Deep learning-based approach for automated detection of lumpy skin disease
* Author to whom correspondence should be addressed.
Sigma Journal of Engineering and Natural Sciences 2026, Vol. 44, Issue 2, pp. 856-868; doi.org/10.14744/sigma.2025.00171
Abstract
Introduction
Lumpy disease, also known as lumpy skin disease (LSD), is a viral disease that affects cattle and is caused by the capripoxvirus. Early detection of Lumpy disease is essential to prevent its spread among cattle and reduce economic losses for farmers [1]. The most common detection method is through visual observation of the characteristic skin lesions. More reliable methods of detection include
laboratory testing of blood or tissue samples using techniques such as PCR (polymerase chain reaction) and ELISA (enzyme-linked immunosorbent assay) [2]. Cattle, domestic water buffaloes across Africa and even camels [3] were more sus- ceptible to a viral illness known as lumpy skin disease. The migration of infected animals may be responsible for the long-distance spread of LSD. Still, the disease’s contrasting seasonal patterns point to the possible involvement of arthropod-borne transmission in its rapid
*Corresponding author. *E-mail address: lokeshmalviya@vitbhopal.ac.in. This paper was recommended for publication in revised form by Editor-in-Chief Ahmet Selim Dalkilic Published by Yıldız Technical University Press, İstanbul, Turkey © Author. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).
Sigma J Eng Nat Sci, Vol. 44, No. 2, pp. 856−868, April, 2026
and aggressive short-distance expansion. Recent research suggests that the common house fly, Musca domestica, and other blood-sucking arthropods may have a role in LSDV transmission; however, this has not been shown in a clinical context [4]. Bangladesh reported the first case of the disease in South Asia in July 2019. In August 2019, India announced the first-ever case of Lumpy skin disease. Since May 2022, LSD has spread over various states in India [5]. Vaccination, bovine animal movement restriction, and quarantine are among of the current preventative measures being used, along with the implementation of biosecurity through vector management. In India, the standard method of treatment is vaccination with a goat pox vaccine that provides 60-70% protection against Lumpy skin disease. This vaccine takes 15-20 days to become fully effective. The Indian Council for Agriculture Research National Research Centre on Equines (ICAR-NRCE) in Hisar (Haryana) and the Indian Vet- erinary Research Institute (IVRI) in Uttar Pradesh have been working together to produce an indigenous vaccine called Lumpi-ProVacInd. There is no proof that it has any effect on people. There is a greater variation in sickness and mortality rates among cattle in the current epidemic in India compared to the previous outbreak [6]. LSD may also reduce the export of livestock and livestock goods [7]. The research mentioned in this article focuses on predicting LSD using machine and deep learning methodologies. Many academics have been working on LSD classi- fication during the last decade in order to develop very efficient machine learning and deep learning models. Because of the increasing accuracy and efficiency of prediction, the use of artificial intelligence and machine learning algorithms has gained promi- nence in recent years. The importance of research in this sector is in the capacity to construct and choose models with optimum accuracy and efficiency while minimising false negatives. The paper utilised image datasets of cattle infected and disinfected are two major classifications. The dataset is pre-processed and the later model is pre-trained using MobileNetv2,
convolution neural network (CNN) models, utilised for feature extraction and classification respectively. Hyperparameter tuning is per- formed for classifiers using a Keras tuner, and the performance of the proposed model is compared with the individual models and the existing work in the same domain over confusion metric parameters; Accuracy, Precision, F1-score, and Recall. The graphical comparison is also shown in the form of ROC and convergence curves. The overall architecture of the lumpy skin disease detection system adapted in this research is explained in Figure 1. Key Contributions • The dataset is pre-processed in order to remove unwanted noise and extract distorted images from the lumpy image dataset. • The LSD dataset is imbalanced, in order to handle imbalanced data synthetic minority oversampling technique (SMOTE) is utilised. • After successful preprocessing, the features are extracted from the images using MobileNetv2, which itself is a neural network-based classification algorithm for image datasets. • The hyper-parameter tuning is performed for classifiers using a Keras tuner. • The classification of the images as infected and disinfected is performed using a two-dimensional convolution neural network model. • The performance of the models is evaluated using confusion metric parameters, ROC and convergence curves.
Methodology
The author demonstrated that the ANN performed better than other machine learning approaches in predicting the presence of Lumpy Skin Infection in unseen data. The model achieved an Area Under Curve (AUC) score of
0.97. and an F1 performance measure score of 0.94.
The authors conducted a qualitative risk assessment to evaluate the likelihood of the LSD virus entering Ukraine and its subsequent transmission in the next year.
A CNN-based method was used for picture analysis to classify symptoms into their respective categories. The system achieved 95% accuracy in symptom categorization.
Debre Markos University, Addis Ababa University-School of Veterinary Medicine, the World Wide Web, and other secondary sources were used for photographic holdings.
The authors determined that the Random Forest method had the highest accuracy 97.7% among the tested algorithms.
Lumpy skin disease dataset with 18,603 instances and 16 features
A pre-trained CNN was employed to extract features from segmented regions of impacted skin color. The ELM classifier was used for classification, achieving an accuracy of 0.9012%.
Including SMOTE and GA to improved performance measure, such as Recall, F1 score, and AUC. The Recall and F1 score improved 0.99, and AUC improved 0.98.
CNN using VGG16, VGG-19, and Inception-v3 for feature extraction
The developed framework successfully predicted both lumpy skin and normal skin using deep CNN, achieving an accuracy of 92.5%. The performance was compared with other advanced methodologies such as kNN, SVM, NB, ANN, and LR.
The suggested method achieved high accuracies in classifying atopic dermatitis (88%), acne vulgaris (85%), and scabies (84.7%) in experimental settings.
Sample data (pictures) from four health care facilities in the Sunyani
Classifier performance was evaluated using ML methods such as ANN, LDA, Naive Bayes, and SVM for various feature combinations. LDA and SVM performed best in a test of four different classification methods.
Patient’s images from the Department of Dermatology, Yenepoya Medical College, Mangaluru
with the comparison of existing surveys and challenges. Section 2 describes the methodologies used in the research. The proposed work is defined in Section 3, along with dataset description, data preprocessing, balancing techniques, and hyper- parameter tuning. Section 4 describes the performance metrics used for the evaluation of the proposed models. The evaluated results are presented and discussed in Section 5. Section 6 concludes the research findings and provides the future scope of further research in this field.
Preliminary
This section includes a description of the methodologies used in this research. It includes the description of data balancing techniques, feature extraction techniques to
extract the features from the image dataset and classification techniques for the identification of infected and disinfected cattle. Synthetic Minority Oversampling Technique (SMOTE) To overcome the issue of class imbalance, the classification problem of lumpy disease detection strongly depends on data balancing. A class imbalance might manifest itself, for instance, in a considerably lower proportion of lumpy image cases compared to non-lumpy image cases or pneumonia patients. To categorise lumpy diseases, we have acquired a lumpy dataset, which brings some degree of irregularity into the dataset. In the proposed system, we have used the SMOTE Algorithm [17, 18] to balance the data. To create a separate line between the samples in the feature space and create a unique example along the
Sigma J Eng Nat Sci, Vol. 44, No. 2, pp. 856−868, April, 2026
line, the SMOTE algorithm first chooses examples that are close to the feature space. First, a sample from the underrepresented group is randomly chosen. The nearest k neighbours to the sample are then found. K is often equal to 5. At some point in the feature space between the two samples, a neighbour is chosen at random, and a synthetic sample is created. The practice is repeated until the proportion of the minority group equals that of the majority group.
sizes. However, the utilization of many branches may cause its computational cost to be greater than that of other structures.
Feature Extraction Models The paper utilises feature extraction techniques to extract the features of the image dataset. This section discusses various feature extraction techniques and their advantages [19, 20]. DenseNet-121 A popular deep-learning architecture for extracting features from images is DenseNet-121 [21]. In 2017, scientists from Facebook’s AI Research Lab unveiled the concept of tightly linked convolutional layers [22]. Each layer in a DenseNet-121 architecture is fed feature maps from the layers above it and contributes its feature maps to the layers below it. This highly interconnected structure reduces the impact of disappearing gradients and promotes feature reuse [23]. DenseNet-121’s dense connections may also be seen as a kind of feature aggregation, enabling the network to pick up more evocative elements. DenseNet-121, like ResNet-121, is a pre-trained model useful for extracting features from images. DenseNet-121’s intermediate features may be utilized as input to a new classifier or for other downstream tasks like object recognition or picture segmentation by eliminating the final classification layer. DenseNet-121 is often more memory-efficient and has fewer parameters during training and inference than ResNet-121. However, owing to its high connection structure, it may be costlier to compute. InceptionV3 Google’s InceptionV3 [24] deep learning architecture was released in 2015. It is a kind of convolutional neural network (CNN) that is often used for feature extraction and image categorization. To provide a more expressive feature representation, InceptionV3 is notable for its novel design, which consists of numerous branches of convolutional layers with varying kernel sizes. This architecture was created with the idea that many pictures identification jobs benefit from capturing both fine-grained and coarse-grained characteristics at distinct sizes. Simply deleting the model’s final classification layer will allow you to use the intermediate characteristics that InceptionV3 collected for downstream tasks like object identification or image segmentation. There are several ways in which InceptionV3 excels above other deep learning models. Parameter efficiency enables it to attain excellent accuracy with fewer parameters than other designs, while the use of numerous branches of convolutional layers allows it to collect information at various
ResNet50 ResNet50 [25] is a deep learning architecture with similarities to ResNet-121 but with additional layers and enhanced feature extraction capabilities. Microsoft Research first introduced it in 2016, and since then, it has seen extensive use in the fields of feature extraction and image categorization. ResNet50 is a 50-layer deep convolutional neural network (CNN). It employs residual blocks, much like ResNet-121, to facilitate the learning of very deep neural networks. ResNet50’s shortcut connections simplify the process of training deep neural networks and reduce the occurrence of vanishing gradients. ResNet50, like the other stated deep learning architectures, may be used as a pre-trained model for image feature extraction. ResNet50’s intermediate features may be used as input to a new classifier or for other downstream tasks like object identification or picture segmentation by deleting the model’s final classification layer. ResNet50 is more capable of feature extraction since it has more layers than ResNet-121. This has the potential to improve performance on increasingly difficult picture identification tasks, but it may need more computing power and training data. VGG19 In 2014, researchers at Oxford University proposed a deep learning architecture called VGG19 [26]. It’s a common tool for image categorization and feature extraction since it’s a convolutional neural network (CNN). VGG19’s intermediate features may be used as input to a new classifier or for other downstream tasks like object recognition or picture segmentation by deleting the model’s final classification layer. There are a number of ways in which VGG19 excels over competing deep learning systems [27]. It uses modest filters to extract characteristics at various scales and has a unified design that makes it easy to learn and execute. The high number of model parameters, however, may make VGG19 costlier to compute than other designs. MobileNetV2 MobileNetV2 [28] is a deep learning architecture that was introduced by Google in 2018. It is a convolutional neural network (CNN) that is designed for mobile and embedded devices, but it can also be used for image classification and feature extraction tasks on larger machines. MobileNetV2 has several advantages over other deep learning architectures. Its lightweight architecture makes it ideal for resource-constrained environments, and its use of depth-wise separable convolutions allows it to achieve high accuracy with fewer parameters and faster inference times. However, MobileNetV2 may not be as accurate as larger architectures such as ResNet50 or VGG19 for more complex image recognition tasks.
Sigma J Eng Nat Sci, Vol. 44, No. 2, pp. 856−868, April, 2026
Classification using CNN To classify the diseased and healthy cattle using images, the section discusses the detailed explanation of the layered deep learning classification models utilised in the experiment. Convolutional neural networks (CNNs) [29] [30] are a popular kind of deep neural network design for image categorization problem. Features are mapped to class probability in the fully connected layers at the network’s output node. Adding additional layers, more filters, or more sophisticated methods like residual connections or attention processes may all make a CNN more complicated. These enhancements allow the network to pick up more complicated information and perform better on difficult picture categorization tasks. A CNN typically consists of multiple layers [31], including: Convolutional layers: Input data, which is often an image, is processed by these layers using a series of filters. When the filters pass over the picture, they calculate dot products between their components and the images. Activation functions: These functions are used to incorporate non-linearity into the model and are applied element wise to the convolutional layer’s output. In convolutional neural networks (CNNs), the rectified linear unit (ReLU) function is the most used activation function. Pooling layers: To reduce the computational burden on the model and limit overfitting, these layers reduce the size of the feature maps.
Flatten layer: Flatten layer is typically used to transform the multi-dimensional feature maps into a one-dimensional vector. Fully connected layers: The final classification or regression operation is carried out by these layers, which compute dot products between the feature maps and a set of weights before feeding the result into an activation function to provide a prediction. Dropout layer: Overfitting may occur in deep neural networks, however regularisation methods like dropout can help. Figure 2 shows the pictorial representation of CNN architecture.
Proposed Methodology
The section describes the proposed methodology of the research paper, along with the description of the lumpy image dataset utilised. The section also narrates the flow of the proposed work in varying sub-sections. Dataset Description Secondary data was utilised for this research. The dataset, titled “Lumpy Skin Disease (LSD)”, had 24,803 records and was downloaded from the Mendeley database [8]. There are only 3,039 records in this dataset about livestock with LSD (12.25%), whereas there are 21,764 records in this dataset about cattle without LSD (87.75%). It indicates an unbalanced dataset. The LSD dataset’s properties are made up of geographical and
Sigma J Eng Nat Sci, Vol. 44, No. 2, pp. 856−868, April, 2026
meteorological features [8]. As regards the characteristics of meteorology, they include: Monthly Cloud Cover in Percent is represented by cld; The term “dtr” is for “diurnal temperature range”; “frs” stands for “frost day frequency”; “pet” stands for “potential evapotranspiration” in millimetres per day; and “pre” stands for “precipitation” in millimetres per month. The following acronyms stand for temperature: tmn, which stands for daily mean, tmp, which stands for temperature, tmx, which stands for monthly average maximum and minimum, vap, which stands for vapour pressure, and wet, which stands for wet day frequency in days. The geographical features are listed as follows: As with the other categories, ”region” refers to the continent, “country” to the nation where the outbreak was first reported, ”reporting date” to the day on which the outbreak was first reported, “elevation” to the height above sea level, and “dominant landcover” to the type of vegetation that predominated at the site in question. The X5 Ct 2010 Da file format provides a brief overview of asymmetric cattle GIS data, while the X5 Bf 2010 Da file format does the same for buffalo, and the lumpy file format classifies whether or not an animal is LSD-infected: 1, no code infection: 0.
Hyperparameter tuning using Keras Tuner The Keras Tuner library was developed as a hyperparameter tweaking tool for Keras- based deep neural networks. [33]: ü Define a model: Tuning a deep learning model requires first defining that model. A model constructed using either the sequential or functional Keras APIs will do. ü Define a hyperparameter search space: Establish a search space for the hyper- parameters you want to tune Next, you need to establish a search space for the hyperparameters you want to tune. ü Define a search algorithm: One need a search algorithm to sift through possible values for the hyperparameters and zero in on the optimal ones. ü Define a tuner: To encapsulate the model, the search space, and the search algorithm in order to execute the hyperparameter search. ü Run the hyperparameter search: Finally, one may start the tuner and look for opti- mal hyperparameter settings. Once the hyperparameter search has been completed, the best hyperparameters may be used to train a final model using the whole training dataset [34], and then the model’s performance can be assessed using the test dataset. Time and effort can be saved using Keras Tuner to automate the process of hyperparameter tweaking, which can lead to higher performance in your deep learning models [35].
Data Preprocessessing There are always no mistakes in the dataset. Text, numbers, time series, continuity, and discontinuity are only a few of the data kinds that may be found in a dataset. Data quality issues such as noise, abnormalities, missing, erroneous, and duplicates may also be present. The data set is either too big or too little. For the model to be effective, the data must agree with it. To make the dataset suitable for the model, pre-processing is required. Data pre-processing includes actions like the elimination of singular features and the handling of missing information. Categorical variables were transformed into numeric values using the one-hot encoding approach so that they could be utilised by ML algorithms. Additionally, min-max scaling was used to standardise the values of several predictive features. In the end, the train test split class in the scikit-learn package was used to divide the dataset into test and training sets. Models were validated using the original training dataset and a new, independent test set. Data Balancing SMOTE is used to handle imbalanced data classes. SMOTE constructs synthetic examples by selecting minority class examples at random and employing interpolation techniques to build examples between the target and surrounding instances. The SMOTE method classifies data before sampling it either more or less thoroughly. These three operational elements are the neighbor count used to enter new cases, the up-sampling rate, and the down-sampling rate, respectively. The dataset is now clean and efficient for classification.
Proposed Work This section explains our proposed approach in order to classify lumpy disease in cat- tle and buffaloes. The lumpy image dataset in our research is preprocessed for the sake of removing distorted and blurred images from the dataset. Then SMOTE is used to handle imbalanced data classes. SMOTE constructs synthetic examples by selecting minority class examples at random and employing interpolation techniques to build examples between the target and surrounding instances. The SMOTE method classifies data before sampling it either more or less thoroughly. These three operational elements are the neighbor count used to enter new cases, the up-sampling rate, and the down-sampling rate, respectively. The dataset is now clean and efficient for classification. Before classification, the various pre-trained models are utilized for extracting features from the images, these are pre-trained neural network architecture for large datasets such as ImageNet, where it learns to extract features that are useful for a wide variety of image classification tasks. The pre-trained models utilized are; MobileNetV2, VGG19, ResNet50, InceptionV3, and DenseNet-121. The proposed work has taken the pre-trained network and removed the final classification layers. This leaves us with a network that can extract features from images but does not classify them. Then we used these features as input to a new classification network.
Sigma J Eng Nat Sci, Vol. 44, No. 2, pp. 856−868, April, 2026
The classification is performed using a two-dimensional Convolution Neural Net- work with the combination of max pooling layers, flatten layers, and dense layers, arranged in a fashion as visible in Figure 2. The parameters of the CNN are tuned using a Keras tuner which automates the process of hyper-parameter tweaking, which can lead to higher performance in the deep-learning classification models. Finally, the dataset is trained with 70% of the lumpy image dataset and tested with the remaining 30% dataset. The experiment is performed using Python language, on jupyter notebook in Anaconda framework. The system with high processing with 32Gb RAM is used for the experiment purpose. In order to construct and train the CNN model, we must first load the required libraries, one of which is TensorFlow. We also import ImageDataGenerator, which is used to generate batches of augmented image data for training. Next, we set up the directories that will hold our test and training images. Next, we create an image generator that uses data augmentation to have our training data ready to go. We imported keras and its keras tuner library to tune the parameter of the classification CNN model. Figure 3 provides the overall flow of the proposed model. After successful execution (training and testing of model) of the models, the results are evaluated and analysed using confusion metric parameters; Accuracy, Precision, Recall, and F1-score. The graphical analysis is also performed using inbuilt Python libraries in order to generate ROC and Convergence curves, which are later discussed in the next section.
Performance Metrics
The research employs the confusion matrix parameters (accuracy, recall, precision, and F1 score) to assess the effectiveness of the proposed model. The following details the confusion matrix [34] terms: To understand the evaluation criteria, one should be familiar with the following four terms:
True Positive (TP): The number of occurrences in which the model correctly classifies them as positive. - True Negative (TN): The number of occurrences in which the model correctly classifies them as negative. - False Positive (FP): The number of negative occurrences in which the model incorrectly classifies them as positive. - False Negative (FN): The number of positive occurrences that the model incorrectly classifies as negative. Accuracy: It is the ratio of the number of correct predictions to the total num- ber of predictions made by the model. [36]. Accuracy = (TP + TN)/ (TP + TN + FP + FN
Precision: It is the ratio of true positives to the total number of positive predic- tions made by the model. [37]. Precision = TP/ (TP + FP)(2) Recall: It is the ratio of true positives to the total number of actual positive cases in the dataset. [16]. Recall = TP/(TP + FN)
F1-score: The F1-score balances accuracy and recall into a single number. It’s helpful when there’s an inequity in the dataset since it strikes a balance between accuracy and recall [38]. F 1 = 2 ∗ (precision ∗ recall)/(precision + recall)
Results And Discussion
In this work, MobileNetV2 was used with a CNN classifier as a feature extractor to diagnose LSD. There are a total of 14 data characteristics and 24,803 data points included in the forecasts. This dataset had 3,039 infected records from set (1) and 21,764 from set (0). Resampling of the dataset is required since it is clear that the current data does not fairly represent cattles who are infected with lumpy compared to those who are not. Before resampling the data, the attributes such as, ‘country’, ‘region’, ‘X5 Ct 2010 Da’, ’X5 Bf 2010
Sigma J Eng Nat Sci, Vol. 44, No. 2, pp. 856−868, April, 2026
Da’, and ‘reportingDate’, will be eliminated, as they are not used in the classification process. The association between attributes was also analysed using a correlation heatmap in an effort to reduce the size of strongly linked features. Some benefits of using MobileNetv2 for feature extraction are; first, since the network is pre-trained, we can take advantage of the knowledge learned from the original dataset to improve the accuracy of our new classification model. Second, by using depth-wise separable convolutions, MobileNetV2 is computationally efficient and requires fewer resources than other, more complex networks. The original dataset should be used to run the training models, and then the LSD dataset should be resampled to evaluate the first model’s efficacy. In order to put the model to the test, we must first partition the original data frame into a 70:30 split of training and validation data. After employing the SMOTE fit resample oversampling strategy, the resampled dataset shape counter (1: 21.764, 0: 21.764), indicates that both the infected and uninfected datasets share the same total number of observations (21.764). Figure 4 depicts the balanced distribution of data that results from SMOTE re-sampling. There are hyperparameters that need to be specified when using CNN for classification, including learning rate, batch size, filter size, pooling type, pool size, dropout rate,
activation function, optimizer, and others. The optimal hyperparameters are optimized using Keras Tuner, which examines all possible combinations within the provided setup. learning rate: [0.0001, 0.001, 0.01, 0.1], Batch size: [16, 32, 64, 128], filter size: [3x3, 5x5,7x7], activation function: [ReLU, LeakyReLU, ELU, Sigmoid, Tanh], pooling type: [max pooling, average pooling], pooling size: [2x2, 3x3], dropout rate: [0.1, 0.2, 0.3, 0.4, 0.5], optimizer: [adam, RMSprop, SGD] are used in this study. The most ideal settings based on CNN are determined after hyperparameter tweaking, and they are as follows: learning rate: 0.001, batch size: 32, filter size: 3x3, pooling type: max pooling, pool size: 3x3, dropout rate: 0.5, activation function: ReLU, optimizer: adam. The accuracy, precision, sensitivity, and F1-Score of the model are assessed using the Confusion Matrix. based on the test scenario utilizing MobileNetV2 and SMOTE with CNN classification on LSD datasets. The experiment’s findings are shown in Table 2. The graphical comparison of the evaluation metrics is illustrated in Figure 5. Receiver Operating Curve (ROC) The confusion metric is used to assess the receiver operating characteristic (ROC) curve and to identify whether or not the identification is successful. The ROC curve establishes the ratio of true positives to false positives in
Table 2. Performance of various feature extraction models with CNN Model
Vgg19-Cnn
Sigma J Eng Nat Sci, Vol. 44, No. 2, pp. 856−868, April, 2026
experimental data. True positive rate (Y axis) and false positive rate (X axis) are plotted on the ROC curve. An ideal ROC (receiver operating characteristic) curve climbs sharply toward the top-left corner, signifying that the model has a high true positive rate and a low false positive rate. The closer the curve gets to this corner, the more effectively the model differentiates between positive and negative classes.
The Area Under the Curve (AUC) is a crucial indicator of performance, with values near 1 indicating excellent classification ability. In contrast, a random or weak model will produce a curve closer to the diagonal, where the AUC approaches 0.5. For optimal performance, the ROC curve should have an AUC near 1, reflecting strong predictive accuracy. Figure 6 depicts how the proposed model ROC stacks up against its competitors. The area under the ROC curve (AUC) is a two-dimensional measure of this connection that shows how well the suggested model outperforms alternative models. Convergence Curve While training and testing a model, the optimal value of the learning parameter is depicted by a convergence curve that plots the accuracy of the model against the loss function. The Figure 7 depicts the convergence curves of the proposed and alternative models. The sub-figures represent 7a DenseNet121, 7b InceptionV3, 7c ResNet50, 7d VGG19, and 7e proposed approach (MobileNetV2), all with the CNN model respectively. The 200-epoch training and validation accuracy curves for the proposed models are displayed. The proposed model outperforms the state-ofthe-art with a training and validation accuracy of 99.88%, revealing a faster convergence rate.
Figure 6. ROC of Various Feature Extraction techniques with CNN.
Comparison with existing work Table 3 gives a state-of-art analysis based on the performance of existing classification models and the proposed models on considering accuracy measures. The paper taken into account are using different classification techniques some of them are artificial neural networks, convolution
Sigma J Eng Nat Sci, Vol. 44, No. 2, pp. 856−868, April, 2026
Figure 7. Convergence curve analysis: (a)DenseNet121, (b)InceptionV3 (c) ResNet50 (d)VGG19 (e) Proposed model.
Table 3. The state-of-art comparison of the proposed model with existing work Ref
Debre Markos University, Addis Ababa University-School CNN of Veterinary Medicine, the World Wide Web, and other secondary sources were used for photographic holdings.
Neural network model using Inception-v3 for feature extraction
Pictures from four health care facilities in the Sunyani Municipality of Ghana.
84.7. (scabies)
Sigma J Eng Nat Sci, Vol. 44, No. 2, pp. 856−868, April, 2026
neural networks, ensemble learning models and individual models such as random forest. The broader implications of our findings extend beyond the immediate scientific community, as the detection model can be implemented in various agricultural sectors, particularly in regions where LSD outbreaks are prevalent, such as Africa, the Middle East, and Asia. By incorporating this technology, veterinary services, livestock producers, and governmental agencies can utilize automated tools to rapidly detect and respond to disease outbreaks, improving animal welfare and reducing the impact of the disease on food security and trade. This work can also inspire further research into applying deep learning models for other transboundary animal diseases, enhancing overall disease control strategies.
Share and Cite
SINGH, R.; DANGI, R.; MALVIYA, L.; JADHAV, A. Deep learning-based approach for automated detection of lumpy skin disease. Sigma Journal of Engineering and Natural Sciences 2026, Vol. 44, pp. 856-868. https://doi.org/10.14744/sigma.2025.00171