A deep learning optimized model for classification and detection of rice leaf diseases

Abstract

Food productivity, quantity and quality are at stake when plant diseases such as rice diseases undermine the food security. Rice leaf disease treatment necessitates accurate and timely diagnosing. This study describes a deep learning model for categorizing and forecasting rice plant diseases. Using the remora optimization algorithm (ROA) on a rice leaf dataset demonstrates its potential for plant disease classification. The ROA-DM method detects rice leaf diseases using the ROA algorithm, a deep maxout network (DMN), and a deep autoencoder (DAE). ROA is applied to the learning parameters of deep model in order to achieve better convergence and avoiding local minima, which usually happens with conventional gradient-based optimizers. Experiments show that the suggested framework is accurate and precise across illness categories. The confusion matrices display the training and validation accuracy, losses of this model. The performance of our optimal learning method with respect to other methods indicated its potential for identifying leaf diseases. The accuracy of the ROA-DM method is 98.5%.

Introduction

Agriculture contributes to feeding the world by producing food for billions of people as their principal source of nourishment. Rice is an important crop worldwide and many people derive most of their calories from rice. However, the global rice industry continues to deal with a persistent issue, plant diseases that threaten crop production and food security. Among these, rice leaf diseases are a serious menace as they can spread fast and cause huge loss in the fields. The skill to recognize plant diseases is mostly visual assessment done by experienced farmers or plant experts^1,2. However, this method is considered to be time-and effort-intensive in predicting performance. There is a growing demand to develop new strategies for disease control. Efforts to promote computer vision and machine learning have resulted in the development of automatic approaches for detecting and classifying rice leaf diseases^3,4.

The gradual exploration in this area can increase rice production and control outbreaks of diseases substantially and thereby decrease losses⁵. Nevertheless, the use of these procedures can greatly increase efficiency and accuracy in identifying diseases from rice plants. Deep learning classification models are powerful data processors^6,7. While these models are capable, they can have weaknesses such as hyperparameter sensitivity, slow convergence and a tendency to be trapped in local minima due to their use of gradient-based optimization algorithms. These issues provide motivation to develop hybrid models combining deep learning and metaheuristic optimization algorithms which can be advantageous in terms of efficiency on training data as well as generalization ability^8,9,10. Figure 1 presents samples of the rice leaf images used in this study, which cover various types.

Fig. 1.

Representative samples of rice leaf images.

The aim of this paper is introduced to detect and classify rice leaf diseases accurately that combines a DMN classier¹¹ and DAE model¹², and the Remora Optimization Algorithm. The DAE is used to learn a noise and redundancy free compact representation of features from input data. The DMN is a more powerful classifier and can effectively learn complex non-linear decision boundary. The ROA is utilized to train the deep model and leads to better convergence performance and higher classification accuracy with respect to traditional gradient-based optimization methods.

This study has three major contributions:

• To create a large variety of rice leaf image datasets made up of multiple disease types and degrees of severity.

• To develop and use a deep learning model as an optimization method (ROA-DM) for the automatic detection and classification of rice leaf diseases.

• Compare the performance of the proposed framework, training and validation accuracy, training and validation loss, precision, and specificity with existing techniques.

The rest of the paper is organised as follows: Sect.  2 reviews existing methods. The proposed algorithm is presented in Sect.  3. In Sect.  4, we present results from the experiments. The conclusion and future research are presented in Sect.  5.

Literature review

The issue of plant and crop diseases has been a serious problem in worldwide food production. Among agricultural products, rice is an essential food staple for a large fraction of the world’s population. It is imperative to control rice diseases to maintain the yield and food security. The development of automated tools for identifying and classifying disease symptoms on rice leaves has become increasingly attractive. This literature review summarizes the existing literature in this domain, investigating key methods such as machine learning, deep learning, and metaheuristics.

Albattah et al.¹³ proposed an overall system for plant disease classification. This approach is achieved by training a Custom CenterNet framework using DenseNet-77 as the base network. The approach reported in this paper can be divided into three consecutive stages. The first phase is to create annotations to select and outline the region of interest. Furthermore, a simpler CenterNet setup in which DenseNet-77 should be used for learning detailed key points. The Internet of things based on plant disease detection and classification is a one stage detector.

Arguesoa et al.¹⁴ proposed an approach based on deep learning, FSL, to detect and classify plant diseases. The critical points were calculated in the Inception V3 network. The features were then employed to train a multiclass SVM. Early rice disease detection using machine learning involved the use of random forest, SVM, and decision tree classifiers for classification of diseased rice leaves from images using handcrafted features. Although these approaches were found to be useful, the data representation failed to accommodate a variety of subtle disease manifestations. The method shows the capability for robustness in classification of plant diseases. But the paper’s findings must be kept in proportion as they stem from a small dataset. Therefore, further evaluation is necessary to assess its effectiveness on a more extensive and diversified corpus. Curing tomato crop disease was proposed by Agarwal et al.¹⁵ by using a CNN-based structure. The method consisted of three convolutional layers with max pooling to learn prominent features of the input data, followed by classification.

Shah et al.¹⁶ introduced ReSTS, a CNN-based approach for predicting the next slide in STASs. An additional decoder is included in this architecture, together with two separate classifiers (ResTeacher and ResStudent). To train both classifiers, the proposed architecture utilizes a reciprocal training technique. The representation propagated between the ResTeacher and the ResStudent serves as a proxy for determining where parts of the image are more valuable for categorization. An autoencoder is employed to preserve the main features in the image. The autoencoder decomposes and re-paints the image, stripping it of noise and retaining only the important information.

Jiang et al.¹⁷ presented an apple leaf disease localization approach, which has impact on the yield of apple fruits. However, such attempts have to cope with the challenge of finding effective and efficient approaches. In this paper, the author presents a Deep Learning approach with upgraded CNN configuration for apple leaf diseases real-time detection. The experiments were conducted on the Apple Leaf Disease Dataset. The picture annotation and also the data Augmentation were used to build-up this dataset. The experimental results seem to indicate that the proposed INAR-SSD has good detection accuracy (about 78.8%) and high detection speed (23.13 FPS). Thus, the designed method can obtain higher accuracy and a faster detection rate compared with previous methods for early diagnosis of apple plant diseases.

To enable agricultural robotics for automatic disease detection and classification of diseases that affect rice leaves, Simhadri et al.¹⁸ suggested transfer learning based using 15 pre-trained CNN models. Efficient, robust methods to diagnose and manage diseases in rice plants are widely searched by plant pathologists. Farmers often use disease-detection methods based on visual observations, which is a time-consuming process and may be accompanied by improper agricultural practices. Agricultural technology has enhanced the detection of pathogenic organisms in rice plant. Deep learning and machine learning have attracted great attention in crop disease studies.

Hossain et al.¹⁹ proposed an original CNN model for rice leaf disease detection. In all, 4199 images were employed for training and testing of the model, corresponding to five typical diseases of rice leaves. The model had decent accuracy for both training and validation rounds. The model is tested by applying it to photographs taken by the model itself, and we have seen that the obtained recognition rates are acceptable under some conditions. Thenmozhi et al.²⁰ presented a deep CNN model with excellent performance in classifying pests. They contrasted their model with other pre-trained transfer learning models. Upadhyay et al.²¹ put forward disease classification of rice. This information is used to validate leaf lesion shape, colour and size from the photographs. The corresponding CNN-based fully connected method achieves 99.7% accuracy on our dataset.

Chen et al.²² introduced a novel approach, BLSNet for detecting and classifying Rice Bacterial Leaf Streak (BLS) disease. This has wide implications for rice cultivation in that it affects yield and the quality of rice. The experimental results demonstrated that BLSNet was superior in identifying damage and locating damaged locations compared to the other leading models. It may be useful as a bioassay to measure the amount of BLS disease. Stephen et al.²³ introduced four CNN models to class all healthy and diseased leaves. To deal with those gradient problems, researchers used ResNet34 and ResNet50 as the architectures. In the multiclass competition, for the task that defined the proposed ResNet34 model to one’s point, obtained accuracy of 0.98.54% surpassing several approaches.

Li et al.²⁴ developed a deep learning system for real-time detection of diseases and pests on a plant leaf in videos. To find low-quality photos, the method employed Faster R-CNN in conjunction with image-training models. Experiments showed that our custom architecture is very strong in identifying the movies on the system it was not trained at, even when compared to latest state of art approaches²⁵.

Kawasaki et al.²⁶ thus introduced a CNN for the detection of cucumber leaf disease. The accuracy of the system is as high as 94.9%, and it is able to identify and classify cucumber leaf disease well. Despite the positive results found in the literature, previous studies based mainly on image datasets with limited variation. Photographic evidence typically comprises photographs taken only in the laboratory or similar controlled environment rather than actual uncontrolled field situations. Field-cultivated images have commonly variable backgrounds and symptom patterns²⁷.

Bedi et al.²⁸ proposed a new U-Net-based model, namely AResUNet, with the help of residual links to facilitate feature propagation and network learning. This has also been used in medical imaging tasks, e.g., PCOS detection. AResUNet excellently learns the details and delicate structures in ultrasound images, thus it can distinguish between the ovaries and PCOS-affected ones. However, the complexity of this model can raise its costs and finally increase its prediction time, which may challenge the suitability of being used in real-time clinical setting.

Alamoudi et al.²⁹ presented a fusion model that combines ultrasound images with information to decide whether or not the patient has PCOS. MobileNet has been specifically tailored for network computation and is well suited for low computational resource applications. We it uses depthwise convolutions to decrease the number of parameters and computation amount. MobileNet, as a result of its architecture and a reasonably good accuracy is eligible to become promising for PCOS detection. The combined attributes based and image feature extraction using MobileNet architecture model obtained an accuracy of 82.46%.

Preti et al.³⁰ provided an extensive review of the development of insect pest monitoring, particularly in relations of techniques and tools adopted. Fiber optic sensors have been used in these traps as the first-generation device for images. The incorporation of IoT and AI, contemporary information technologies, has facilitated the creation and customization of diverse devices for pest surveillance. Saranya et al.³¹ created a method to identify insects by their thermal radiation emissions using a passive infrared sensor and picture processing. To confirm the pest’s existence in the field, image processing is used to create visual representations of it.

Edelaar et al.³² conducted a study on Azure Sand Grasshoppers to examine their color-changing behavior and their ability to blend into their surroundings as a defensive response to predators. In the context of deep convolutional neural networks, Chen et al.³³ proposed using transfer learning to identify plant leaf diseases. The decision has been made to use the VGGNet model trained on the ImageNet dataset, along with the Inception module, rather than starting training from random weights. This choice is based on using pre-trained networks trained on the well-annotated, massive ImageNet dataset for weight initialization.

Although deep learning has shown promising results in plant disease identification, specialized frameworks that address the challenges of rice leaf diseases are necessary. Our work bridges this gap, by offering a comprehensive deep-learning approach for the identification and categorization of plant disease in rice leaves^34,35. In this work, we apply state-of-the-art deep learning methods on a well-curated dataset with the purpose of providing efficient and accurate resolution to this agricultural issue. Table 1 Comparison study of the drawbacks of available disease-identifying methods.

Table 1.

Comparison study of existing disease identification methods.

Reference	Method	Objective	Dataset Used	Limitations
Taji et al.36	Ensemble hybrid CNN features with BDA, ACO, and MFO optimizers	Improve plant disease classification accuracy using ensemble optimized features	PlantVillage dataset (apple, maize leaves)	High computational complexity and dependency on multiple metaheuristics
Naveenkumar et al.37	Adaptive hybrid segmentation and IRV-WSA optimized transfer learning	Enhance segmentation and classification of plant leaf diseases	UCI, Kaggle, benchmark datasets	Complex architecture and higher preprocessing cost
Khan et al.38	Bayesian optimized CNN and ML hybrid models	Accurate tomato leaf disease classification with optimized hyperparameters	PlantVillage tomato dataset	High dependency on hybrid pipelines and multiple classifiers
Sun et al.39	Improved EfficientNet with CBAM attention and multi-scale fusion	Enhance maize leaf disease recognition accuracy	Maize leaf disease dataset	Focused on single crop and deep model tuning complexity
Al-Gaashani et al.40	Transfer learning with Gravitational Search Algorithm	Optimize feature selection for efficient plant disease classification	Multiple plant disease image datasets	Computational overhead due to metaheuristic optimization

Methodology

The methodology used for the design and evaluation of the deep learning model for detection and classification of rice leaf diseases is presented in this section. The methodology includes data collection, preprocessing, model architecture, as well as training details. The Fig. 2 illustrates the complete ROA-DM pipeline starting from rice leaf image acquisition, followed by preprocessing, DAE-based feature extraction, ROA-driven optimization, and DMN-based disease classification. It shows how ROA optimizes the deep learning framework to achieve accurate detection of healthy and diseased rice leaves with final performance evaluation.

Fig. 2.

ROA-DM framework for rice leaf disease detection.

Data collection and preprocessing

The dataset collected from https://www.kaggle.com/datasets/emmarex/plantdisease (PlantVillage dataset) for algorithm testing in plant disease diagnosis³⁵. The selected rice leaf disease samples from PlantVillage dataset consisting of 3050 colour leaf images across four classes. These categories consist of 500 images of brown spot, 1100 images of common rust, 1000 images of healthy leaf, and 450 images of leaf blight. All images were performed under controlled conditions, with a blank background and no significant changes in light intensity. Several images of the same leaf from random directions are taken. Quality and diversity of the dataset are crucial for the success of any deep learning model for image classification. High-quality datasets are essential for training and evaluating the efficacy of deep learning models for plant disease detection. Dark-field-featured plant diseases, including the PlantVillage datasets, have played a significant role in advancing research in this domain. These datasets include several images of various diseases with different severity levels. The images are high quality rice leaf images having different diseases with various levels of severity.

Before the training, preprocessing was applied to all rice leaf images for normalization and better model performance. All images were rescaled to the same size of pixels. The dataset diversity was increased, and the model’s robustness was improved through data augmentation techniques. To ease convergence during training, image pixels are normalized to a common range, [0, 1] or [− 1, 1]. Table 2 presents the detailed class-wise distribution, percentage representation, and dataset split used in this study.

Table 2.

Detailed description of rice leaf disease dataset.

Class name	Total Images	Percentage (%)	Training images (80%)	Validation images (10%)	Testing images (10%)	Data acquisition type
Brown Spot	500	16.39	400	50	50	Lab-based
Common Rust	1100	36.07	880	110	110	Lab-based
Healthy Leaf	1000	32.79	800	100	100	Lab-based
Leaf Blight	450	14.75	360	45	45	Lab-based
TOTAL	3050	100	2440	305	305	Lab-based

ROA-based deep model intrusion detection

The deep model is used for intrusion detection, incorporating a DMN classifier and a DAE architecture. The Hybrid deep model has demonstrated high efficacy in attaining superior detection results. The deep learning model’s training technique is executed using the suggested ROA. Nature, and more specifically the parasitic habits of remoras, serves as an inspiration for the ROA algorithm, a meta-heuristic method. When two separate host properties interact, the result is the best possible outcome: exploitation and exploration. The process of updating the switch between different modes involves identification and selection.

Moreover, the remora aspect has been designed to support efforts to enhance area regeneration. In contrast, ROA is an optimization methodology that mimics remoras’ hunting methods. The ROA improved its detection results by efficiently mitigating optimization challenges. The deep model is trained to utilize the developed optimization algorithm, ROA-DM.

DMN architecture

To successfully execute the disease detection procedure, the DMN is used. The DMN is designed to reduce the size of the image components considerably. Using the minimization factor is designed to enhance the model’s reliability under limited data conditions. The architectural perspective of the DMN is shown in Fig. 3. In addition, the DMN has several layers, each of which can generate a concealed activation via the max-out function. The leaf image classifier consists of multiple layers, shown in Fig. 3, to classify features. The deep max-out classifier is built by linking these layers in a sequence that proceeds from lowest to highest. This network takes a set of solutions as input and processes them through multiple layers to produce detection results.

Fig. 3.

Architecture of the deep max-out network classifier.

The ultimate output layer performs the computation of the detection result, while the hidden unit’s function is evaluated as follows:

1

2

3

4

The variable represents multiple units located at the layer, while weight is denoted by , and bias is indicated by in the Eqs. (1–4). represents the output of the model of predicted class label and denotes the iteration index.

DAE architecture

The DAE architecture comprises three distinct layers: the output, hidden, and input layers. The training dataset is the source of the data fed to the input layer. The initial layer of the DAE encodes the characteristics of the given dataset. The present model is devised by collectively daisy-chaining auto-encoders. This classifier’s primary objective is to effectively identify the disease within a limited timeframe. The first hidden layer’s job is to sift through all incoming data and extract the most important details for subsequent steps. Within the deep network, the output of the hidden layer is used as input to the hidden layer. Only after finishing auto-encoder training at layer does the layer-wise method begin training an auto-encoder at layer . This is done in order. The final covert layer is a supervised layer that uses a softmax classifier to group attacks into their respective categories. In conclusion, the output layer serves as the model’s output, completing the entire DAE representation in Fig. 4.

Fig. 4.

Architecture of deep autoencoder (DAE).

Moreover, using a nonlinear activation function in the DAE model enables the extraction of more valuable features. The output generated by the DAE model is denoted . represents the outcome of the comprehensive model.

The suggested ROA is used to conduct the training procedure for the deep learning method. The ROA [9] method is based on the parasitic behaviour of the remora, a model organism. Maximal efficacy is attained by using two separate classes: whales and swordfish. The switching between different approaches is efficient by choosing the most easily recognizable movement style. Moreover, the factors of remora are developed to assist in accurately modifying regional revitalization. In contrast, it aims to optimize solutions by leveraging the strategies employed by these marine creatures. The combination of ROA and WOA improved detection outcomes by effectively mitigating optimization difficulties.

The model has two distinctive types of mobilization account for exploitation and exploration. Such self-constructed graduality is applied to mode-switching. During the exploration, remora factors are appended to ensure accuracy and convergence. Thanks to these good measures, ROA’s research is fruitful. The following are a few of the key factors that explain why ROA has such large optimization potential:

The population is initialised by selecting a predetermined number of solutions, denoted by ‘’.

5

The number of solutions is represented by in Eq. (5), and the solution is denoted by .

This strategy of the remora used in SFO is a method on how to achieve the goal: attaching itself with the swordfish to get an update position about its movement through algorithm improved by elite concept. The elitism concept had improved the TLO algorithm’s location update formula as in Eq. (6).

6

iteration be denoted by the variable t, ‘r’ epsilon [0,1] an arbitrary position. In a classical sense with the perspective of updating. Their placement method is historically associated with remora species. For the efficient and effective exploration of the search space, one must have a procedure to explore some location randomly. Remora selects a host according to its historical feeding behavior and whether the current fitness function value is larger than that in previous generations. The experience attack determines the current value of a fitness function. As a behavioral inclination, Tuyu species will normally take few steps around their host(s) to decide if they must switch hosts. The concept is articulated as,

7

denotes spatial location of the generation and is a temporary phase in question. Given that remoras so shift in their behaviour, the host-choices of a remora looks like a broadly-applicable global phenomenon. At this moment it is time to decide. After that, it’s random yay or nay on switching hosts. The inspection consists of a comparison of the measure of fitness with the solution. This type of behavior is reflected by the remora fish, which can feed as its environment dictates and do so not only while seeking such food over time through natural selection in order to optimize its current solution. phenomenon is referred to as

8

if the fitness exceeds that of the current one, it would be required to restart from choosing host as described by next protocol:

9

Exploitation: The WOA algorithm and host feeding technique are also applied in the exploitation stage. The re-sited location of the remora attachment is subject to compliance with the WOA standard.

10

11

12

13

The distance at which the prey is located relative to the predator. The rational number from 0 to 1, decaying reasonably quickly between range of 0 to 2.

Host feeding: The exploitation entails host-feeding and the attention with the focus on the host’s location at that point in time. Moving closer to where the host’s feeding is in small steps that can be represented mathematically for modelling purposes.

14

15

16

17

The variable T indicates the upper limit of iterations, while t represents the overall magnitude. The variable X represents a minute increment of motion, linked to the spatial dimensions of both the remora and the host. To limit the search space for identifying and distinguishing remora from the host organism, the C parameter, also known as the remora factor, is used.

Fitness Function: A fitness assessment determines the optimal solution by finding the minimum value. The fitness function is computed using an expression.

18

The fitness measure is represented by , while the intended output is indicated by . The terms “variables” stand for the sum of all training samples. together with the classifier’s representation of the classification result.

The steps listed are continually carried out until the best solution is reached. The provided pseudo-code, in Algorithm 1 illustrates the ROA that was developed.

Algorithm 1.

ROA-DM pseudo-code.

Experimental result analysis

The system was generated and assessed on an Intel i7 system equipped with a W-2125 CPU and 128 GB of RAM, operating at 4.00 GHz with 4 cores, a high-performance system. The construction and testing procedures were performed using MATLAB R2019a. The experiments used deep models, specifically the DMN Classifier and DAE, trained on the PlantVillage dataset comprising 3050 classes. The dataset is usually split into three groups, i.e., training, testing, and validation. The ratio for a combined set is 80% for training, 10% for testing, and 10% for validation. Adjusting the initial and final layers customizes these models for our use case. According to the model’s architecture, convolutional, softmax, fully connected, and output layers are modified. Most pre-trained models capture generic traits at lower levels, while the top layers are better at capturing task-specific information. The pre-existing model is customized for new jobs by being adjusted to higher levels, thereby capitalizing on information acquired from the first job, and all models are trained with identical hyperparameters. 40 epochs were used for training, selected based on how well the models performed during training and validation.

Using a variety of data partitioning strategies, it was determined whether the models could generalize to a larger, out-of-sample testing set. An examination of the ROA-DM model is presented, using assessment metrics such as F1-score, recall, precision, and testing accuracy. From empirical analysis, find that a two-level binary classification is sufficient for image classification. Termination of training results in evaluating the approach on the test data using popularized classification recall metrics.

Evaluation performance

For the experimental phase of this work, we proceeded to train and test, for disease detection and crops classification, a deep learning framework that we’ve developed. Both the model training and validation are performed on dataset by proposed approach. Furthermore, to verify the effectiveness of our proposed approach, we also compare it with other three state-of-the-arts approaches such as ReSTS¹⁶, CenterNet¹³ and Transfer Learning²² in experiments. Results This section describes the findings of extensive experimental results such as model performance metrics, confusion matrix and comparative evaluations.

True Positives (T_Pos): refers to the count of correctly recognized positive messages, namely disease leaf, that is classified as positive.

True Negative (T_Neg): indicator measures the number of negative communications, namely those categorized as disease leaf, that are correctly recognized as negative.

False Positives (F_Pos): is the calculation of the number of negative messages, which are really healthy leaves, that are mistakenly detected as positive, meaning they are classed as diseased leaves.

False Negative (F_Neg): is the count of positive signals, or diseased leaves, that are mistakenly identified as negative, or healthy leaves.

Precision (P): The percentage of accurate predictions relative to all optimistic predictions is called predictive accuracy. Divide the total number of true positives and false negatives by the total number of true positives to get the accuracy measure.

19

Recall (R): Percentage of positive situations with correct optimistic projections. Recall is the number of truly positive incidents divided by the total number of genuine positives and false negatives.

20

Specificity:

21

Accuracy: Accuracy measures the proportion of correctly classified samples in the testing set. It provides an overall indication of the model’s classification performance.

22

The superiority of the suggested methodology over other state-of-the-art approaches, even when using the best classifier, is evident from the results shown in Table 2, obtained via experimentation on the public dataset. This strategy is mostly motivated by the use of the pre-trained ROA-DM on a dataset which contain an inception module. As a result, this approach fairly takes advantage of the good qualities of both parts. It has a primitive feature extraction function and the other layers for high-dimensional feature extraction and classification. Although pre-trained weights were utilized, rather than training the models from scratch, the best results still must be realized. The proposed method achieves a training accuracy 98.5, validation accuracy of 92.2, and loss of 0.225 after 40 training epochs as depicted in Table 3. Moreover, the model trained using the suggested methodology is used to classify plant disease images outside its training scope.

Table 3.

Various methods of accuracy and loss function after 40 epochs.

Methods	Training accuracy	Training loss	Validation accuracy	Validation loss
ReSTS	72.7	0.858	69.2	0.861
CenterNet	77.9	0.911	70.9	0.934
Transfer Learning	91.6	0.546	82.1	0.662
ROA-DM	98.5	0.089	92.2	0.225

Figure 5 shows the numerical comparison of the performance indices acquired through various approaches. As shown in Fig. 5(a), the proposed ROA-DM model has the highest training performance of 98.5%, which is much better than Transfer Learning 91.6%, CenterNet is 77.9% and ReSTS is 72.7%. Similarly, Fig. 5(b) shows that our model reaches higher validation accuracy of 92.2% versus 82.1%, 70.9%, and 69.2% obtained by Transfer Learning, CenterNet, and ReSTS, indicating the better generalization performance. In terms of the loss performance, Fig. 5(c) indicates that the training loss of ROA-DM is the minimum, which is 0.089 and Transfer Learning, CenterNet and ReSTS obtain higher losses of 0.546, 0.911 and 0.858 respectively. Likewise, Fig. 5(d) verifies that the model with the proposed components have best validation accuracy of 0.225 which is much less than Transfer Learning 0.662, CenterNet is 0.934 and ReSTS is 0.861. These findings demonstrate that the fusion of the Remora Optimization Algorithm enhances convergence effectiveness, increases classification accuracy, and lowers model error when compared to traditional deep learning methods.

Fig. 5.

Comparison the performance analysis of various methods (a) Training accuracy (b) Validation accuracy (c) Training loss comparison (d) Validation loss.

In Table 4, compares the performance of the proposed ROA-DM with existing approaches using standard evaluation metrics such as Accuracy, Precision, Recall and F1-Score and IoU. The experimental results prove the effectiveness and robustness of our proposed ROA-DM model for precise rice leaf disease classification.

Table 4.

Performance comparison of various existing models with proposed model.

Model	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	IoU (%)
ReSTS16	69.2	68.5	67.8	68.1	52.3
CenterNet13	70.9	70.1	69.4	69.7	54.6
Transfer Learning22	82.1	81.3	80.6	80.9	67.8
Proposed ROA-DM	92.2	91.5	90.8	91.1	83.7

Confusion matrix

The confusion matrix of leaf disease classification by our model is provided. Figure 6 shows the confusion matrix of the developed ROA-DM model for four classes; Leaf blight, Brown spot, Common rust and Healthy leaves. The class accuracy for the Leaf blight class was 47 and reflected that in the number of samples being correctly classified, however, Healthy 31 samples and Common rust 16 samples where the classes into which Leaf blight samples mostly misclassified with them but not with other classes having a combined classification. As for Brown spot class, 78 of 100 samples were classified correctly 78% accuracy, with some confusions between Common rust 8 samples and Healthy 12 samples. For the Common rust class 46 of 100 samples were classified correctly while 47 as Healthy, leading to an accuracy of 46%, which shows this class has similar features with healthy. Moreover, the Healthy class had perfect classification of 100% accuracy as each of 100 samples was correctly assigned to its respective class and showed that the model had very good performance on distinguishing healthy from unhealthy cases. In overall, the confusion matrix indicates that resulting ROA-DM model provides the high performance in classification, especially for Healthy and Brown spot classes.

Fig. 6.

Predicted class confusion matrix.

Figure 7 shows the confusion matrix detection results of the developed ROA-DM model. For the Bacterial leaf streak class, 19 of the samples out of 20 were classified accurately with an accuracy rate of 95% and one misclassification. Also, we obtained an 80% accuracy in Rice leaf scald class with 16 correct predictions out of 20 samples and a White tip in Rice where the result was of 90% of accuracy with 18 correct predictions. The Rice leaf smut was 85% accurate in classifying the two classes with a positive prediction of 17 for one of the two categories that were correctly classified. In total, the optimal ROA-DM model correctly predicted a total of 70 out of 80 specimens with an overall accuracy of 87.5%. It proves the strength and competence of the proposed approach for accurate rice leaf diseases detection and classification.

Fig. 7.

Detection of rice leaf diseases using a confusion matrix.

The evaluation metrics for these results are calculated and are given in Table 5. According to Table 5, the average predictive accuracy consistently exceeds 86.61% across multiple experiments. The proposed ROA-DM approach possesses a notable ability to identify rice leaf diseases. The proposed approach effectively identifies healthy and diseased plants and accurately distinguishes among different classifications of rice leaf diseases.

Table 5.

Performance evaluation of various types of leaf disease.

Type of disease	Accuracy	Precision	Specificity
Leaf blight	83.47	47	86.87
Brown spot	89.21	78	92.46
Common rust	91.45	46	95.74
Healthy	82.32	100	83/67

Conclusion

This paper proposed a ROA-DM method, which consists of the Deep Autoencoder, the Deep Max-out Network and Remora Optimization Algorithm for effective identification of rice leaf disease. Through matching superior feature extraction with metaheuristic-guided parameter tuning, the proposed model effectively alleviates the shortcomings of gradient-based light deep learning methods and obtains better convergence properties along with enhanced classification rates. The empirical findings confirm that the ROA-DM method outperforms existing deep-learning models and its accuracy is 98.5%. The new method performed remarkably well compared to a set of existing methods. The collected data of rice diseases results into an overall accuracy of 86.61% under complex environmental background in class prediction. Extensive experiments and comparisons have demonstrated the effectiveness of our framework, which presents an applicable solution to rice disease management.

Although the ROA-DM model has achieved high classification ability it has relatively higher computational complexity than metaheuristic and deep learning-based techniques due to the incorporation of metaheuristic with deep learning that makes its training time longer and more resource consuming. Moreover, its effectiveness will be impaired by the size of dataset, image quality and differences in real field. Future work will involve development of a light model which should be faster, acceleration the ROA convergence speed and validation of our framework on larger real time filed collected data sets for robustness poo to apply in smart agriculture system.

Author contributions

Conceptualization, S. C., U.K., B. M. S.,methodology, S. C., U.K., B. M. S.,software, S. C., U.K., B. M. S.,validation, S. C., U.K., B. M. S.,formal analysis, S. C., U.K., B. M. S.,investigation, S. C., U.K., B. M. S.,resources, S. C., U.K., B. M. S.,data curation, S. C., & B. M. S.,writing—original draft preparation, S. C., U.K., B. M. S.,writing—review and editing, S. C., U.K., B. M. S.,visualization, B. M. S., & U.K.,project administration, S. C., U.K., B. M. S.,All authors have read and agreed to the published version of the manuscript.

Funding

The authors have no competing financial interests in this paper.

Data availability

Dataset collected from https://www.kaggle.com/datasets/emmarex/plantdisease.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

not applicable.

Consent for publication

Not applicable.