Classifying Radio Galaxies with Convolutional Neural Network

• ASKAP is expected to find about 70 million radio sources, making manual handling of that volume impracticable.The paper also describes the data volumes expected from SKA precursor facilities as formidable.Source: §1 · pp. 1–2
• Visual classification of extended sources becomes impracticable as survey sizes increase, motivating automated classification of extragalactic radio images.Source: §1 · p. 2
• Deep learning learns directly from data rather than engineered features, addressing the concern that hand-crafted features may not capture all relevant properties.Source: §1 · p. 2
• Morphology-only classification avoids requiring ancillary data, making it applicable to deep surveys and regions beyond available ancillary-data coverage.Source: §1 · pp. 2–3
• The study targets FRI, FRII, and bent-tailed galaxies; bent-tailed sources can trace galaxy clusters, particularly where optical or X-ray information is sparse.Source: §1 · p. 3

• The FRI sample combined confirmed CoNFIG sources with FRICAT additions to address the relative scarcity of FRIs in the initial sample.FRICAT contributed 219 FRIs before image-based exclusions.Source: §2 · p. 4
• Bent-tailed examples came from visually confirmed WAT and NAT sources in Proctor’s catalog, excluding uncertain identifications.Source: §2 · p. 4
• The selection process removed images with strong artefacts, multiple sources, insufficient structural detail, or source sizes too large for the cutout.Additional cross-matched sources with potentially confusing morphologies were excluded.Source: §2 · pp. 4–5
• After image-quality and morphology-confusion exclusions, the final sample contained 178 FRIs, 284 FRIIs, and 254 bent-tailed radio galaxies.Source: §2 · pp. 4–5

• CNNs process raw data by replacing the conventional neural-network dot product with convolution, using kernels or filters over image regions.Source: §3 · p. 5
• Weight sharing lets CNNs learn features that remain consistent across translations and rotations of the same class’s examples.Source: §3 · pp. 5–6
• Successive convolutional layers learn increasingly complex feature hierarchies, with multiple filters producing distinct feature maps from an input image.Figure 1 illustrates filters and the sliding kernel that generates values in feature maps.Source: §3 · p. 6, Figure 1 · p. 6
• The study uses a deep CNN for radio-galaxy classification and identifies its model as a potential basis for future transfer-learning applications.The paper describes transfer learning as adapting a pre-existing network to a new classification problem with few examples.Source: §3 · p. 6

4. NETWORK MODEL

• The modified AlexNet has 12 layers, including five convolutional layers, three pooling layers, three fully connected layers, and a scoring layer.The model uses ReLU activations, normalization, 50% dropout, and a two-class Softmax output; simpler models tested by the authors achieved accuracies below 60%.Source: Figure 2 · p. 7, §4 · p. 7

4.1. Pre-processing Sample Images

• Preprocessing suppresses pixels below a 3σ background threshold and uses rotations and flips to balance the augmented classes.Images are cropped from 300×300 to centered 150×150 patches; thresholds other than 3σ yielded accuracy below 60%.Source: §4.1 · p. 9, Figure 4 · p. 9

4.2. Training

• The original dataset is split 70/30 before augmentation, and validation images are kept unseen during training.Augmented training samples are then split 80/20 for training and testing, so those two sets overlap in their underlying source images; training used 30 epochs and a 0.01 base learning rate.Source: §4.2 · p. 10
• After 10 epochs, accuracy saturated at high values and training loss fell steeply across all three binary models.With four GPUs, training each test case took around 1.2 hours; the authors attribute rapid learning to the simplicity of the binary classifications.Source: Figure 5 · p. 10, §4.2 · p. 11

4.3. Filter Visualization

• First-layer filters increasingly distinguish the radio galaxy from its background as training progresses.Filter 12 learns sharp, high-frequency features while Filter 93 learns smoother features; the authors note that deeper-layer filters are harder to explain.Source: §4.3 · pp. 11–12, Figure 6 · p. 11

5. CLASSIFICATION MODEL

• The single three-class DCNN trained inefficiently on highly imbalanced samples, with training error rising and test accuracy below 60% across epochs.Source: §5 · p. 12
• Breaking the three-class task into three binary classifications performed better for individual classifications.The three pairings were FRI vs FRII, FRI vs bent-tailed, and FRII vs bent-tailed.Source: §5 · p. 12
• Changing the loss to Info-Gain and tuning its parameters still left training and validation results poor.Source: §5 · p. 12
• The fusion classifier combines binary-model predictions and probabilities, rejecting mixed or low-confidence votes as strange objects with probability zero.The figure depicts the three binary predictions being fed into the fusion classifier for the final classification.Source: §5 · p. 12, Figure 7 · p. 12

6. RESULTS & DISCUSSION

• Validation averages were 88% precision, 86% recall, and 86% F1; results were comparable to manual classification while many times faster.The table reports validation precision, recall, F1-score, and class support.Source: §6 · p. 13, Table 3 · p. 15
• The fusion classifier assigned a class only when a source was identified as belonging to it in both individual classifications featuring that class.The reported validation metrics are weighted averages.Source: §6 · p. 13
• Bent-tail precision was 95% and recall 79%; FRII recall was 91% and precision 75%, while FRI showed high precision and recall.Many bent-tailed sources were misclassified as FRII; the authors suggest bright spots may outweigh diffuse emission in those cases.Source: §6 · p. 13, Figure 8 · p. 14
• Average precision, recall, and F1 improved as the training sample increased; the comparison kept the validation sample fixed.The study also reports that results were comparable to manual classification while many times faster.Source: §6 · p. 13, Table 4 · p. 15
• The CNN was sensitive to preprocessing: sigma clipping was used for training and had to be repeated for predictions.One bent-tail validation image was misclassified as FRII without clipping but correctly classified with high confidence after preprocessing; effects can vary with resolution and noise statistics.Source: §6 · p. 17, Figure 10 · p. 17

7. WIDER APPLICATION OF DEEP LEARNING MODEL

• Transfer learning can reuse a trained network by freezing initial layers and retraining later ones, whose complex features may depend on survey resolution and other factors.The authors describe initial layers as learning basic features that are relatively similar across surveys.Source: §7 · p. 18
• The study does not demonstrate transfer learning because radio-image optimization lacks established guidance and depends on sample size, layer selection, layer size, and learning rate.The authors note that the network may overfit depending on the new dataset's size and content.Source: §7 · p. 18
• The authors say their model can support classification and other applications with data from other surveys, and could integrate into upcoming telescopes' science-processing pipelines.They identify transfer learning and fine-tuning as relevant to adapting a trained network to new survey data.Source: §7 · p. 18

8. CONCLUSIONS

• Bent-tailed precision is 95%, FRI precision is 91%, and FRII precision is 75%; FRI and FRII recall is 91% each, while bent-tailed recall is 79%.The authors report that these results are comparable to manual classification while being much faster.Source: §8 · p. 18
• The end-to-end DCNN removes handcrafted feature extraction and classifies a single image in less than 0.17 seconds, with inference speed potentially improved by faster GPUs or batch-size changes.The method learns directly from data rather than relying on handcrafted features.Source: §8 · p. 18
• Bootstrapping available images may narrow the feature space, and results on data from different observations may suffer; preprocessing also affects inference across surveys.Inference in non-stationary environments remains an open problem.Source: §8 · p. 19
• Unequal numbers of representative class samples can confine learned features and reduce generality and overall accuracy; the authors propose retraining with future catalog samples via transfer learning.Synthetic samples and modified loss functions had limited success in improving accuracy.Source: §8 · p. 19
• The authors plan to make the Caffe model and code publicly available, and an image-upload classification service is under construction.The service is intended to support feedback and retraining with additional samples.Source: §8 · p. 19