Agriculture is vital to boosting the economy of any nation. In India, agriculture alone makes up 18.8% of the gross domestic product (Economy Survey, 2021). Rice is a leading food crop and the most consumed agrarian product. Paddy is a primary source of sustenance for half the world’s population. Approximately 20% of the world population’s daily calorie demand is fulfilled by rice. Rice is cultivated almost everywhere; 10% of the world’s total agricultural land is used solely for the cultivation of the rice crop. This equates to 164.19 million hectares of land, of which 44 million hectares are found in India. Furthermore, 78% of the total rice production is directly used for human consumption, of which 90% is consumed in Asia only. Rice is traditionally the most substantial part of an Indian meal, and so a major portion of India’s cultivated land is used for rice cultivation. India is the second-largest producer of rice after China and the Indian economy is heavily reliant on rice production. A significant portion of total rice production is exported to other countries. A rapid growth in population increases the demand for agricultural products exponentially. Ultimately, this puts pressure on the agricultural industry to increase productivity. As agricultural land is limited, to increase production, work should be carried out to decrease losses in rice yield. Currently, approximately 20%–100% of rice crop yields are devastated by rice diseases (Dhiman et al., 2023).

Rice is a Kharif crop that performs best in warm, humid weather and flooded areas. This creates a favorable environment for a variety of diseases to thrive. Based on cause, diseases in rice crops can be divided into two groups. Diseases caused by biotic factors or an organism are parasitic. Parasites include pathogens, pests, and weeds. Pathogens, including viruses, bacteria, and fungi, can cause a wide range of diseases. Out of all these diseases, some of the rice diseases are more likely to take hold and severely affect the yield of the crop. Other factors that cause a reduction in the yield of a field are non-parasitic diseases. Non-parasitic diseases are caused by unfavorable temperatures, irradiation, deficiencies of specific nutrients, and water. Alkalinity, bronzing, cold injury, panicle blight, straighthead, and white tip are examples of abiotic diseases.

Diseases can also be classified based on the part of the crop affected. Symptoms of the disease can appear on the stem, panicle, sheath, or leaf (Dutta et al., 2023). All these diseases cause a loss in the yield of the crop. The magnitude of the reduction in yield is directly influenced by the level of severity of the disease. A disease has different levels of severity: mild, moderate, or severe. To increase the production of rice it is essential to prevent these diseases from occurring or to detect the disease and ascertain its severity level before it affects the yield of the crop. The identification or detection of disease requires careful and in-depth observation of different parts of the plant. Previously, it has been difficult to detect diseases as it requires manpower and expertise to identify the disease from the symptoms. However, the recent expansion of the application of computational approaches, facilitated by the rapid development of computer vision techniques, has meant that computer vision-based automation has become a popular method for diagnosing and monitoring plant diseases (Kashif et al., 2023) so that they can be cured before spreading across the whole plant and destroying the panicle. In practical terms, detecting a disease and classifying its severity helps farmers to prevent or cure the disease and determine the potential loss in yield of the crop.

It is very difficult to categorize the intensity of paddy leaf disease. A wide range of studies has produced varied results. This section discusses several of these findings. Lamba, Baliyan, and Kukreja (Lamba et al., 2022a) proposed a novel hybrid classification model combining two popular classification approaches, a convolutional neural network (CNN) and a long short-term memory (LSTM) network. Classification is based on the type of rice leaf disease. The three rice leaf diseases considered in this article are bacterial blight, blast, and leaf smut. An overall accuracy of 97% was achieved by the hybrid model. The images were collected from both fields and online resources. After data collection, the dataset was augmented using a generative adversarial network (GAN). The authors (Lamba et al., 2022b) also investigated the effect of generative adversarial network augmentation on the CNN classification algorithm. For model testing and training, three paddy leaf infections were being considered. In predicting the disease, the classifier had a high accuracy of 98.23%. Patil and Kumar (2022) combined multi-layer perceptrons with the classification algorithm to classify three diseases. Using an image set of 3,200 images, the classifier achieved a correctness rate of 95.31%. Deng et al. (2021) used CNN to classify six different paddy diseases with 91% accuracy. The dataset contained 33,026 self-generated images of six diseases. The training precision was approximately 92% with a CNN architecture model. Priyangka and Kumara (2021) used VGG19, a CNN model, to categorize images according to seven different paddy diseases. The dataset comprised 105 images (15 images for each disease) from three diverse sources. Overall, the success rate was 95.4%. The authors used data extension to increase the size of the image set. Mekha and Teeyasuksaet (2021) used the random forest method and attained a prediction rate of 69%. The dataset contained 120 images of infected leaves. Luo et al. (2019) categorized four diverse paddy illnesses using a model combining CNN and support vector machine (SVM). Using a self-collected dataset of 6,637 images, the authors achieved 96.8% accuracy. Ghosal and Kamal (2020) used CNN to classify three rice ailments and achieved 92.46% accuracy. Goluguri et al. (2021) used CNN combined with SVM for feature extraction and prediction. There were 1,600 images in the dataset. The model had a 97.5% accuracy rate. Bhattacharyya and Mitra (2019) used CNN classification to predict three paddy diseases. The classifier achieved a correctness rate of 94%. There were 1,500 images in the dataset, i.e., 500 of each disease (Baroudi et al., 2021).

Dastider et al. (2021) used lung ultrasounds to classify the severity of COVID-19 illnesses using a CNN–LSTM hybrid model. The auto-encoder network with CNN and LSTM used in this study was proposed as a reliable and noise-free model. Maragheh et al. (2022) developed a hybrid approach for multi-label text classification by combining the most precise features of LSTM with a spotted hyena optimizer. The model was tested on four different datasets. This article also compared six other fusion approaches using LSTM to produce the best performance. She and Zhang (2018) employed the CNN–LSTM hybrid technique for text classification. The LSTM algorithm was used to store historical data. For text classification, Zhang et al. (2018) used an LSTM–CNN hybrid approach. Overall, the success rate was 91%. The classifier was tested against plain CNN models, LSTM simple models, and LSTM–CNN using various filter sizes. The best results were obtained using an LSTM–CNN hybrid model with a filter size of 5X 600 pixel. Tee et al. (2022) provide an overview of various action recognition strategies. The paper describes the mixed system created by combining CNN and LSTM and provides a brief summary of studies that used both strategies. Waldamichael et al. (2022) reviewed 45 publications on the diagnosis of plant disease, which included information on classification techniques, datasets, accuracy, and strategies. Dhiman and Vinod (2022) rused the CNN approach to identify and categorize paddy illnesses. The classifier divided the dataset into three categories: healthy, unhealthy but curable, and unhealthy and incurably intense. The dataset contained 650 sample photos. Hassan and Maji (2022) used the CNN algorithm to classify four paddy leaf diseases. The diseases under consideration were paddy blast, blight, tungro, and brown spots. Babu et al. (2022) used the CNN technique to classify four rice diseases (Adedoyin et al., 2022).

Saidi et al. (2020) employed a CNN–SVM combined approach to detect depression. The CNN–SVM cross-classifier produced a precision rate of 68% using the Distress Analysis Interview Corpus/Wizard-of-Oz (DAIC-WOZ) dataset. The database comprised a training set and evaluation set 2,480 and 560 bytes in size, respectively. In another study, the CNN–SVM integration was used by the researchers to recognize and catalog brain tumors using magnetic resonance imaging (MRI) images (Zhou et al., 2019). According to the experimental results, brain tumors can be categorized with 98.49% accuracy. The combined approach was also compared with other classification approaches. A new adaptive machine (Sun et al., 2017) has been suggested for the categorization of MRIs. The new proposal has an estimated accuracy of 99.5%. The authors used data from Haxby’s functional MRI dataset from 2001. The study by Ahlawat and Choudhary (2020) sought to distinguish between manually written digits. A composite model of CNN–SVM was applied to the Modified National Institute of Standards and Technology (MNIST) image set. This method had an accuracy of 99.28%. For analyzing hyperspectral images, Leng et al. (2016) used a CNN–SVM combined technique. In a study by Niu and Suen (2012), handwritten characters are recognized and classified. The MNIST dataset was utilized to train the classifier, and the digit classification accuracy was 99.81%. The studies reviewed are presented in Table 1 , along with a summary of the cataloging methodology utilized. Jiang et al. (2020) applied a CNN–SVM mixed algorithm to 8,912 images of four paddy diseases with the aim of classifying the diseases. All conditions were labeled as leaf diseases. The CNN–SVM model achieved 96.84% accuracy. Hasan et al. (2019) used CNN and SVM to identify nine paddy crop diseases. A dataset of 1,080 infected leaf images was created. The model categorized the diseases with 97.5% accuracy. Liang et al. (2019) achieved a 99.2% accuracy rate for the detection of paddy blast disorder. The dataset included 3,010 images of healthy and diseased plant leaves (Upadhyay et al., 2022).

A proposed hybrid model for detecting the diseased area and disease severity level of paddy leaves for three infectious diseases, bacterial blight, blast, and leaf smut, was constructed and consists of six modules. The first module is dataset preparation, which comprises dataset collection from both primary and secondary data sources and the distribution of the image set according to the type of disease. It targets the collection of images of paddy leaf infected by three paddy infections. The data is then pre-processed in module two. Pre-processing standardizes the pictures from numerous possessions in an identical format. The third module increases the size of the dataset exponentially using GAN amplification techniques. In the fourth module, disease severity is determined for every leaf in the image set using the segmentation technique. The disease severity level is also annotated on the image in module four. In the fifth module, a multi-class hybrid classification model is generated by combining the characteristics of CNN and SVM. In the last module, the hybrid classifier is trained and verified against the training set and test set individually.

The proportion of clearly specified data points in the set that is being trained is known as the training accuracy. Similar to resolution accuracy, validation accuracy describes the share of data samples that are correctly resolved from another sample. There were two sets in the dataset. The training images were in one set and the validation images were in the other set. Model training and validation were carried out using an 80–20 cross-validation procedure. Multiple mixed-image studies were performed for validation. The productivity of the classifier was tested using a new randomly selected image. The sparse categorical cross-entropy is used as loss function. The accuracy is 98.43% accomplished by the prototypical was 98.43%.

The cross-entropy function of the classifier was optimized using the Adam optimizer. Cross-entropy loss is the most widely used function in deep-learning or machine-learning classification. It aids in evaluating a model’s accuracy in terms of 0s and 1s, from which we may later deduce the probability percentage. Out of the three diseases of the paddy, the model can identify leaf smut with an accuracy of 98.88%, precision value of 91%, recall of 97.7%, and F1-score of 94.23%. For bacterial blight disease, the model recorded an accuracy of 97.766%, precision value of 85%, recall of 84%, and F1-score of 84.5%. For the paddy blast disease, the model recorded an accuracy of 96.77%, precision value of 79%, recall of 84.75%, and F1-score of 81.77%. Figure 9 shows the confusion matrix of the proposed model. The sample size used for the evaluation of the performance parameters was 20% of the images from each category of rice leaf disease.

The diagonally highlighted cells in the confusion matrix specify the number of images of bacterial blight, blast, and leaf smut that were predicted correctly. Other cells provide the number of images of the three rice diseases that were predicted incorrectly, meaning that the prediction was either a true negative or false positive. The proposed model correctly identified 1,991 images of bacterial blight from the dataset. In addition, 1,724 and 189 images of blast and leaf smut, respectively, were correctly identified. Cell (1,2) indicates that 48 images of blast were identified as bacterial blight. Similarly, three images of blast-infected leaves were identified as leaf smut by the proposed model, as shown in cell (3,2). In the case of cells (1,3) and (2,3), the model identified 21 and 50 images of bacterial blight and blast, respectively, as leaf smut-infected images. In cells (2,1) and (3,1), 43 images and 1 image of bacterial blight were wrongly identified as non blast and leaf smut, respectively.

Figure 10 gives the epoch-wise training and validation accuracy and loss value. Figure 10A shows an accuracy graph according to each epoch. The accuracy graph encompasses both the training and validation phase. Figure 10A shows that as the epoch increases, the accuracy of the fusion classifier’s prediction increases. This is because the model is being trained with each epoch. In Figure 10B , the loss curve is shown for the training and validation phase, according to each epoch. As the number of epochs increases and the classifier is trained, the loss function decreases. The precision of the proposed classifier for the classification of paddy disease type and severity increases as the epoch increases.

The performance of the machine was further compared with existing cataloguing approaches using the same image set. Figure 11 shows the accuracy curve and loss curve of various deep-learning classification models for the multi-classification of paddy diseases according to the type and severity of the disease when applied to the dataset created in this study. Two further algorithms were tested in this study: standard CNN and standard SVM for multi-classification. Figure 11 indicates that the proposed hybrid model performed better than either individual approach.

Figure 12 shows the accuracy of various classification models with and without GAN augmentation. Four approaches were compared for accuracy with the proposed classifier model using the same dataset. The correctness achieved by the basic CNN classifier is 96% without GAN augmentation. GAN increased the accuracy of the basic CNN categorizer to 97.17%. A similar effect was seen on the standard SVM classification approach, with a GAN-augmented accuracy of 95.87% and an accuracy of 93% without GAN.

The performance of the proposed hybrid classifier was compared with the most prominent classifiers from the reviewed literature. Table 5 shows the comparison of various approaches used for classification problems with the proposed hybrid model. The greatest accuracy attained with a CNN classifier was 99.45%, which was the best result out of all approaches. After CNN, a hybrid of CNN and SVM performed best with an accuracy of 99.20%. After that, the proposed model achieved an accuracy of 98.43%. A standard CNN with GAN augmentation achieved 98% accuracy in the classification problem. A Deep Convolutional Neural Network (DCNN) and SVM fusion model achieved 97.50% accuracy. A CNN combined with either the Internet of Things or VGG19 classifier attained an accuracy 95% and 95.40%, respectively, in classification problems. A 91% accuracy was achieved with a hybrid model of CNN and LSTM. CNN with transfer learning achieved 92.49% accuracy and random forest model attained 69% accuracy in classification problems.

The proposed classifier was trained over a different number of epochs to study the influence of epochs on classification accuracy. Figure 13 shows the accuracy curves of the proposed hybrid multi-class classifier at different numbers of epochs. Figure 13A shows the accuracy curve at 30 epochs, Figure 13B shows the accuracy curve at 50 epochs, and Figure 13C shows the accuracy curve at 100 epochs of training. There was no major difference in the precision and accuracy of the multi-class classifier at different epochs. The accuracy remained the same and had no major effect of epochs count. The optimal number of epochs for the training dataset is 50. GAN-augmented images from primary sources

GAN has a major influence on the size of the dataset and hence on the accuracy of the model. Figure 14 demonstrates the effect of GAN on the image set for the three paddy infections considered. As shown in Figure 14 , 4% of the total dataset came from primary sources and 27% of the total dataset came from secondary sources. The GAN’s augmentation of the primary images constituted 9% of the total dataset and the GAN’s augmentation of the secondary images constituted 60% of the total dataset.

Table 6 shows the results of the ablation study performed on the proposed model. Its shows that the accuracy of the proposed model is highest using the hybrid model and GAN augmentation. In machine learning, models have many different components, each of which affects the performance as a whole. Therefore, it is crucial to have a means of gauging how much these components contribute to the overall model. This is where the idea of an ablation study comes from, where specific components of the implementing model are removed to better understand the behavior. The proposed model consists of three components. In the ablation study, the effect of each component on the accuracy of the model was evaluated. The three components are GAN augmentation, feature extraction using CNN, and classification using SVM. The model’s accuracy was rated after training and testing using several GAN, CNN, and SVM combinations. The accuracy of the model without GAN augmentation and using SVM as a feature extractor and classifier while importing the same dataset was the lowest, at 93%. The accuracy of the model trained on the same dataset using CNN as the feature extractor and classifier and without GAN augmentation was 96%. Without GAN augmentation, the model using CNN as a feature extractor and SVM as a classifier achieved 97.2% accuracy. The CNN and SVM classifiers achieved 97% and 95% accuracy, respectively, with GAN augmentation.

The biggest threat to agricultural progress is pathogenic diseases, which have a strong influence on overall production quality and quantity. As a result, a computer vision-based automatic diagnosis of rice leaf infections and the extent of infection is increasingly desirable in analytics. Deep-learning techniques, particularly CNNs and hybrid models with a CNN, have shown a promising ability to solve the difficulties in identifying infections. The combination of CNN and SVM was investigated to improve the ability to diagnose blight, blast, and leaf smut diseases in paddy leaves according to four disease severity levels. The image set comprised pictures of all three rice diseases and was compiled from both primary and secondary sources of data. A total of 533 images—202 images of bacterial blight, 218 images of rice blasts, and 113 images of leaf smut—were collected from primary sources. Four standard online repositories were used for secondary data collection. A total of 3,535 images—1,856, 1,599, and 80 images of bacterial blight, blast, and leaf smut infection, respectively—were collected from secondary sources. The dataset was then augmented using a GAN. The GAN increased the dataset from 4,068 images to 9,175 images. The augmented dataset was then pre-processed. Pre-processing comprised the standardization, normalization, and rescaling of the images. All these operations were implemented using the ImageDataAugmentor function of the Keras library. The severity level was then calculated using segmentation techniques. In this study, five segmentation techniques were used: grayscale, threshold segmentation, edge detection, masking, and histogram segmentation. The leaf detection process was accomplished using grayscale, threshold, edge detection, and mask segmentation techniques. The severity level was then annotated on the image. The severity evaluation was carried out using the pixel information from the histogram segmentation.

These images were then fed into the CNN–SVM fusion model for the categorization of the infection type. SVM was used as a classifier, and CNN was employed as a feature extractor. The test accuracy for blight, blast, and leaf smut disease on a sample of randomly chosen photos was 97%, 96%, and 98%, respectively. The results from the proposed hybrid multi-class classifier were compared with other approaches using the same dataset. When compared with supplementary algorithms tested on the same image set, the proposed model yielded the best results. The approaches used for the comparison were standard CNN, standard SVM, standard SVM with GAN, standard CNN with GAN, and CNN–SVM without GAN, and their respective accuracies were 96%, 93%, 95.87%, 97.17%, and 97.23%. To increase its size, 69% of the dataset was generated using GAN augmentation techniques. Secondary sources constituted 27% of the dataset, and primary sources constituted 9% of the dataset. The proposed model helps in the identification of rice leaf disease and the level of disease severity, which can help farmers to apply the appropriate remedies to stop the spread of the disease to other healthy plants. The identification of disease and determination of an exact severity level also enables farmers to predict the degree of loss of crop productivity.

In the future, this methodology can be utilized for the multi-categorization of other plant infections for the same or different crops. The proposed model works with various other datasets for a variety of crops. The proposed approach is useful for predicting the crop yield of a field based on losses due to various crop infections.

A limitation of this study is that the dataset used contained images showing only sections of infected plant leaves. A better method would be to choose only images showing complete leaves from the infected plant for the calculation of disease severity.

The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding authors.

Conceptualization: SL, VK, JR, AB. Methodology: SL, VK, JR, AB, TG. Formal analysis and investigation: SL, VK, JR, TG, JK, AB, DG, SS. Writing—review and editing: SL, VK, JR, TG, JK, AB, DG, SS. All authors contributed to the article and approved the submitted version.