-19, VGG-16, CNN, and Alexnet—for rapid severity classification of common rust (CR) and pesticide dose recommendation based on the level of damage occurring in maize leaves. We have applied the proposed technique to five levels of common rust disease [4].

To achieve our goals, we utilize various image preprocessing techniques, such as noise removal, image resizing, image enhancement, image segmentation, and feature extraction techniques. Moreover, we apply model performance optimization techniques such as batch normalization, dropout, and early stop techniques collectively. We use these techniques to minimize execution time and improve accuracy. To show the performances of the classification of our proposed mode on the classification of common rust maize disease, we compare it to other CNN models using different metrics. Besides, we demonstrate the pesticide dose-recommendation system using a prototype. The followings are the main contributions of this work: Design a deep-learning model for automatic severity classification and pesticide dose recommendation for common rust maize disease. We improve the proposed model using batch normalization, dropout, early stop, and a good learning rate to achieve higher classification accuracy than other state-of-the-art CNN models.

The structure of the remaining sections is shown as follows: Relevant surveys are revised in Section II. Section III presents the proposed method and data collection method. Section IV presents the experimental setup and data collection techniques. Findings and discussions are presented in Section V. Section VI concludes the paper and outlines future work

Related Works

Severity quantification of common rust refers to determining the degree of damage caused by the rust infection in the host plant. The rust infection is caused by various species of fungi, which attack plants and result in significant losses in crop yields. The severity of a rust infection can be assessed through various measures, such as visual examination, counting the number of rust spots on leaves, measurement of the leaf area affected by rust, and molecular assay techniques [5]. Factors affecting the severity of rust infection include the susceptibility of the host plant, environmental conditions, and presence of other pathogens. The measurement of leaf area affected by rust is a common method of severity quantification. However, although severity quantification of common rust is an important aspect of disease management for crop production, most of the analysis used manual quantification or did not use state-of-the-art techniques. Several methods (cultural, host resistance, and chemical) are available to manage common rust maize disease, but none of them is used to quantify the dosage of the pesticides accurately [6]. Hence, adaptive and robust techniques are required to easily quantify the maize common rust disease [7]. proposed a network for maize crop disease severity classification based on the model artificial neural network and achieved an accuracy of 93.5%. However, the researchers focused on only two classes. They did not address the identification of severity levels, which were divided by the pathologists into healthy, low, medium, high, and very high [8] used CNN to classify maize-leaf disease symptoms and got an accuracy of 98.2%. However, the researchers used several folds of the exited datasets, and this technique leads to overfitting issues [9] developed a smartphone application for the diagnosis of maize leaf diseases using SVM with a success rate of 90%. However, it is important to note that the effectiveness of this technology is heavily dependent on the quality of images captured by the smartphone camera, which may not always be consistent or accurate. Besides, the SVM model is not applicable for larger datasets [10] the scholars proposed to use Resnet50 to detect the Northern Corn Leaf Blight disease in maize crops. For the proposed model evaluations, they used publicly available datasets, and the model performance was an F1 score of 0.99, accuracy of 0.99, precision of 0.98, and recall of 1.00 using publicly available datasets. However, when the model was tested on fieldwork data, the performance declined quickly. Moreover, this work focuses on recognizing early signs of plant pathogens only rather than focusing on the severity level of the maize crop [11]. Proposed a CNN model to detect plant diseases using unmanned aerial vehicles (UAVs). Their framework extracted phenotypic traits to detect and estimate the severity of a leaf disease at the leaf level, and the schemes found 73

Design Science Research (DSR) is a problem-solving paradigm that aims to improve human and organizational capabilities through the creation of innovative artifacts [13]. These artifacts are designed to solve problems and enhance the environment in which they are used. Generally, there are four potential entry points for DSR. The first is a problem-centered approach, where researchers begin by identifying a problem and developing a solution for it. The second is an objective-centered approach, which involves addressing an industry or research need with a newly developed artifact. The third is a design- and development-centered approach, where artifacts that have not yet been formally thought through as solutions are utilized to solve problems. The fourth is a client- or context-initiated approach, which involves observing a practical solution that has worked and retroactively applying rigor to the process to develop a new solution.

For this study, a problem-centered initiation approach is selected as the entry point of the research because of the observed severity classification problems in the agricultural field and suggestions from prior research on this topic. The goal of the study is to extend human and organizational capabilities by designing new and innovative artifacts that will address the above problems and improve the agricultural environment. Specifically, the main goal of this work is to improve maize severity classification and pesticide dose recommendation by applying different regularization and preprocessing techniques.

Fig. 1 shows the structure of the proposed system from data collection up to implementation. The required images are collected from fieldwork and stored in the image database. Then, domain experts that are working at Haramaya University have done the ground truth data, such as severity levels of the maize diseases. Then, image preprocessing is applied to optimize data quality. Then, apply different regularization techniques to different ratios of datasets (training, validation, and testing ratios) to develop the proposed model. Finally, we test different CNN algorithms and select the optimum one. The details of the proposed system are discussed below:

Figure 1: The architecture of the proposed system.

Pre-processing

We use image pre-processing, such as noise removal and image normalizing, for better analysis and interpretation. The selection of the filter type depends on the type of noise present in the image.

Image Equalization 

Before all of the operations, we apply histogram equalization, as shown in Equation 1, to adjust image intensities to enhance contrast. Histogram equalization is used for contrast enhancement, and then in the image crop, the center square method is used for processing the set of databases from the image [14]:

N (P(x y)) = round                                                        (1)

Where, N (P (x y)) refers to histogram equalization, cdf refers to cumulative frequency, cdfmin= minimum value of cumulative distribution function, Cdf (p(x,y)) refers to intensity of the current pixel, Rx & Cx are product of number of pixels in rows and columns, and L refers to the number of intensities.

Noise Removal 

Several filters are commonly used for image filtering and noise removal, including the median filter, Gaussian filter, FFT filter, bilateral filter, and Wiener filter. In this research, a Gaussian filter and a median filter are applied together. Although Gaussian filters tend to blur the image, resulting in a loss of detail, median filters do not blur the image and replace the value of each pixel with the median value of the surrounding pixels. Thus, for salt and pepper noise removal, the median filter is the appropriate choice when compared to the Gaussian filter. The median filtering is defined as shown in Equation (2):

                                                              (2)

Where, Median {} denotes the median operation, and i, j are the pixel offsets from the center of the window. The equation for Gaussian filtering is shown in Equation (3):

                                                             (3)

Where, G(x,y) refers to the filtered image at pixel coordinate (x,y), σ refers to the standard deviation, e refers to the mathematical constant (approximately equal to 2.71828).

Normalization/Resizing

Image resizing is performed using bilinear interpolation to standardize the dimensions of images and fit them to the required input size of 224x224. This step helps to eliminate irrelevant information and reduce computational complexity during model training. Thus, we use bilinear interpolation to resize the maize common rust image while maintaining its quality by estimating the values of new pixels based on their nearest neighbors in the original image to avoid any loss of important details. The bilinear interpolation can be represented by Equation (4):

                                                                        (4)

Where, refers to the pixel value at a specific point on the resized image, & are the coordinates of the point on the original image, and a, b, c & are coefficients calculated based on the nearby pixel valuesin the original image.

Segmentation

As indicated in Fig. 2, we segment the content of maize common rust from background images based on colored image segmentation techniques that are used to segment images based on a specified color range. For segmentation 

                   (5)

Where segmentation mask (i, j) is the binary mask indicating whether a pixel is part of the segmented region or not, H (i, j), S (i, j), and V (i, j) are the Hue, Saturation, and Value color values of the pixel (i, j), and Hlow, Hhigh, Slow, Shigh, Vlow, and Vhigh are the lower and upper threshold values for each color channel.

                                                                                         A) Original input image                       B) Segmented Image

Figure 2: Image Segmentation (From the experiment)

Data Augmentation

The goal of data augmentation is to increase the diversity and variability of the training data, thereby reducing overfitting and improving the generalization ability of the model [15]. These results highlight the importance of leveraging data augmentation techniques to enhance machine-learning performance and generalize to new data. To show the proposed algorithm has better performance in augmented data than in original datasets, we tested it in two scenarios: Scenario 1 (original data) and Scenario 2 (augmented data). 

CNN Algorithm Usages

In this study, we evaluated the performance of five CNN algorithm types: Resnet50, VGG19, VGG16, CNN, and AlexNet, regarding classifying the severity of maize common rust and recommending pesticide doses. Majorly, we conduct two experiments: without optimization and with optimization of algorithms. Then, we select the optimum model, which outperforms other algorithms in accuracy, recall, precision, and f1-factor cumulatively. 

Model Evaluations

For this work, we use different metrics for evaluating the performance of the proposed models. These are accuracy, precision, recall, and F1-score. The mathematical formula for each evaluation is shown in Equations (6-9). 

Precision =                 (6)

Recall =                    (7)

                (8)

F1-score =             (9)

Experimental Setup and Data Collection

Image Acquisition 

We captured images from Haramaya University Rare Research Center using different smartphones (Samsung 20 AS, Huawei, and Galaxy S8). A total of 5000 images of maize leaves were collected from the field, and 3000 images were augmented with respect to 5 classes from the real maize field, and the images were prepared in a well-defined png image format. Image acquisition was performed to collect a database of images. The main data set preparation focused on the severity of maize common rust, which was classified into five levels by plant pathologists according to the CIMMYT scale. The types and severity of the diseases are identified based on their symptomatic characteristics. These symptoms are: healthy (no symptoms), low severity (disease symptoms cover > 2.5%), medium severity (disease symptoms cover >10% of the total leaf area), high severity (disease symptoms cover > 35% of the total leaf area), and very high severity (disease symptoms cover > 75% of the total leaf area), as provided in Fig 3. Thus, the collected images were labeled by expert knowledge at Haramaya University into five categories: health, low, medium, high, and very high status.

Figure 3: Sample images for maize common rust disease [1].

The data set was organized by stage of severity level, with 1600 images per class, resulting in 8000 images for training, validation, and testing. The plant pathologists classified the maize according to the level of disease present in the leaves. The researcher then allocated different ratios to training and testing sets: 70%:30%, 80%:20%, 85%:15%, and 90%:10%. The data set used for training, validation, and testing is presented in Table 1 below.

Table 1: Experimental Collected Dataset

For systematic computing, the Anaconda environment, which is an open-source offering for the Python programming language, is used. Jupiter Notebook has been used to write the Python code. Keras (a free-source CNN public library) and TensorFlow were used as backend to implement the prototype. 

Experimental Setup

In this experimental test, we assessed the impact of using batch normalization, dropout, and early stop techniques on image severity classification. In each experiment, the researcher used image pre-processing, data augmentation, image segmentation, feature extraction, and classification with a "Softmax" activation function and "Adam" optimizer. We chose the Adam optimizer due to its adaptive learning rate and better regularization techniques to prevent overfitting. We compared Adam with SGD, and Adam outperforms SGD in accuracy and time complexity. The model was trained for 50, 100, 150, and 200 epochs with a batch size of 32. However, when the epoch size is increased, the performance declines and the computational complexity increases. For smaller epochs, the accuracy becomes lower. Thus, we used 100 epochs for this work.

From different ratios of the experiments, the data ratio with 85%:15% ratios for training and testing outperforms 70%:30%, 80%:20%, and 90%:10

This paper presents a Resnet50 model for classifying the severity of maize common rust and recommending appropriate pesticide doses using image processing techniques. The model achieved higher accuracy and F1-score when using the Resnet50 algorithm with techniques such as dropout, batch normalization, and early stop. The Gradio interface was used to accurately recommend the appropriate dose of fungicide. This research demonstrates the potential of using CNNs to improve the monitoring and control of maize common rust, providing valuable insights for maize growers and researchers globally. Our proposed method shows better performance in average precision, accuracy, recall, and accuracy of maize common rust disease than the recent state-of-the-art works because of the appropriate data preprocessing and algorithm optimization.

Declaration of competing interest

The authors declare no conflict of interest.

Acknowledgment

This research is not supported by any financial sources. However, Haramaya university plant science pathology lab members are highly recognizing for their valuable knowledge experts in data sharing and ground truth data preparation.