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

Abstract

Maize is one of the most widely grown crops in Ethiopia and is a staple crop around the globe; however, common rust maize disease (CRMD) is becoming a serious problem and severely impacts yields. Conventional CRMD detection and treatment methods are time-consuming, expensive, and ineffective. To address these challenges, we propose a real-time deep-learning model that provides disease detection and pesticide dosage recommendations. In the model development process, we collected 5000 maize leaf images experimentally, with permission from Haramaya University, and increased the size of the dataset to 8000 through augmentation. We applied image preprocessing techniques such as image equalization, noise removal, and enhancement to improve model performance. Additionally, during training, we utilized batch normalization, dropout, and early stopping to reduce overfitting, improve accuracy, and improve execution time. The optimal model recognizes CRMD and classifies it according to scientifically established severity levels. For pesticide recommendations, the model was integrated with the Gradio interface, which provides real-time recommendations based on the detected disease type and severity. We used a convolutional neural network (CNN), specifically the ResNet50 model, for this purpose. To evaluate its performance, ResNet50 was compared with other state-of-the-art algorithms, including VGG19, VGG16, and AlexNet, using similar parameters. ResNet50 outperformed the other CNN models in terms of accuracy, precision, recall, and F-score, achieving over 97% accuracy in CRMD classification—surpassing the other algorithms by more than 2.5% in both experimental and existing datasets. The agricultural experts verified the accuracy of the recommendation system across different stages of the disease, and the system demonstrated 100% accuracy. Additionally, ResNet50 exhibited lower time complexity during model development. This study demonstrates the potential of ResNet50 models for improving maize disease management.

1. Introduction

Agriculture is a cornerstone of Ethiopia’s economy, providing livelihoods for 85% of the population, contributing 50% to the Gross Domestic Product (GDP), and generating 90% of export earnings [1]. It plays a crucial role in the country’s socio-economic development. However, despite its importance, the agricultural sector remains weak, particularly in rural areas, and is vulnerable to numerous challenges, including pathogens like viruses, fungi, and bacteria, as well as biotic factors like pests and abiotic stresses such as water scarcity, heat, and salinity.

Maize is one of the most widely cultivated crops globally, both in developing and developed countries. According to Zhou et al. [2], maize ranked among the most cultivated cereals worldwide; approximately 1.14 billion tons of maize was produced in 2019. It plays a crucial role in global food security within the agricultural sector but is frequently threatened by various diseases, including maize leaf spot, pathogenic spot disease, maize aphids, and maize borers. These challenges significantly impact food security, particularly in developing regions such as Africa and tropical Latin America [3]. In Ethiopia, maize is one of the most critical cereal 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 [4].

Maize crop production faces several biotic challenges, including common rust maize diseases (CRMDs), maize leaf spot disease, pathogenic spot disease, maize aphids, and maize borers, all of which contribute to significant yield losses [2.5]. Among these, CRMD is the most devastating foliar disease, thriving in cool to moderate temperatures and high relative humidity. In Eastern Ethiopia’s Hararghe areas, where weather conditions are conducive to the development of common rust, the disease poses severe threats to maize cultivation. The temporal development of common rust and its impact on grain yield and yield components were analyzed in this area during the 2013 and 2014 cropping seasons [6]. The results revealed that the disease caused a yield loss of up to 60.5%, even though the analysis utilized only 230 images. Such losses exacerbate food shortages and famine in the area, underscoring the importance of timely disease management. Monitoring and early-stage identification of diseases like CRMD is crucial for successful cultivation [7]. Early detection and control can significantly minimize potential yield losses, ensuring sustainable food production and improved food security in vulnerable regions.

Identifying the CRMD and classifying it by severity using conventional methods poses significant challenges, including requiring highly trained experts to perform said identification and classification, which is impractical when monitoring large areas. Traditional techniques for quantifying maize leaf diseases are time-consuming, labor-intensive, and prone to errors, largely due to the absence of standardized diagnostic guidelines [8]. As a result, there is a pressing need for a low-cost, fast, and automated solution, such as machine learning, to effectively identify disease at various stages. This technology offers the potential to analyze disease severity and recommend precise pesticide dosages [9,10]. To date, no existing studies have addressed CRMD severity classification or pesticide dosage recommendations comprehensively. In this research, five convolutional neural network (CNN) algorithms—Resnet50, VGG-19, VGG-16, CNN, and Alexnet—were applied to rapidly classify the severity of CRMD and to recommend appropriate pesticide dosages based on the damage levels observed on maize leaves. The proposed approach classifies CRMD into five severity levels, which are determined by the extent of visible spots on maize leaves. The main objective of this work is early detection of maize diseases, which enables the rapid implementation of control measures. These include selecting appropriate pesticides tailored to the disease severity, which ultimately improves crop yields and produce quality. Hypothesis: Early detection of crop diseases using deep learning models, such as CNN, will facilitate timely implementation of control measures, including the selection of suitable pesticides, significantly enhancing crop yields and produce quality.

To achieve our goals, we utilized a range of image preprocessing techniques, including noise removal, image resizing, image enhancement, image segmentation, and feature extraction. Moreover, we implemented performance optimization techniques such as batch normalization, dropout, and early stopping, which collectively minimized execution time and improved model accuracy [11]. To validate the pesticide dose recommendation system, we developed a prototype, which was thoroughly evaluated by field experts to ensure its accuracy and effectiveness. The key contribution of this work includes the development of a deep learning model for automatic severity classification and pesticide dose recommendation for CRMD. The model’s performance was significantly enhanced through batch normalization, dropout, early stopping, and careful tuning of the learning rate, and it achieved higher classification accuracy than existing state-of-the-art CNN models. Furthermore, extensive fieldwork was conducted in Ethiopia to collect experimental data, ensuring that the model is tailored to local agricultural conditions and practical needs.

2. Related Works

The severity of a rust infection can be assessed through several methods, including visual examination, counting the number of rust spots on leaves, measurement of the leaf area affected by rust, and molecular assay techniques [12,13]. Among these, measuring the leaf area affected by rust is a commonly used approach for quantifying severity. However, although severity quantification of common rust is an important aspect of disease management for crop production, most analyses rely on manual quantification methods, which are time-consuming and prone to errors. Additionally, the adoption of cost-effective, state-of-the-art technologies, such as AI-driven applications, remains limited. In many developing countries, traditional methods such as cultural practices, host resistance, and chemical treatments are commonly used to manage CRMD, rather than technology-based early detection methods [14]. Advanced methods that tailor treatment based on severity levels could offer significant advantages with regard to the management of fungal infections and mitigation of yield reductions. Therefore, adaptive and robust techniques are critical to the effective quantification of CRMD, enabling improved disease management and enhancing crop yields, particularly for smallholder farmers.

In a review of work published from 2016 to 2021, Nunoo-Mensah et al. [5] highlight the significant advancements provided by deep learning in terms of improving the accuracy of maize leaf rust disease detection. Despite these improvements, several challenges persist, including the lack of expert annotation, suboptimal model architectures, issues with hyperparameter tuning, and limited training resources. Their study highlights transfer learning’s potential to reduce training time and enhance model performance but stresses that achieving optimal outcomes still depends on access to sufficiently large and diverse datasets and the careful selection of algorithms. Similarly, the review articles in [15] examined 100 review studies and concluded that the CNN algorithm is particularly effective for early detection of leaf diseases using image datasets. The results showed 91% accuracy in disease region segmentation and 74% accuracy in leaf instance segmentation. Kundu et al. [16] utilized the K-means clustering algorithm to extract the region of interest and used a deep learning algorithm for maize leaf disease detection, achieving the highest accuracy of 98.50% with 2996 images. However, this work relied on a relatively small dataset tailored to the K-means algorithm and did not incorporate optimization techniques. As a result, the findings have limited applicability to larger datasets, making it challenging to generalize the results for broader real-world scenarios.

Nagaraju and Chawla [17] proposed using deep learning, particularly Inception-v3, for maize crop disease severity classification, and they achieved a maximum accuracy of 95.99% using a total of 5939 digital images of maize. However, the study focused on only two classes, and while four classes were used, the unbalanced class sizes may have introduced biases, potentially leading to inaccurate conclusions. In the work of Zhou et al. [2], CNN-ShuffleNetV2 architecture was employed to detect maize leaf disease symptoms, achieving an accuracy of 98.4%, which marked a 4.1% improvement over the original model. The dataset included 2200 experimental images, which were augmented to 4000 images, as the results were enhanced through algorithms and hyperparameter optimizations. Additionally, Khan et al. [18] developed a smartphone application for the diagnosis of maize leaf diseases in 2675 images using YOLOv8, with a success rate of 99.4%. The researchers compared their proposed work with YOLOv3-tiny, YOLOv4, YOLOv5s, and YOLOv7s, and demonstrated that YOLOv8 outperforms the other state-of-the-art algorithms.

Richey et al. [19] proposed using ResNet50 to detect Northern Corn Leaf Blight disease in maize crops using publicly available datasets, and the model achieved an F1 score of 0.99, accuracy of 0.99, precision of 0.98, and recall of 1.00. However, when tested on fieldwork data, the model’s performance declined significantly due to variations in image quality. Additionally, this work focused exclusively on recognizing early signs of plant pathogens without addressing the severity levels of the maize crop disease. In the work of Hernández and López [20], a combination of Bayesian probabilistic and deep learning algorithms was proposed to detect maize leaf diseases. While the combination reduced the performance variability observed in deep learning models, the accuracy of the proposed algorithm was not numerically reported, leaving its effectiveness unclear.

Reddy et al. [21] proposed a deep learning-based approach for maize leaf disease detection using a CNN model with image sizes of 168 × 168 and 224 × 224, achieving accuracies of 84.66% and 85.23%, respectively, utilizing the Adam optimizer. The researchers trained their model on 13,300 images sourced from Kaggle [22] and tested it on 2660 field images. However, the study had limitations, including a lack of data preprocessing, algorithm optimization, and hyperparameter tuning, resulting in lower accuracy compared to other state-of-the-art performances.

In the agricultural sector, Design Science Research (DSR) has been particularly effective in advanced precision agriculture practices [23]. For instance, machine learning models applied to drone imagery and multispectral data have enhanced pest detection and crop yield predictions, leading to improved resource allocation and reduced crop losses [24]. These systems leverage CNN or similar architectures alongside transfer learning techniques, which results in scalable and adaptive solutions that can be used to resolve complex agricultural challenges [25].

3. Methodology

In this study, we leveraged the problem-centered DSR approach to develop a deep learning-based system aimed at improving disease severity classification and optimizing pesticide dose recommendation for CRMD. Our system integrates CNN-based image classification with dynamic pesticide recommendation models, addressing the limitations of traditional CRMD management. This adaptive solution not only optimizes pesticide application but also minimizes resource waste, aligning with the principles of precision agriculture and sustainable farming practices. The goal of the study is to extend human and organizational capabilities by designing an innovative tool that tackles these challenges and improves agricultural precision.

3.1. Proposed Architecture

Figure 1 illustrates the structure of the proposed system, encompassing all stages from data collection to implementation for disease detection and severity classification using deep learning. The required images were collected from fieldwork, and the ground truth data, including maize disease severity levels, were provided by domain experts and pathologists from Haramaya University. To ensure high-quality data, image preprocessing techniques were applied, enhancing the dataset’s reliability. Additionally, regularization techniques were employed on datasets with varying ratios to improve model robustness. The data were then split into training and testing sets, facilitating the development of the proposed model. The model’s performance was subsequently evaluated against other CNN models, and it demonstrated superior accuracy and adaptability, outperforming state-of-the-art algorithms.

3.2. Preprocessing

We applied varies image preprocessing techniques, including image resizing, noise removal and image normalization, to enhance the image quality, enabling better analysis and interpretation. The filter type was selected according to the specific noise characteristics present in each image. To address different noise types, such as ‘salt-and-pepper’ noise and normally distributed Gaussian noise, we employed noise reduction techniques grounded in the mathematical principles of median filtering and Gaussian filtering (details are provided in Section 3.2.2).

3.2.1. Image Equalization

Before proceeding with further operations, we applied histogram equalization, as shown in Equation (1), to adjust the image intensities and enhance contrast. This technique ensures a more uniform distribution of pixel intensities, thereby improving the clarity and visibility of the processed images. Following the equalization process, we implemented the center square cropping method for additional image processing, which focuses on the most relevant image region while reducing noise and irrelevant background details [26].

$$N (P(x y))=round\left({\displaystyle \frac{\mathit{c}\mathit{d}\mathit{f}\left(\mathit{p}\left(\mathit{x}\mathit{y}\right)\right)-{\mathit{c}\mathit{d}\mathit{f}}_{\mathit{m}\mathit{i}\mathit{n}}}{\left({\mathit{R}}_{\mathit{x}}\times {\mathit{C}}_{\mathit{x}}\right)-{\mathit{c}\mathit{d}\mathit{f}}_{\mathit{m}\mathit{i}\mathit{n}}}}\times \left(\mathit{L}-1\right)\right)$$

(1)

where $N \left(P\right(x y\left)\right)$ refers to the new intensity value after the histogram equalization has been applied at the pixel value (x,y) in the image. This function ensures that pixel intensities in the image are redistributed to enhance the contrast. The cumulative distribution function $cdf\left(p\right(x,y\left)\right)$, which maps the original image intensity to new intensity, indicates the cumulative probability of pixels with intensity values from 0 up to the maximum intensity at position $(x,y)$. The minimum CDF, ${cdf}_{min}$, refers to the minimum non-zero value in CDF, preventing images from becoming too dark by avoiding low contrast in dark areas. The term $R_x$ × $C_x$ represents a product of rows $R$ and columns $C$, accounting for the total numbers of pixels in the image. The value $L-1$ represents the maximum possible intensity level, which ensures that the normalized CDF fits within the desired intensity range. The function $round(\dots )$ is applied to obtain an integer value, ensuring that the output intensity level is valid within the range from 0 to $L-1$.

3.2.2. Noise Removal

In this research, a Gaussian filter and a median filter were applied collectively to achieve comprehensive noise reduction while preserving image quality. Gaussian filtering smooths the image by applying a weighted average across neighboring pixels, effectively reducing high-frequency noise. However, it is less effective at addressing impulsive noise and tends to blur edges. Conversely, the median filter excels at removing impulsive noise by replacing each pixel with the median value of its surrounding neighbors, thereby preserving the edge sharpness. By combining these techniques, Gaussian filtering was utilized for general noise reduction and image smoothing, while median filtering was applied to retain edge sharpness and address impulsive noise. This hybrid approach successfully mitigates both high-frequency and impulsive noise, ensuring minimal edge blurring and preserving critical image details to facilitate improved analysis and interpretation [27]. The median filtering is defined as illustrated in Equation (2):

$$\mathit{O}(\mathit{x},\mathit{y})=\mathit{M}\mathit{e}\mathit{d}\mathit{i}\mathit{a}\mathit{n}\left\{\mathit{I}\right(\mathit{x}+\mathit{i},\mathit{y}+\mathit{j}),\mathit{i},\mathit{j}\mathit{€}\mathit{W}\}$$

(2)

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

$$\mathbf{G}(\mathbf{x},\mathbf{y})={\displaystyle \frac{1}{2\mathsf{\pi }\mathsf{\sigma }²}}{\mathit{e}}^{\left\{{\displaystyle \frac{-(\mathbf{x}²+\mathbf{y}²)}{2\mathsf{\sigma }²}}\right\}}$$

(3)

In this formula, $G(x,y)$ represents the Gaussian function applied at a specific pixel coordinate $(x,y)$. This function assigns weight to each pixel based on its distance from the center of the Gaussian kernel, with pixels closer to the center having a greater influence and those further away having a smaller effect. This weighted influence creates a smooth transition across pixel values, preserving the main structure of the image while reducing fine details. The standard deviation $\sigma$ controls the spread or width of the Gaussian curve. A larger $\sigma$ results in greater smoothing, as the filter incorporates a wider area around each pixel, thereby softening more details. In contrast, a smaller $\sigma$ provides less smoothing and preserves finer details within the image. Adjusting $\sigma$ thus offers control over the degree of blurring applied to the image. The $\sigma$ refers to the standard deviation, whereas the $e$ refers to the mathematical constant (approximately equal to 2.71828) [27]. The term ${\displaystyle \frac{1}{2\mathsf{\pi }\mathsf{\sigma }²}}$ in the Gaussian formula normalizes the function, ensuring that the area under the Gaussian curve sums to 1. This normalization is essential for maintaining the overall brightness or intensity of the image after filtering. Without normalization, applying the Gaussian filter could unintentionally alter the image’s brightness, either darkening or lighting it.

3.2.3. Normalization

Image resizing is performed using bilinear interpolation to standardize the dimensions of images and fit them to the required input size of 224 × 224 [17]. This resizing method ensures uniformity across the dataset, facilitating compatibility with the deep learning model’s architecture while reducing computational complexity during training. Bilinear interpolation calculates the value of a new pixel by taking a weighted average of the four nearest neighboring pixels in the original image. This technique effectively preserves critical image details and maintains overall quality by minimizing distortions or artifacts that could arise during resizing. By standardizing the dimensions while retaining the image integrity, bilinear interpolation ensures that the resized CRMD images remain representative of their original characteristics, optimizing them for analysis and interpretation in the model. The bilinear interpolation can be represented by Equation (4):

$$\mathit{f}(\mathit{x},\mathit{y})=\mathit{a}\ast \mathit{x}\ast \mathit{y}+\mathit{b}\ast \mathit{x}+\mathit{c}\ast \mathit{y}+\mathit{d}$$

(4)

where $\mathit{f}(\mathit{x},\mathit{y})$ refers to the pixel value at a specific point on the resized image, $\mathit{x}$ and $\mathit{y}$ are the coordinates of the point on the original image, and $\mathit{a},\mathit{b},\mathit{c}$ and $\mathit{d}$ are coefficients calculated based on the nearby pixel values in the original image.

3.2.4. Segmentation

As shown in Figure 2, we segment the CRMD region from the background using color-based image segmentation techniques. This separates image regions based on a specified color range, allowing us to isolate the target area effectively. The segmentation process is defined as follows:

$$\mathrm{m}\mathrm{a}\mathrm{s}\mathrm{k}(\mathrm{i},\mathrm{j})=(H_{\left(\mathrm{i},\mathrm{j}\right)}\le H_{\mathrm{h}\mathrm{i}\mathrm{g}\mathrm{h}})\mathrm{˄}\left(H_{\left(\mathrm{i},\mathrm{j}\right)}\ge H_{\mathrm{l}\mathrm{o}\mathrm{w}}\right)\mathrm{˄}(S_{\left(\mathrm{i},\mathrm{j}\right)}\le S_{\mathrm{h}\mathrm{i}\mathrm{g}\mathrm{h}})\mathrm{˄}\left(S_{\left(\mathrm{i},\mathrm{j}\right)}\ge S_{\mathrm{l}\mathrm{o}\mathrm{w}}\right)\mathrm{˄}(V_{\left(\mathrm{i},\mathrm{j}\right)}\le V_{\mathrm{h}\mathrm{i}\mathrm{g}\mathrm{h}})\mathrm{˄}\left(V_{\left(\mathrm{i},\mathrm{j}\right)}\ge V_{\mathrm{l}\mathrm{o}\mathrm{w}}\right),$$

(5)

where mask (i, j) is the binary mask indicating whether a pixel $(i,j)$ is part of the segmented region of CRMD. Here, $H (i, j)$, $S (i, j)$, and $V (i, j)$ are the Hue, Saturation, and value components of the pixel at position ($i, j)$. The thresholds $H_{low}$, $H_{high}$, $S_{low}$, $S_{high}$, $V_{low}$, and $V_{high}$ are the lower and upper bounds for each respective color channel.

3.2.5. Data Augmentation

The goal of data augmentation is to increase the diversity and variability of the training data, which helps to reduce overfitting and improve the model’s generalization capability [28]. This process is vital for improving machine learning models’ ability to perform effectively on unseen data, as demonstrated by [29]. We used experimental datasets to prepare ground truth data for specific locations, tackle specific problems, and validate models’ performance within those contexts. This approach enabled us to customize the data to reflect the unique characteristics of the region and the problem, ensuring that the model is both relevant and accurate with regard to the intended task. To validate the effectiveness of this approach, we compared the model’s performance in two scenarios: Scenario 1, which used the original dataset, and Scenario 2, which incorporated augmented data.

3.3. Algorithm Usages and Model Optimization Methods

ResNet-50 is a CNN comprising 50 layers, which is specifically designed to address challenges in deep learning architectures, such as the vanishing gradient problem observed in other CNNs [9,30]. It processes images at a uniform size, with a default input of 224 × 224 pixels. The network begins by extracting primary features through a 7 × 7 convolution kernel, followed by deeper feature extraction through the bottleneck residual block structure for efficient feature extraction across layers. To enhance convergence speed and improve learning efficiency, each convolutional layer is paired with batch normalization and ReLU activation functions, ensuring consistent data distribution between training and testing images. In its final stage, ResNet-50 employs the Softmax function in the fully connected classification layer to map output values to a probability range of 0 to 1, which is critical for accurate loss function computation. This architecture enables ResNet-50 to excel in complex image recognition tasks, offering robust and efficient feature extraction and classification [31].

To optimize the proposed model, we implemented early stopping, batch normalization, and dropout techniques. Batch normalization is a training optimization method that standardizes the input data for each layer, improving the stability and efficiency of the training process. This technique normalizes each batch of data to have a mean of 0 and a variance of 1 before passing the batch to the next layer, ensuring that the model learns consistently and mitigates issues related to vanishing or exploding gradients. For a given layer with d-dimensional input $x=(x^{\left(1\right)}, \dots \dots \dots , (x^{\left(d\right)}),$ each dimension is normalized individually as Equation (6):

$${\widehat{x}}^k={\displaystyle \frac{{\mathit{x}}^{\mathit{k}}-\mathit{E}\left[{\mathit{x}}^{\left(\mathit{k}\right)}\right]}{\sqrt{\mathit{v}\mathit{a}\mathit{r}\left[{\mathit{x}}^{\left(\mathit{k}\right)}\right]}}}$$

(6)

where ${\widehat{x}}^k$ represents the normalized value of the $k-th$ dimension, $E\left[x^{\left(k\right)}\right]$ is the mean, and $var\left[x^{\left(k\right)}\right]$ is the variance of the $k-th$ dimension within the batch.

Early stopping is a regularization technique that prevents overfitting by halting the training process when the validation accuracy no longer improves after a predefined number of epochs. This approach ensures that the model does not overfit to noise in the training data, thereby enhancing its generalization capabilities for unseen data. By avoiding unnecessary training epochs, early stopping also reduces computational costs and training time. Moreover, it improves the model’s robustness by minimizing the effects of noise and fluctuations in the training dataset. When combined with techniques like batch normalization, early stopping stabilizes and accelerates the training process by ensuring consistent inputs to each layer, ultimately improving the model’s efficiency and accuracy.

3.4. CNN Algorithm Usages

In this study, we evaluated the performance of five CNN algorithms—Resnet50, VGG19, VGG16, CNN, and AlexNet—for classifying the severity of CRMD and recommending appropriate pesticide doses. We conducted two sets of experiments: one without optimization and another with algorithmic optimization. Subsequently, we identified the optimal model, which demonstrated superior performance across key metrics, including accuracy, recall, precision, and F1 score.

For the expert system, we utilized a rule-based decision-making approach grounded in predefined rules established by experts from Haramaya University. For example, when the model identifies a high-severity case of maize rust disease, the system cross-references these expert-provided rules to determine the appropriate fungicide and its recommended dosage. This methodology is consistently applied across all severity levels, ensuring that treatment recommendations are aligned with expert knowledge and tailored to the specific severity of the detected disease. By integrating expert knowledge, the system delivers precise and contextually accurate recommendations, enhancing the reliability and practical utility of the model.

3.5. Model Evaluations

In this work, we use different metrics for evaluating the performance of the proposed models: accuracy, precision, recall, and F1-score. Precision is the ratio of true positives (TPs) to the sum of true positives and false positives (FPs). It focuses on the accuracy of the correct prediction of CRMD, measuring how often the model correctly identifies CRMD cases without mistakenly classifying other classes as a specific class of CRMD. Recall is the ratio of TP to the sum of TP and false negatives (FNs). It emphasizes the model’s ability to identify all relevant positive cases, so a high recall is achieved when the model captures most or all the actual positives. Recall increases when the model correctly captures more instances of the positive CRMD class, thereby minimizing the number of positive cases misclassified as negative. F1-score is the harmonic mean of precision and recall. It is particularly useful when both the precision and recall of CRMD are high, as this indicates that the model is effectively classifying CRMD cases while minimizing classification errors across CRMD classes. Accuracy provides an overall measure of the model’s correctness across all CRMD’s classes. Since the CRMD classes do not present any issues of class imbalance, accuracy is a reliable metric for evaluating our proposed model. The mathematical formulas for each evaluation metric are provided in Equations (7)–(10).

$$\mathrm{Precision}={\displaystyle \frac{TP}{TP+FP}}$$

(7)

$$\mathrm{Recall}={\displaystyle \frac{TP}{TP+FN}}$$

(8)

$$Accuracy={\displaystyle \frac{TP+TN}{TP+FP+FN+TN}}$$

(9)

$$\mathrm{F}1-\mathrm{score}=2\ast {\displaystyle \frac{(Recall\ast Precision)}{(Recall+Precision)}}$$

(10)

To assess the effectiveness of the expert system, a panel of agricultural experts reviews its outputs. These experts evaluate whether the system’s recommendations align with established knowledge and best practices for diagnosing and treating maize rust disease. The evaluation emphasizes the accuracy and appropriateness of the recommended fungicides and dosages based on varying severity levels. By comparing the system’s suggestions with the real-world expertise of the specialists, the evaluation ensures that the expert system provides reliable, practical, and field-ready solutions for effective disease management.

4. Experimental Setup and Data Collection

4.1. Image Acquisition

We collected 5000 maize leaf images from the Haramaya University Rare Research Center using smartphones (Samsung 20 AS (Samsung, Thai Nguyen, Vietnam), Huawei p40 Pro (Huawei, Shenzhen, China), Galaxy S8 (Samsung Electronics, Retro City, China)) under the university’s authorization. The center, located at 9°0′ north latitude and 42°03′ east longitude in Ethiopia, served as the primary site for image collection. To enhance the dataset’s size and variability, an additional 3000 images were generated using data augmentation techniques, resulting in a comprehensive dataset of 8000 images across five severity classes (Figure 3). The classification of the dataset was based on the CIMMYT scale [32], with plant pathologists categorizing the images into five levels: healthy (no visible symptoms), low severity (symptoms covering > 2.5% of the leaf area), medium severity (symptoms covering > 10%), high severity (symptoms covering > 35%), and very high severity (symptoms covering > 75%).

The dataset was evenly organized into 1600 images per class—healthy, low, medium, high, and very high—resulting in a total of 8000 images, as outlined in

Table 1. Classes and datasets.

Severity	Datasets
Healthy	1600
Low	1600
Medium	1600
High	1600
Very high	1600
Total	8000

. To assess the model’s performance under varying conditions, the dataset was split into various ratios (70:30, 80:20, 85:15, and 90:10) for training, validation, and testing. These splits provided insights into the model’s robustness and generalizability across different data allocations.

4.2. Experimental Setup

This experimental study evaluated the effects of batch normalization, dropout, and early stopping techniques on image severity classification, focusing on improving generalization and preventing overfitting. Key stages of the experiments included image preprocessing, data augmentation, segmentation, and feature extraction. The classification process employed the Softmax activation function and the Adam optimizer, which was chosen for its adaptive learning rate and superior regularization capabilities. A comparative analysis with stochastic gradient descent (SGD) revealed that Adam outperformed SGD in terms of both accuracy and computational efficiency, highlighting its effectiveness in minimizing overfitting and enhancing model performance.

Table 2. RaseNet50 Performances in different data ratios.

Training–Testing Ratios (%)	Accuracy (%)	Precision (%)
70:30	44	36
80:20	82.5	86
85:15	97.94	97
90:10	77	72

presents the accuracy and precision of the ResNet50 algorithm under various training-to-testing ratios, including 70:30, 80:20, 85:15, and 90:10. The 85:15 ratio consistently delivered the best performance, achieving the highest accuracy and precision. Based on these findings, the final model configuration was optimized using 100 epochs, a batch size of 32, and the 85:15 data split, ensuring robust training and evaluation outcomes.

As shown in

Table 3. Time complexity of Resne50 on Adam and SGD optimizers in different epoch size.

Epoch Size	Accuracy/Time Complexity in Different Optimizers and Epochs
	Adam		SGD
	Accuracy	Time	Accuracy	Time
50	91.94	218 s, 2 s/step	62.94	265 s, 5 s/step
100	97.94	217 s, 2 s/step	72.53	221 s, 2 s/step
150	95.01	282 s, 5 s/step	79.33	300 s, 5 s/step
200	96.70	345 s, 5 s/step	85.80	360 s, 5 s/step

, the model was trained using various configurations, with the number of epochs set at 50, 100, 150, and 200, while maintaining a batch size of 32. Increasing the number of epochs beyond 100 resulted in diminishing returns, with minimal accuracy improvements and a significant rise in computational complexity. Conversely, fewer epochs yielded lower accuracy, establishing 100 epochs as the optimal choice for balancing performance and efficiency. Table 3 further highlights ResNet50’s time complexity and accuracy across different training-to-testing ratios. At 100 epochs, ResNet50 exhibited lower time complexity and superior accuracy compared to larger epoch sizes. However, when the epoch sizes increased from 100 to 150 and 200, the time complexity rose significantly, and the model’s learning rate slowed compared to that at 100 epochs. Based on these findings, 100 epochs were selected as the ideal configuration for evaluating the proposed system.

This systematic computing was conducted in the Anaconda environment, an open-source platform for the Python programming language. Jupyter Notebook was used to write the Python code, with Keras (a free-source CNN library) and TensorFlow serving as the backend to implement the prototype.

To evaluate the expert system, we relied on domain specialists’ expertise in maize cultivation and CRMD diagnosis. These experts contributed documented guidelines and best practices for identifying and managing the disease, forming the foundation of the expert system’s decision-making rules. The expert system was seamlessly integrated with the classification model, enabling it to combine machine learning outputs with expert-driven decision rules. This hybrid approach ensured precise disease classification and tailored treatment recommendations, enhancing the system’s reliability and accuracy in managing maize rust disease.

5. Results and Discussions

5.1. Experiment One: Without Optimizers

This experiment was conducted without incorporating three widely used optimization techniques in deep learning: batch normalization, dropout, and early stop.

Table 4. Different algorithms performances (before optimization).

Algorithm	Using Experimental Datasest				Using Existing Dataset (Accuracy)
	Accuracy (%)	Precision (%)	Recall (%)	F1-Factor (%)	Accuracy (%)
ResNet50	80.20	83	80	80	80.30%
VGG-19	85.33	88	87	87	81.75%
CNN	87.83	88	87	87	84.01%
VGG-16	76.16	80	76	76	81.00%
Alexnet	86.58	88	87	87	80.12%

presents the severity classification accuracies under these conditions, offering a baseline understanding of the model’s performance across different datasets. To evaluate the proposed models, we tested them on both experimental and publicly available datasets. The publicly available dataset, sourced from the Kaggle data repository and comprising 4188 data points, was used alongside the experimental dataset. For the experimental dataset, the accuracy, precision, recall, and F1-score ranged between 76% and 88%. Performance metrics across both datasets—including accuracy, precision, recall, and F1-score—remained below 90%, largely due to the absence of optimization techniques. The publicly available dataset achieved accuracies ranging from 80% to 85% across different models. This was primarily attributed to the smaller dataset size, the lack of data augmentation, and class imbalances, where underrepresented classes negatively impacted the models’ overall performance and generalization capabilities. These findings establish a baseline for comparison with Experiment Two, where optimization techniques such as batch normalization, dropout, and early stopping were applied to improve the models’ accuracy and robustness.

Figure 4 is the confusion matrix used to show the experimental results for the Resnet-50 algorithm. The matrix reveals some misclassified instances, such as maize leaves with very high severity being incorrectly classified as having low severity (three items). Additionally, a number of maize leaves with low severity were misclassified as having high (136 items) or medium (17 items) severity. This misclassification could lead to incorrect pesticide treatments, potentially harming the maize crops or leading to resource wastage. Identifying and addressing these errors is critical for improving the model’s precision and ensuring accurate recommendations for severity-based pesticide applications.

Figure 5 reveals the training and validation performance across various models. ResNet50 and CNN show relatively better performances in terms of accuracy and lower variability compared to other algorithms; however, their accuracies still fall below 90%. Furthermore, significant fluctuations in the training and validation curves suggest overfitting issues, where the models perform well on training data but struggle to generalize effectively to unseen data.

5.2. Experiment Two: Using Optimizers

In this experiment, we trained the proposed model using batch normalization, dropout, and early stopping to enhance performance and mitigate overfitting, as shown in

Table 5. Selected parameter setups.

Parameter	Value
Optimizer	Adam
Epochs	100
Batch size	32
Learning rate	0.0001
Dropout	0.2

. We fixed the initial learning rate at 0.0001 (1 × 10⁻⁷) and applied a dropout rate of 0.2. The selected parameters are shown in Table 5.

Table 6. Summary of different algorithms’ performances (after optimization).

Using Experimental Datasets					Using Existing Dataset
Tested Algorithms	Accuracy	Precision	Recall	F-Score	Accuracy
Resnet50	97.94%	97%	97.16%	97%	95.8%
VGG-19	94.99%	95%	95%	95%	90.66%
CNN	92%	92%	92%	92%	89.22%
VGG-16	93%	94%	93%	93%	93%
AlexNet	93.5%	95%	94%	94%	92.3%

summarizes the performance of various algorithms after applying regularization techniques to both experimental datasets (8000 images) and publicly available datasets (4188 images). Among the models tested on the experimental dataset, the ResNet50 model demonstrated the best performance, achieving an impressive accuracy of 97.94%, along with a precision of 97%, recall of 97.16%, and an F1-score of 97%. This demonstrates its superior ability to classify maize rust disease severity compared to other algorithms. VGG-19 followed closely behind with an accuracy of 94.99% and consistent precision, recall, and F1-score values of 95%, highlighting its reliability as a strong alternative. The CNN model performed well but with slightly lower metrics, achieving 92% across all evaluation metrics. Similarly, VGG-16 and AlexNet delivered solid results with accuracies of 93% and 93.5%, respectively, alongside strong precision, recall, and F1-scores, although they lagged behind ResNet50. The results from Experiment 2, with regularization techniques applied, show significant improvement over Experiment 1, where no such optimizations were included. ResNet50 consistently outperformed all other algorithms due to its advanced feature extraction and architectural advantages, proving to be the most suitable model for maize rust severity classification. Its accurate predictions will allow farmers to determine appropriate pesticide dosages based on severity levels, enhancing disease management efficiency and reducing unnecessary pesticide use, thereby aligning with sustainable agricultural practices.

After applying optimization techniques to the existing dataset, we observed notable performance improvements. However, the results from the publicly available datasets were less accurate compared to those obtained from the experimental datasets. This discrepancy is likely attributed to the smaller size of the publicly available dataset, which limits the diversity and variability of the training samples. Despite these challenges, the model exhibited good generalization across various data samples, demonstrating improved overall accuracy and robustness in each dataset type.

Figure 6a reveals that out of 1200 tested images, only 38 were misclassified, indicating that ResNet50 achieved over 97% accuracy in classifying maize rust disease severity levels. Figure 6b shows that VGG19 misclassified 60 images, demonstrating slightly lower accuracy than ResNet50. Figure 6c,d show even higher error rates, with 95 and 84 wrongly classified images, respectively, while Figure 6e indicates 69 misclassified images for another algorithm. These results highlight that ResNet50 consistently outperforms the other models across confusion matrices and evaluation metrics, benefiting from the integration of optimization techniques like dropout, batch normalization, and early stopping.

In our experiment, ResNet50 achieved higher accuracy than other models due to its deeper architecture, which allows it to capture more complex features from image data. Additionally, the use of residual connections in ResNet50 helps mitigate the vanishing gradient problem that often occurs in deep networks. This, combined with the advantages of pre-trained models, enabled ResNet50 to outperform the standard CNN model in our tests.

As indicated in

Table 7. ResNet50 test result comparison: before and after data augmentation.

Scenarios	Accuracy (%)
Scenario 1	91
Scenario 2	97.94

, we compared the model’s performances in various scenarios to verify its effectiveness. Scenario 1 represents the model’s performance before data augmentation, while Scenario 2 reflects its performances after applying data augmentation, as demonstrated in the experiment. The comparison highlights a notable improvement in Scenario 2, confirming the benefits of data augmentation in enhancing model accuracy and overall performance.

Table 8. Time complexity of different models in various scenarios.

Models	Average Speeds per 100 Epoches (Before Early Stopping)	Average Speeds per 100 Epoches (After Early Stopping)
	Training	Training
Resnet50	227 s	18 s
VGG-19	654 s	36 s
CNN	173 s	105 s
VGG-16	355 s	45 s
AlexNet	108 s	108 s

presents the time complexity of various model development processes, comparing the average training speed over 100 epochs before and after applying early stopping optimization techniques. ResNet50 exhibited a dramatic reduction in training time, decreasing from 227 s to just 18 s with early stopping, as early stopping effectively eliminates unnecessary iterations. Similarly, VGG-19 showed a substantial decrease, reducing the training time from 654 s to 36 s. The CNN model experienced a more modest reduction, with the training time dropping from 173 s to 105 s. VGG-16 also benefited significantly from early stopping, reducing the training time from 355 s to 45 s. In contrast, AlexNet’s training time remained consistent at 108 s per 100 epochs, suggesting that early stopping had no effect on its speed. Overall, the results indicate that early stopping optimization can significantly accelerate training for most models, although the impact depends on the specific model architecture.

5.3. Maize Common Rust Pesticide Dose Recommendation

We developed two methods for recommending fungicides for CRMD using Gradio, a web-based platform for creating personalized interfaces for deep learning models. The first method utilizes a pre-trained model integrated into a Gradio prototype interface, as shown in Figure 7. This interface enables users to input the severity level of rust disease, which may be classed as “healthy”, “low”, “medium”, “high”, or “very high”. Based on this input, the system recommends the appropriate fungicide dose. This method offers a user-friendly approach for selecting the disease stage and receiving tailored treatment recommendations, assisting farmers in applying the correct pesticide dosage.

The second method involves using Gradio interfaces in real time for image-based fungicide recommendation. In this case, users can upload an image of a maize leaf using a smart phone or from saved images. The system processes the uploaded image to determine the disease severity and subsequently recommends an appropriate fungicide dose based on the assessment. This method simplifies the process by enabling users to receive recommendations directly from the image, offering seamless and efficient solutions for real-time diagnosis and treatment. Both methods provide accessible and user-friendly platforms for farmers and agricultural experts, effectively aiding in the management of rust disease in maize.

The agricultural experts verified the accuracy of the recommendation system, which demonstrated 100% accuracy for both model output testing and real-time image application. The system was tested across various stages of maize leaf health, consistently delivering precise results in each scenario by leveraging the rule-based recommendation approach. These findings confirm the system’s reliability in diagnosing maize rust disease and recommending appropriate treatments tailored to the severity of the infection.

6. Conclusions

This paper presents a deep learning-based solution that leverages the ResNet50 model to classify the severity of CRMD and recommend appropriate pesticide doses through image processing techniques. Advanced techniques such as dropout, batch normalization, and early stopping were integrated to enhance the model’s performance, leading to significant improvements. The Gradio interface was employed to provide a user-friendly platform that accurately recommends the appropriate fungicide dosage based on the disease severity, enabling practical real-time applications. The ResNet50 model demonstrated an accuracy of over 97%, with minimal variability between training and validation, highlighting its ability to generalize effectively while reducing overfitting issues. The system’s accuracy and reliability were validated through real-time testing and expert evaluations, achieving 100% accuracy in classifying maize leaf conditions and recommending appropriate treatments. By leveraging CNNs, this study showcases the potential of AI-driven tools to revolutionize the monitoring and management of CRMD, offering substantial benefits for farmers and researchers globally. Furthermore, this research underscores the critical role of AI in improving agricultural productivity and food security, particularly for smallholder farmers. It provides a scalable framework for effective disease management while laying the groundwork for future expansions to other crops and agricultural contexts, broadening its impact on global agriculture.