An Ensemble-Based Approach for Accurate Classification of Astronomical Objects Using Deep

Learning
ABSTRACT
The rapid development in telescopic technologies and space missions in the field of astronomy has led
to an exponential increase in astronomical data. The vast amount of data is collected from satellites in
the form of audio, video and images. The classification of space images manually has become infeasible
and more time consuming. Hence the new ensemble model has been proposed which combines
ResNet101V2 and VGG19 architectures through soft voting technique to classify - asteroids, black
holes, comets, constellations and planets. ResNet101V2 is well known for its residual connections
which resolves the problem of vanishing gradient and enable efficient in training the deeper architecture
models. VGG-19 architecture uses convolutional and pooling layers for effective feature extraction and
its 19-layer depth ensures that model efficiently learn both low-level and high- level features. By
combining ResNet101V2 and VGG19 through soft voting the proposed ensemble model leverages the
depth and gradient optimization of ResNet101V2 with the structured feature extraction of VGG19 and
achieves an accuracy of 99.7% on training data and 95.7% on validation data. The outcome of this
ensemble model demonstrated its efficiency in multi-class image classification in the field of astronomy.
INDEX TERMS: VGG-19, ResNet101V2, Ensemble model, Soft Voting, Image Classification
1. INTRODUCTION
The field of astronomy shows an tremendous progress in recent years due to discovery of advanced
telescopes, space exploration missions that provide exceptional insights into celestial phenomena and
the eagerness to explore vastness of the universe .The recent improvements have led to the collection
of high resolution images, audio and videos data from satellites and observatories revolutionizing the
way we study universe.
The notable advancements in the field of astronomy include the development of high-resolution
telescopes such as James Webb Space Telescope which provides clear images of distant galaxies,
nebula, black holes, stars, and other cosmic objects. In addition to that space missions like NASA's
Voyager and ESA's Gaia have provided knowledge about the Milky Way's structure and planetary
system. The data collected from such satellites and observatories contains valuable information about
universe but using traditional classification methods or manual analysis of large volume of data is time
consuming and infeasible.
Hence, to classify the images from these resources an ensemble image classification model has been
proposed which is an machine learning technique where multiple models are combined to improve
predictive performance. By combining the strength of different models, ensemble method reduce the
errors and increases robustness.
The previous works in ensemble model of astronomical image classification focus on specific object
types or datasets such as galaxies or star clusters. They lack a unified framework to classify multiple
object types which limits their applicability.
In this study, an ensemble approach is proposed that combines the ResNet101V2 and VGG19
architectures using the soft voting technique. Initially, the algorithms VGG19, ResNet101V2,
InceptionV3, EfficientNetV2M, and Xception were chosen for evaluation. Among these, ResNet101V2
demonstrated the highest accuracy, followed by VGG19. Consequently, an ensemble model was
developed combining VGG19 and ResNet101V2, achieving an accuracy of 95.7%.
This proposed ensemble model accurately classifies five classes of astronomical images: asteroids,
black holes, comets, constellations, and planets. The automated image classification model offers an
efficient way to process high-resolution images from the universe quickly and accurately. By leveraging
the depth and gradient optimization capabilities of ResNet101V2 alongside the structured feature
extraction of VGG19, the model achieves high accuracy in multi-class classification tasks. Together,
these architectures complement each other, making the ensemble model well-suited for classifying
complex astronomical images.
1.1 RELATED WORKS:
Over the years, various studies have been proposed for classifying and detecting astronomical objects
which leads to tremendous advancements in machine learning and deep learning.
Star-Galaxy Classification have been proposed as ensemble approach which combines multiple
classification techniques such as morphological classifiers, supervised learning (e.g., random forests),
unsupervised learning (e.g., self-organizing maps), and Bayesian template-fitting. This is an meta-
classification framework which combines the strengths of individual classifiers which outperforms
standalone methods. By using CFHTLenS survey data, this frameworks exhibit high performance even
with low-quality of training sets , proves its value for ongoing and future photometric surveys like the
Dark Energy Survey and LSST(Canada-France-Hawaii Telescope Lensing Survey).
The method of transfer learning and fine tuning has been used in asteroid classification. The deep
learning algorithms has been used to classify Near-Earth-Asteroids from other astronomical objects
which can be used to mitigate potential threats to Earth. The deep learning neural networks such as
InceptionV3, Xception, InceptionResNetV2 and ResNet152V2 has been evaluated. InceptionV3
achieves high accuracy and minimized loss in asteroid classification.
The detection of comets in multispectral astronomical surveys has been achieved with Faster R-CNN a
deep learning-based framework. The method accommodates images across multiple wavelengths,
achieves an average precision of 80% on NEOWISE 4-band images. This study shows that multispectral
data enhances precision by 170% compared to single-wavelength models.
The recent research uses Hubble Space Telescope data introduced HR-CelestialNet a deep learning
model designed for high resolution galaxy and NSC classification. To overcome the struggle of real
time classification of celestial images for Chinese Space Station Telescope (CSST) survey, this method
has been introduced. The deep learning models such as HR-CelestialNet, AlexNet, VGGNet, and
ResNet have been evaluated. In which HR-CelestialNet shows high accuracy of 89.09% on testing data
and outperformed other models. It shows high performance even if the quality of the image is low and
blur.
The research of diagnosing Alzheimer’s disease (AD) shows greater accuracy by using advanced deep
learning and ensemble techniques. It combines models like ResNet101V2, VGG19, and a custom CNN,
applying weighted soft voting to achieve an impressive classification accuracy of 95.125% which uses
same algorithms as the proposed architecture of this paper. The study highlights the effectiveness of
cutting-edge architectures and ensemble methods in advancing AD diagnosis and prognosis through
medical imaging.
These studies together highlights the growing role of machine learning and deep learning in
astronomical object classification and detection which paves the way for efficient and high quality
solutions in upcoming large-scale surveys and missions.
1.2 RESEARCH GAPS:
Despite significant advancements in astronomical data analysis, several challenges persist that limit the
applicability of existing methodologies. Current approaches often focus on specific celestial objects,
such as galaxies or star clusters, without providing a unified framework capable of handling diverse
classes like asteroids, black holes, comets, constellations, and planets. The another limitation lies in the
reliance on individual models, which, while effective in isolation, fail to leverage the complementary
strengths of ensemble techniques. Furthermore, much of the existing research emphasizes binary
classification tasks, leaving a substantial gap in addressing multi-class classification scenarios that are
critical for analysing the vast and diverse datasets generated by modern telescopes and space missions.
Addressing these gaps requires a robust and efficient model capable of multi-class classification while
combining the strengths of different architectures to enhance accuracy and generalization.
1.3 RESEARCH OBJECTIVES
Build an Ensemble Model to Classify Astronomical Objects

 From a technical perspective, this study aim to develop an ensemble-based deep learning
model which includes ResNet101V2 and VGG19 architectures capable of performing multi-
class classifications on astronomical objects.
Combine the Power of ResNet101V2 with VGG19

 To combine the effective depth and gradient optimization capabilities of ResNet101V2 and the
structured feature extraction potential of VGG19 for efficient and accurate classification.
Handle Multi-Class Astronomical Data
 This study aims to address the research gap of binary classification dominance by focusing on
multi-class classification of five distinct categories: asteroids, black holes, comets,
constellations, and planets.
[Link] METHODOLOGY:
In the proposed methodology two models are implemented - VGG19 and ResNet101V2. The models
VGG-19 and ResNet101V2 has been chosen because it outperformed among chosen algorithms –
InceptionV3,EfficientV2M,Xception,VGG-19 and ResNet101V2.
Table1. Displays the accuracy of the evaluated model

Algorithm Accuracy
ResNet101V2 94.4%
VGG -19 91.1%
InceptionV3 90.4%
Xception 82.76%
EfficientNetV2 46.6%

2.1 DATA COLLECTION AND PREPROCESSING :

The dataset has been collected from Kaggle . The Spacenet image dataset has been utilized here. The
pre-processing of image has been done by resizing them to the required input dimensions (224*224)
and pixel values has been normalized.
 Table 2. The below table shows the distribution of images in a dataset.

Since the distribution of images across the classes is not balanced , class weights has been used to
manage the imbalance distribution of dataset.
 Figure [Link] image of asteroid, black hole, comet, constellation and planet from the Spacenet
dataset.

Training Dataset : From the dataset given 80% of the dataset are used for training the models.
Validation Dataset : From the dataset given 20% of the dataset are used for validation purpose.

Dataset balancing technique:

The class weights is a technique to handle imbalanced datasets by assigning numerical weights to each
class. These weights adjust the importance of each class in the loss function during model training.
This method ensures that minority classes contribute equally to model optimization which adjust the
loss function that allow the machine learning models to focus on minority classes without altering the
dataset structure.
The aim of the class weights is to modify the loss funciton. For a classification task, the weighted loss
can be expressed as :

Wc=N/(C*nc)
Where , Wc is the weight for class c
N is the total number of samples in the dataset
C is the total number of classes
nc is the number of samples in class c
Analysis before and after applying class weights :
The below table shows the impact of class weights in the ensemble model
Table 3. Analysis of class weights
Class weights Accuracy
Before Applying class weights 94.1%
After applying class weights 95.7%
The values of hyperparameters used are given in the below table.
Table 4. Hyperparameters used
Parameter VGG19 ResNet101V2
Epoch 60 60
Batch size 32 32
Learning rate 1e-4 1e-5
Optimizer Adam Adam
Loss Function Sparse Categorical Crossentropy Sparse Categorical Crossentropy

2.2 FEATURE EXTRACTION USING VGG19 :

The VGG-19 has been used here because of its 19 layer depth balances both computational efficiency
and accuracy. It captures the essential features hence it is effective for datasets with limited
[Link] top layer has been removed and base layers has been frozen in this model.
The input image with required dimension (224*224) has been passed through the model. The pre-trained
VGG-19 model has been loaded . The custom CNN layers has been added to the model to extract the
features efficiently. The flatten layer has been added to convert the output from the VGG19 model into
1-D vector. The dense layer has been added with activation function as ReLU which helps the model to
capture complex patterns. The dropout rate here is 0.5 which indicates 50% of the neurons has been
dropped which allow the model to learn new patterns. The dense layer is added with the activation
functions as softmax which indicates the maximum class probablity from the class probablity of five
classes. The optimizer used here is Adam which shows high performance through the training process.
 Figure 2. Architecture of VGG19
Input Layer : The size of the input image is 224*224*3.
Convolutional Layers (in pre-trained model) : The block dimensions are standard for VGG-19
architectures is given below ,

 Block 1 : The output image from this convolutional block has an image size of 112*112*64
 Block 2 : The output image from this convolutional block has an image size of 56*56*128
 Block 3 : The output image from this convolutional block has an image size of 28*28*256
 Block 4 : The output image from this convolutional block has an image size of 14*14*512
 Block 5 : The output image from this convolutional block has an image size of 7*7*512
Flatten Layer: This layer reshapes the multi-dimensional output from the VGG19 feature extractor into
a one-dimensional vector. This layer connects the convolutional layers to the dense (fully connected)
layers. The output size of the image from this layer is 7*7*512.
Dense (1024 units, ReLU):A fully connected layer with 1024 neurons and ReLU (Rectified Linear
Unit) activation. This layer helps in learning complex and non-linear relationships from the flattened
feature vector.
Dropout (rate = 0.5):Randomly deactivates 50% of the neurons during training to prevent overfitting
and improve the model's generalization ability.
Dense (num_classes units, Softmax):A fully connected layer with a number of neurons equal to the
number of classes in the classification task. The softmax activation function outputs a probability
distribution across the classes, with the highest probability indicating the predicted class. This dense
layer represents the final output of the model, providing the probabilities for each class. The output size
of the vector is 1*5 .The class with the highest probability is selected as the model's prediction.
The confusion matrix in Figure 3 visualizes the predictions of the model which shows the distribution
of correct and incorrect classification across all classes.

Figure 3. Confusion Matrix of VGG19

The confusion matrix displays the performance of a classification model across five classes: asteroid,
black hole, comet, constellation, and planet. Each cell shows the number of predictions where the true
label matches or mismatches the predicted label. The diagonal entries (e.g., 45 for "asteroid" and 277
for "planet") represent correct classifications, while off-diagonal entries indicate misclassifications.
The classification report in Figure 4. provides precision, recall and F1-score for each class.
 Figure 4. Classification report of VGG19
The classification report summarizes the model's performance for each class with metrics like
precision, recall, and F1-score, along with support (the number of true samples for each class). The
model achieves a weighted average F1-score of 0.91 and an accuracy of 91% across 876 samples.

2.3 FEATURE EXTRACTION USING RESNET101V2 :

The ResNet101V2 has been used here as it uses residual connections or skip connections to address
vanishing gradient problem which enables effective training of deeper architectures. This is particularly
important for extracting intricate patterns and features in high-resolution astronomical images.
The input image, resized to the required dimensions (224×224), is passed through the pre-trained
ResNet101V2 model. The layers up to layer 160 are frozen to retain learned features, while the
remaining layers are fine-tuned. A Global Average Pooling (GAP) layer is added to aggregate spatial
features globally, followed by a dense layer with ReLU activation to capture complex patterns. To
prevent overfitting, a dropout layer with a rate of 0.5 is included, randomly dropping 50% of the
connections during training. Finally, a dense layer with a softmax activation function outputs the class
probabilities, identifying the most probable class among the five categories. The Adam optimizer is
used to ensure high performance during the training process.
 Figure 5. Architecture of ResNet101V2
Input : The size of the input image is 224*224*3
Convolutional layer (in Pre-trained model) : The block dimensions are standard for ResNet101V2
architectures is given below ,

 Block 1 : The output image from this convolutional block has an image size of 56*56*256.
 Block 2 : The output image from this convolutional block has an image size of 28*28*512.
 Block 3 : The output image from this convolutional block has an image size of 14*14*1024.
 Block 4 : The output image from this convolutional block has an image size of 7*7*2048.
Global Average Pooling (GAP): This operation reduces the spatial dimensions of the feature maps by
computing the average of each feature map. This operation outputs a single value per feature map that
makes the model more robust to spatial translations. The output size from this layer would be 1*1*2048.
Dropout (rate = 0.5): This dropout rate randomly deactivates 50% of the neurons during training to
prevent overfitting and improve generalization.
Dense Layer : This is a fully connected layer with 128 neurons and ReLU activation to learn complex
patterns and non-linear relationships from the extracted features.
Dropout (rate = 0.5): Again, the model randomly drops 50% of the connections during training to
enhance the robustness of the model and prevent overfitting.
Dense Layer : The softmax activation normalizes the outputs into a probability distribution, ensuring
the highest probability corresponds to the predicted class. This final dense layer represents the final
output of the model, providing the probabilities for each class. The class with the highest probability is
selected as the model's prediction. The size of the output would be 1*5.

Figure 6. Confusion Matrix of ResNet101V2

This confusion matrix shows the performance of a ResNet101V2 model on five classes: asteroid,
black hole, comet, constellation, and planet. Correct predictions are along the diagonal (e.g., 58 for
"asteroid" and 286 for "constellation"), while off-diagonal values indicate misclassifications. The
model performs well overall, with minor errors in neighboring classes like "comet" and
"constellation."

Figure 7. Classification report of ResNet101V2

The classification report summarizes the performance of a multi-class model, showing precision,
recall, and F1-score for each class (e.g., asteroid, black hole). Overall, the model achieved high
accuracy (0.94) with a weighted F1-score of 0.94 across 876 samples.
2.4 ENSEMBLE MODEL ARCHITECTURE
The input images are processed through both VGG19 and ResNet101V2 models to obtain probability
of each class from both models. These models produces arrays of probabilities for the five classes,
where each probability represents the probability of the image that belonging to a specific class. The
softmax activation is used to compute these probabilities in each model, which ensures that the outputs
are normalized and sum to 1.

Figure 8. Architecture of Ensemble model

An ensemble technique known as soft voting is applied to improve classification accuracy. Soft voting
averages the class probabilities generated by VGG19 and ResNet101V2, and the class with the highest
averaged probability is selected for the final prediction. This technique combines the strengths of both
models which balance their biases and enabling a more robust decision-making process.
The performance of the ensemble model has been evaluated using metrics such as validation accuracy,
validation loss, confusion matrix, precision, recall, and F1-scores. The probabilities of each model has
been averaged to perform soft voting. The ensemble approach reduces overfitting, handles extreme
predictions, and lowers variance in classification decisions. This enhances the model's generalization
capability to perform well on unseen data.

Figure 9. Confusion Matrix of the Ensemble model

The confusion matrix of the ensemble model displays a more balanced distribution of correct
classifications across all classes. The confusion matrix illustrates the classification performance,
showing correct predictions along the diagonal and misclassifications off-diagonal. Most predictions
are accurate, with minor misclassifications such as a few comets misclassified as constellations or
planets. The precision, recall, and F1-scores values are increased compared to individual models and
fewer false positives and false negatives indicate the ensemble model's ability . The misclassifications
are significantly reduced in this ensemble approach.
By combining the strengths of ResNet101V2 and VGG19, the ensemble approach achieves higher
performance across all evaluation metrics. The ability of the ensemble models to generalize effectively
and minimized misclassifications makes it a preferred choice for multi-class classification tasks. This
ensemble method not only improves accuracy but also ensures a more reliable and robust performance
for real-world applications.
The ensemble model achieves higher precision, recall and F1-scores indicating its ability to make fewer
false positives and negatives while maintaining balance across all classes. The misclassifications are
reduced in the ensemble approach, particularly for classes that were challenging for individual models.

Figure 10. Classification report of the Ensemble Model

This classification report highlights improved performance with a 95% overall accuracy and a
weighted F1-score of 0.95. Each class shows strong precision, recall, and F1-scores, particularly for
"planet" and "constellation."
The ensemble approach effectively combines the strengths of ResNet101V2 and VGG19, leading to
improved performance across all evaluation metrics. Its ability to generalize better, reduce
misclassifications, and stabilize predictions makes it the preferred choice over individual models for
this multi-class classification task.
3. EVALUATION :
The ensemble model of VGG19 and ResNet101V2 achieved 95.7% accuracy on validation data with
a validation loss of 0.1911 indicate strong generalization on unseen data. The confusion matrix shows
higher performance by correctly classifying all classes. The precision, recall, F1-score shows the
model's ability to handle the false positives and false negatives effectively while the model balance
predictions for all classes.
 Table 5. Comparison of Model’s performance

Model Accuracy (in Loss Precision Recall F1- score

%)
VGG19 91.1 0.3199 0.91 0.91 0.91
ResNet101V2 94.4 0.3013 0.94 0.94 0.94
Xception 82.76 0.5144 0.85 0.83 0.82
InceptionV3 90.4 0.5071 0.91 0.90 0.91
EfficientNetV2M 46.4 1.5604 0.51 0.46 0.46
Ensemble Model 95.7 0.1911 0.95 0.95 0.95

The metrics mentioned in the above table can be calculated using the formulas given below.
Formula for calculating the metrics :
Accuracy : Accuracy measures the proportion of correctly classified instances out of the total instances.

Precision : Precision measures the proportion of correctly predicted positive instances out of all
instances predicted as positive.

The classification report gives the precision of each class so the precision of the model (Weighted
average precision) can be calculated using below formula :

Recall : Recall measures the proportion of correctly predicted positive instances out of all actual
positive instances.

The classification report gives the recall of each class so the recall of the model (Weighted recall) can
be calculated using below formula :

F1-score : The F1-score is the harmonic mean of precision and recall. It provides a balanced measure
when there is an uneven class distribution.
The classification report gives the F1-score of each class so the overall F1-score of the model (Weighted
F1-score) can be calculated using below formula :

The ensemble model consistently outperforms both individual models (VGG19 and ResNet101V2)
across all evaluation metrics which indicates the strength of combined architecture.
Training and Validation Graphs :
The training and validation accuracy graphs and loss graphs ensures that ensemble model avoids
overfitting which makes it perfect model for real-world applications. The proposed ensemble model
shows superior performance than individual models which paves the way for its application in large-
scale astronomical surveys.

Figure 11. Training and Validation accuracy and loss graphs of VGG19

Figure 12. Training and Validation Accurracy and loss graphs of ResNet101V2

4. RESULTS AND DISCUSSION :

The proposed ensemble model, which combines the strengths of ResNet101V2 and VGG19
architectures through a soft voting mechanism, shows high performance in classifying astronomical
objects into five categories: asteroids, black holes, comets, constellations, and planets. The ensemble
model achieved a validation accuracy of 95.7% which outperforms other chosen algorithms. The
ensemble model exhibited a lower validation loss of 0.1911which indicates model’s generalization
ability to unseen data and robustness. The confusion matrix analysis indicate that the ensemble model
significantly reduced misclassifications. The metrics like precision, recall, and F1-score further
confirmed the ensemble model's reliability, showing higher values across all classes, ensuring both high
confidence and sensitivity in predictions.

The training and validation accuracy graphs indicated effective convergence, with dropout layers (rate
of 0.5) helping to mitigate overfitting. The ResNet101V2 architecture contributed its ability to handle
complex astronomical patterns through residual connections, while VGG19 excelled in extracting
structured features, particularly in datasets with less variability. The ensemble model utilize these
strengths and achieves balanced predictions by averaging class probabilities, thereby reducing bias and
variance. In addition the ensemble approach provided more consistent and robust classifications,
making it suitable for practical applications in astronomical surveys and research. Although a few
misclassifications were observed, their frequency was significantly lower compared to individual
models, underscoring the ensemble's overall effectiveness in real-world scenarios.

5. LIMITATIONS :
Class Imbalance:
Despite employing class weighting to mitigate dataset imbalance, the model's performance for minority
classes may still be suboptimal due to limited representation and diversity. This can lead to disparities
in precision, recall, and generalization for underrepresented categories. Hence, residual effects of class
imbalance can still be considered a limitation

Sensitivity to Weighting Choices:

The effectiveness of class weighting depends on the chosen weighting strategy. The improper weighting
could lead to overcompensation, where the model overemphasizes minority classes, potentially
misclassifying majority class samples.
6. FUTURE DIRECTIONS :

 The future studies can integrate explainable AI techniques to provide insights into how the
model makes classification decisions.
 The misclassification can be addressed by incorporating advanced error analysis methods
which can help refine the model architecture or preprocessing steps to mitigate these issues.
 The model can be integrated into end-to-end pipelines for astronomical surveys which
automates tasks such as anomaly detection, object tracking, and classification in a unified
framework.
7. CONCLUSION :
This study presents an ensemble-based approach combining ResNet101V2 and VGG19 architectures
through a soft voting mechanism to classify astronomical objects into five categories: asteroids, black
holes, comets, constellations, and planets. The ensemble model achieves superior performance
compared to individual models, with a validation accuracy of 95.7% and reduced misclassifications. By
leveraging the gradient optimization of ResNet101V2 and the structured feature extraction of VGG19,
the proposed approach ensures robustness and generalization to unseen data. These results highlight the
potential of ensemble deep learning models for large-scale astronomical surveys, paving the way for
more efficient and accurate classification of complex celestial objects.
REFERENCES :
[1] Edward J. Kim, Robert J. Brunner, Matias Carrasco Kind, A hybrid ensemble learning approach to
star–galaxy classification, Monthly Notices of the Royal Astronomical Society, Volume 453, Issue 1, 11
October 2015, Pages 507–521, [Link]
[2] S. Manthira Moorthi, I. Misra, R. Kaur, N. P. Darji and R. Ramakrishnan, "Kernel based learning
approach for satellite image classification using support vector machine," 2011 IEEE Recent Advances
in Intelligent Computational Systems, Trivandrum, India, 2011, pp. 107-110, doi:
10.1109/RAICS.2011.6069282.

[3] S. A. Şakaci and S. Ertürk, "Superpixel based spectral classification of hyperspectral images in
different spaces," 2018 5th International Conference on Electrical and Electronic Engineering (ICEEE),
Istanbul, Turkey, 2018, pp. 384-388, doi: 10.1109/ICEEE2.2018.8391367.
[4] B. B. Damodaran, R. R. Nidamanuri and Y. Tarabalka, "Dynamic Ensemble Selection Approach for
Hyperspectral Image Classification With Joint Spectral and Spatial Information," in IEEE Journal of
Selected Topics in Applied Earth Observations and Remote Sensing, vol. 8, no. 6, pp. 2405-2417, June
2015, doi: 10.1109/JSTARS.2015.2407493.
[5] N. Aldahoul, H. A. Karim, M. A. Momo, F. I. F. Escobara and M. J. T. Tan, "Space Object
Recognition With Stacking of CoAtNets Using Fusion of RGB and Depth Images," in IEEE Access,
vol. 11, pp. 5089-5109, 2023, doi: 10.1109/ACCESS.2023.3235965.
[6] Y. Zhang, G. Cao and X. Li, "Multiview-Based Random Rotation Ensemble Pruning for
Hyperspectral Image Classification," in IEEE Transactions on Instrumentation and Measurement, vol.
70, pp. 1-14, 2021, Art no. 5500314, doi: 10.1109/TIM.2020.3011777.
[7] S. Sudharson, R. Annamalai, A. A. Reddy and P. Varsha, "Stacking Ensemble Model for Celestial
Object Classification: Galaxies, Stars and Quasars," 2023 Seventh International Conference on Image
Information Processing (ICIIP), Solan, India, 2023, pp. 746-752, doi:
10.1109/ICIIP61524.2023.10537787.
[8] B. S. Panigrahi, S. Artheeswari, W. Khan, P. G, D. Sushil Kumar Pathak and R. Bharanidharan,
"Artificial Intelligence in Astrophysics: Automated Detection of Celestial Objects and Anomaly
Detection," 2024 5th International Conference on Recent Trends in Computer Science and Technology
(ICRTCST), Jamshedpur, India, 2024, pp. 210-214, doi: 10.1109/ICRTCST61793.2024.10578402
[9] Roy Krishty, Barsha & Faruk, Md & Islam, Md Nazmul & Srizon, Azmain & Hasan, S M & Mamun,
Md. Al & Hossain, Md & Hossain, Md. (2024). A Cutting-Edge Ensemble of Vision Transformer and
ResNet101v2 Based Transfer Learning for the Precise Classification of Leukemia Sub-types from
Peripheral Blood Smear Images. 49-54. 10.1109/ICEEICT62016.2024.10534388.

[10] S. Chatterjee, J. -M. Gil and Y. -C. Byun, "Early Detection of Multiclass Skin Lesions Using Transfer
Learning-Based IncepX-Ensemble Model," in IEEE Access, vol. 12, pp. 113677-113693, 2024, doi:
10.1109/ACCESS.2024.3432904.
[11] T. Adar, E. K. Delice and O. Delice, "Detection of COVID-19 From A New Dataset Using
MobileNetV2 and ResNet101V2 Architectures," 2022 Medical Technologies Congress (TIPTEKNO),
Antalya, Turkey, 2022, pp. 1-4, doi: 10.1109/TIPTEKNO56568.2022.9960225.
[12] Fabricio Aparecido Breve,COVID-19 detection on Chest X-ray images: A comparison of CNN
architectures and ensembles,Expert Systems with Applications,Volume 204,2022,117549,ISSN 0957-
4174,[Link]
0957417422008673)
[13] Ali, S., Agrawal, J. Automated segmentation of brain tumour images using deep learning-based
model VGG19 and ResNet 101. Multimed Tools Appl 83, 33351–33370 (2024).
[Link]
[14] Catalina Gómez, Mauricio Neira, Marcela Hernández Hoyos, Pablo Arbeláez, Jaime E Forero-
Romero, Classifying image sequences of astronomical transients with deep neural networks, Monthly
Notices of the Royal Astronomical Society, Volume 499, Issue 3, December 2020, Pages 3130–
3138, [Link]
[15] N. Chihara, T. Takata, Y. Fujiwara, K. Noda, K. Toyoda, K. Higuchi, M. Onizuka,Effective
detection of variable celestial objects using machine learning-based periodic analysis,Astronomy and
Computing,Volume 45,2023,100765,ISSN 2213-1337,[Link]
([Link]
[16] Barman, A.U., Shah, K., Kashyap, K.L., Sandilya, A., Desai, N.P. (2021). Binary Classification of
Celestial Bodies Using Supervised Machine Learning Algorithms. In: Prateek, M., Singh, T.P.,
Choudhury, T., Pandey, H.M., Gia Nhu, N. (eds) Proceedings of International Conference on Machine
Intelligence and Data Science Applications. Algorithms for Intelligent Systems. Springer, Singapore.
[Link]
[17] C. Mu, J. Liu, Y. Liu and Y. Liu, "Hyperspectral Image Classification Based on Active Learning
and Spectral-Spatial Feature Fusion Using Spatial Coordinates," in IEEE Access, vol. 8, pp. 6768-6781,
2020, doi: 10.1109/ACCESS.2019.2963624.
[18] J. Wang, Y. Zheng, M. Wang, Q. Shen and J. Huang, "Object-Scale Adaptive Convolutional Neural
Networks for High-Spatial Resolution Remote Sensing Image Classification," in IEEE Journal of
Selected Topics in Applied Earth Observations and Remote Sensing, vol. 14, pp. 283-299, 2021, doi:
10.1109/JSTARS.2020.3041859.
[19] P. Li, Y. Pei, J. Li and H. Xie, "Medical Image Recognition Using a Novel Neural Network
Construction Controlled by a Kernel Method Encoder," in IEEE Access, vol. 12, pp. 146807-146817,
2024, doi: 10.1109/ACCESS.2024.3474742.
[20] J. J. Bird and A. Lotfi, "CIFAKE: Image Classification and Explainable Identification of AI-
Generated Synthetic Images," in IEEE Access, vol. 12, pp. 15642-15650, 2024, doi:
10.1109/ACCESS.2024.3356122.