1

Uncertainty-focused on Classification of Astronomical Events using

EfficientNetB0

SARKESAN S

Anna University Regional Campus, Coimbatore. Email id : sarkesansarkesan@[Link]

ROHITH A

Anna University Regional Campus, Coimbatore. Email id : a.rohit11th@[Link]

ABSTRACT

UncertaintyNet is a neural network-based system developed to classify astronomical

objects accurately while also assessing the certainty level of its outputs. This capability is vital in

modern astrophysics, where precise identification of transient phenomena can significantly influence

scientific conclusions. It effectively classifies a wide range of astronomical events occurring in space

(Ex: supernovas, gamma-ray bursts, neutron stars, and black hole),producing both the predicted label

and a confidence score. Leveraging transfer learning with the EfficientNetB0 architecture,

UncertaintyNet can learn from limited and imbalanced datasets. To address class imbalance, it

employs focal loss, which emphasizes harder-to- classify examples. A key feature is its confidence

thresholding mechanism, which labels low-certainty predictions as “unknown,” ensuring that

ambiguous cases receive further expert review. Using 5-fold cross-validation, an ensemble of models

achieved over 90% accuracy, demonstrating strong robustness and generalization. Scalable and

adaptable, UncertaintyNet can be extended to handle additional object categories and integrate with

future astronomical datasets, making it a valuable tool for reliable, automated analysis in

astrophysical research.

[Type here]
 2

INTRODUCTION:

The rapid growth of astronomical imaging data from modern telescopes and platforms like NASA’s

database, manually classifying celestial objects has become increasingly impractical. Although

identifying various astronomical phenomena is crucial for scientific research, the process is often time-

consuming and requires specialized expertise. Deep learning provides a promising solution to automate

this classification, but it still faces several challenges.

One major challenge is data imbalance, as some astronomical events are much rarer than others. This

can bias models and reduce their ability to generalize. Another important issue is the lack of reliable

uncertainty estimation—traditional deep learning models often make predictions without indicating

how confident they are, which is critical in scientific applications where false positives can have

significant consequences. Additionally, variability in data quality—differences in image resolution,

noise, and clarity—further complicates accurate classification.

To tackle these challenges, we propose UncertaintyNet, a framework that integrates transfer learning,

uncertainty modeling, and ensemble techniques. The key contributions of this work include:

• A transfer learning–based architecture that incorporates uncertainty awareness.

• A mechanism to flag predictions with low confidence as “unknown.”

• An ensemble approach to improve overall model reliability.

• A comprehensive evaluation strategy using cross-validation and detailed performance

metrics.

This framework aims to provide a more robust, reliable, and scientifically meaningful way to classify

astronomical data.

[Type here]
 3

LITERATURE REVIEW:

Deep learning is becoming an important tool in astronomy, helping scientists automatically classify

many different types of events that occur in space. Instead of focusing only on specific tasks like

identifying galaxies or detecting supernovas, modern deep learning methods are now being used to

study a wide variety of space phenomena. This makes them very useful for analyzing the huge

amount of data collected from telescopes and space missions. Still, most existing methods mainly

concentrate on improving accuracy, while paying less attention to measuring uncertainty—

something that is very important for making reliable scientific decisions.

Transfer learning has emerged as a particularly valuable approach in astronomical image

analysis due to the limited availability of labeled datasets. Ackermann et al.【4】 showed that

ImageNet-pretrained models can be effectively adapted for galaxy morphology classification, while

Gómez-Vargas and López-Cruz【5】 achieved promising results applying EfficientNet architectures

to stellar classification tasks. These studies confirm the benefits of leveraging pre-trained models to

accelerate convergence and improve generalization in astronomy-related tasks.

Uncertainty estimation in deep learning is an active research area. Methods such as Monte

Carlo dropout【6】 and ensemble learning【7】 have been shown to provide useful confidence

measures. However, their adoption in astronomical classification has been relatively limited. In

particular, few works have addressed the challenge of identifying truly “unknown” objects, which is

critical for enabling discovery of rare or previously unseen astrophysical phenomena.

[Type here]
 4

MODEL DESIGN AND EXPERIMENTAL ANALYSIS

1. Dataset and Preprocessing

1.1 Dataset Description

Our dataset consists of astronomical images representing a wide range of events occurring

in space. These images were collected from publicly available astronomical datasets and organized

according to their respective object classes. The dataset reflects the natural imbalance in the

occurrence rates of different astronomical phenomena, as some events are far more common than

others..

1.2 Image Preprocessing

All images undergo a standardized preprocessing pipeline:

1. Resizing to 224×224 pixels to match the input requirements of the EfficientNetB0

architecture.

2. RGB normalization with scaling to the [0,1] range

3. Application of EfficientNet-specific preprocessing to match the distribution of the pre-

trained weights.

1.3 Data Augmentation

To improve model robustness and address class imbalance, particularly important for

astronomical images. we employ an aggressive data augmentation strategy including:

• Random rotations up to 30 degrees.

• Width and height shifts up to 15%.

• Shear transformations up to 15%.

[Type here]
 5

• Zoom variations up to 20%.

• Horizontal and vertical flips.

• Brightness variations of ±20%.

2. Methodology

2.1 Model Architecture

UncertaintyNet is built upon the EfficientNetB0 architecture, which provides an

excellent balance between computational efficiency and its ability to capture complex patterns in

data.. We use two training phases:

1. Feature Extraction Phase:

• The pretrained EfficientNetB0 acts as a fixed feature extractor.

• Convolutional layers remain frozen.

• Only a custom classification head is trained.

• Allows the model to map extracted features to target classes quickly.

• Prevents disturbance of the pretrained weights.

2. Fine-Tuning Phase:

• After the classification head stabilizes, the last 20 layers of EfficientNetB0 are unfrozen.

• These layers are trained together with the classification head.

• A smaller learning rate is used.

• Helps the model adjust higher-level representations to capture unique characteristics of

astronomical images.

• Improves accuracy and generalization.

[Type here]
 6

The classification head consists of:

Focal Loss Function

To address class imbalance, we implement a custom focal loss function:

FL(pₜ) = -αₜ (1 - pₜ)^γ log(pₜ)

In this formula, pₜ represents the predicted probability of the true class, αt is a weighting

factor that addresses uneven distribution of categories, and γ is a focusing parameter that reduces the

loss contribution from well-classified examples and emphasizes harder, misclassified ones.

[Type here]
 7

Why Focal Loss Was Used

• The dataset suffers from class imbalance, with some astronomical events being rare.

• Traditional loss functions (e.g., cross-entropy) tend to bias the model toward majority
classes.

• Focal Loss reduces the weight of well-classified (easy) samples.

• It emphasizes hard and minority-class examples, improving recognition of rare

• Uses two parameters:

o α (alpha): balances the contribution of each class.

o γ (gamma): increases focus on misclassified or difficult samples.

• Enhances the model’s robustness, generalization, and accuracy across all categories.

2.2 Uncertainty Quantification

Our approach introduces a confidence threshold mechanism that evaluates each prediction as
follows:

1. Extract the maximum probability across all classes.

2. Compare this confidence value against an optimal threshold determined during validation.

3. If the confidence is below the threshold, the prediction is labeled as "unknown".

By adopting this approach, the model avoids overconfident errors by flagging uncertain cases

instead of assigning unreliable labels.

2.3 Ensemble Learning

To further enhance reliability, we implement an ensemble approach by:

1. Training multiple models using k-fold cross-validation.

2. Averaging prediction probabilities across all models.

3. Applying the confidence threshold to the averaged probabilities.

[Type here]
 8

This ensemble approach reduces variance and improves overall prediction reliability.

2.4 Training Procedure

Our training procedure consists of:

1. Stratified 5-fold cross-validation.

2. Two-phase training for each fold (feature extraction followed by fine-tuning).

3. Optimal threshold determination for each fold.

4. Final model training on the complete dataset.

5. Generate final predictions by averaging probabilities across fold models and applying the
confidence threshold.

Early stopping stops training once the model stops improving, avoiding unnecessary

[Link] rate reduction gradually decreases the step size, helping the model learn more

precisely and stabilize.

3. Experiments and Results

3.1 Experimental Setup

All training was conducted using TensorFlow on GPU-enabled hardware. For final

evaluation, the dataset was split into 85% training and 15% testing, with stratification to preserve

class distribution.

Hyperparameters were tuned as follows:

• Learning Rate: 1×10⁻³ during Phase 1 (feature extraction) and 1×10⁻⁵ during Phase 2 (fine-

tuning).

• Batch Size: 32 in Phase 1, reduced to 16 in Phase 2 for more precise updates.

• Epochs: Up to 10 in Phase 1 and 15 in Phase 2, with early stopping to prevent overfitting.

[Type here]
 9

3.2 Classification Performance

The model performed strongly across all categories, with particularly high accuracy for

supernovas and black holes. Performance metrics for each class are shown in Table 1.

Table 1: Classification Performance Metrics

Class Precision Recall F1-Score

Supernova 0.94 0.92 0.93

Gamma Ray Burst 0.87 0.89 0.88

Neutron Star 0.91 0.86 0.88

Black Hole 0.95 0.93 0.94

Weighted Avg 0.92 0.90 0.91

.3.3 Uncertainty Threshold Optimization

Determining an optimal confidence threshold is central to our framework, as it balances

classification accuracy with the rejection rate. Through systematic validation experiments, we

identified a threshold of 0.70 as optimal, yielding:

• 92% of above-threshold predictions being correct

• 15% of test samples being flagged as "unknown"

• F1-score improvement of 4% on the remaining predictions

[Type here]
 10

Figure 1 provides a visual example of this analysis, illustrating the effect of varying

confidence thresholds on accuracy and F1-score. The plot clearly shows the trade-off between

rejection rate and classification performance, with the optimal threshold marked at 0.70.

3.4 Ensemble Performance

To further enhance reliability, we implement an ensemble approach by:

• Training multiple models using k-fold cross-validation.

• Averaging prediction probabilities across all models

• Applying the confidence threshold to the averaged probabilities.

This ensemble method reduces variance and improves prediction stability. Compared to individual

models, it achieved about a 3% accuracy improvement and produced more consistent confidence

levels, making the uncertainty estimates more trustworthy.

[Type here]
 11

4. Discussion

4.1 Significance of Uncertainty Awareness

• The model can indicate when it is uncertain about a prediction.

• This prevents blindly trusting incorrect outputs and reduces the risk of errors.

• In astronomy, this is crucial since wrong classifications can waste telescope time,
resources, or mislead research.

• Confidence signals allow scientists to prioritize uncertain cases for expert review.

• This leads to more reliable results and better use of observational resources.

4.2 Limitations

Despite the promising results, several limitations should be acknowledged:

1. Generalization to other datasets – The model’s performance on data from different

telescopes or instruments may be limited if the training set does not include sufficient

diversity in imaging conditions.

2. Binary thresholding – The current accept/reject threshold is a simple binary decision.

More nuanced measures of uncertainty, such as probabilistic ranges or multiple

uncertainty tiers, could provide richer information.

3. Handling extremely rare events – Phenomena that are exceptionally rare may still be

misclassified or consistently flagged as “unknown” due to insufficient representation in

the training set.

[Type here]
 12

Result:

[Type here]
 13

4.3 Future Work

4.3.1 Subtype Classification

While UncertaintyNet effectively classifies astronomical objects into four major

categories—Supernovas, Gamma Ray Bursts, Neutron Stars, and Black Holes—it currently lacks

the ability to identify detailed subtypes. Future development will focus on enhancing model to

improve scientific insights.

[Link] Supernovas:

• Differentiate Type Ia subtypes (normal vs. peculiar) for improved cosmological

modeling.

• Classify core-collapse types (II-P, II-L, Ib, Ic) based on stellar evolution.

• Detect rare superluminous supernovas.

• Enable early-phase detection for timely follow-up studies.

[Type here]
 14

[Link] Gamma Ray Bursts (GRBs):

• Distinguish short (<2s) vs. long (>2s) GRBs.

• Classify optical/X-ray/radio afterglows.

• Incorporate host galaxy data for improved accuracy.

• Identify ultra-long GRBs as a distinct class.

[Link] Neutron Stars:

• Identify various pulsar types (e.g., millisecond, magnetar).

• Detect neutron stars in binary systems.

• Classify X-ray binaries and isolate interacting cases.

[Link] Black Holes:

• Classify by mass (stellar, intermediate, supermassive).

• Detect accretion states and jet morphologies.

• Identify binary black hole systems near merger stages.

While these are only illustrative examples, our broader goal is to build a framework that can

adapt to all kinds of astronomical events. To achieve this, the model will need larger, well-annotated

datasets and architectural improvements that support hierarchical classification

[Type here]
 15

6.3.2 Methodological Improvements

Bayesian Neural Networks – In future work, we plan to explore Bayesian neural networks,

as they can capture uncertainty in a more reliable and mathematically sound way.

Active Learning – We also intend to use active learning, which helps the model focus on the

most useful and confusing samples, making the training process more efficient.

Advanced Ensemble Strategies – Another direction is to develop smarter ensemble methods

that give more weight to models performing better during validation, leading to more

dependable predictions.

Self-Supervised Learning – Finally, we aim to apply self-supervised learning so the model

can first learn from large amounts of unlabeled astronomical data and then fine-tune on

labeled datasets, boosting generalization.

CONCLUSION:

In this work, we introduced UncertaintyNet, a deep learning framework that not only

performs accurate classification of astronomical objects but also quantifies the confidence of its

predictions. This combination of accuracy and uncertainty awareness makes the system particularly

well-suited for modern astrophysical research, where the consequences of misclassification can be

significant. By flagging uncertain cases as “unknown”, the model enables astronomers to focus expert

attention on ambiguous events while relying on automated classification for the majority of cases.

Our approach integrates multiple proven techniques—transfer learning with

EfficientNetB0, focal loss for handling class imbalance, and ensemble learning for improved

stability—while adding a crucial layer of transparency through its confidence thresholding mechanism.

[Type here]
 16

Experimental results, validated through 5-fold cross-validation, show that UncertaintyNet

consistently achieves over 90% accuracy while providing interpretable confidence metrics.

The principles demonstrated here are not limited to astronomy. The same framework can

be adapted to other fields where both high accuracy and trustworthy decision-making are required,

such as medical diagnostics, remote sensing, and high-stakes industrial applications. Looking ahead,

future work will focus on expanding the model’s capabilities to include subtype classification,

integrating more advanced uncertainty estimation methods, and exploring self-supervised learning to

leverage the vast amounts of unlabeled astronomical data available.

REFERENCES

1. Dieleman, S., Willett, K. W., & Dambre, J. (2015). Rotation-invariant convolutional neural

networks for galaxy morphology prediction. Monthly Notices of the Royal Astronomical

Society, 450(2), 1441–1459.

2. Cabrera-Vives, G., Reyes, I., Förster, F., Estévez, P. A., & Maureira, J. C. (2017). Deep-

HiTS: Rotation invariant convolutional neural network for transient detection. The

Astrophysical Journal, 836(1), 97.

3. Gieseke, F., Bloem-Reddy, B., Altaf-Ul-Amin, M., Kanaya, S., & Ster, C. (2017). Fast

classification of small astronomical objects. Astronomy and Computing, 19, 118–129.

4. Ackermann, S., Schawinski, K., Zhang, C., Weigel, A. K., & Turp, M. D. (2018). Using

transfer learning to detect galaxy mergers. Monthly Notices of the Royal Astronomical

Society, 479(1), 415–425.

5. Gómez-Vargas, G. A., & López-Cruz, O. (2020). Stellar classification using EfficientNet:

Deep learning in astronomy. arXiv preprint arXiv:2012.14211.

[Type here]