MiraBest FR-I/FR-II Dataset

• The dataset standardizes radio galaxy classification by manually labeling AGN as FR-I, FR-II, hybrid, or unclassifiable, enabling benchmark ML experiments.
• It incorporates precise pre-processing, normalization, and batching techniques to ensure compatibility with deep learning libraries like PyTorch and Keras.
• MiraBest’s detailed subclass structure and confidence annotations support nuanced analysis of low-redshift radio galaxy morphology.

Searching arXiv for MiraBest and closely related catalog papers. The MiraBest FR-I/FR-II dataset is a publicly available batched dataset of 1256 radio-loud AGN from NVSS and FIRST, filtered to $0.03 < z < 0.1$, manually labelled by Miraghaei and Best (2017) according to the Fanaroff-Riley morphological classification, and created for machine learning applications compatible with standard deep learning libraries (Porter et al., 2023). It is designed around the morphological distinction between FR-I and FR-II radio galaxies, while also retaining hybrid and unclassifiable cases during curation. Within the literature summarized here, MiraBest occupies a specific role: it standardizes a manually classified radio-galaxy sample for benchmarking automated classification methods, and it is frequently discussed alongside low-redshift morphology-focused catalogues such as FRII${\sl CAT}$, which was constructed from related survey infrastructure but with different selection priorities (Capetti et al., 2017).

1. Definition and relation to Fanaroff-Riley morphology

MiraBest is specifically created for the morphological classification of radio galaxies using the Fanaroff-Riley scheme, which separates radio-loud AGN into two primary classes based on the location of their brightest regions of emission relative to the core (Porter et al., 2023). In the dataset description, FR-I sources are defined as core-brightened sources, with peak luminosity within half of the overall source radius, whereas FR-II sources are edge-brightened, with peak luminosity beyond half the radii. The dataset offers manual expert labels distinguishing sources as FR-I, FR-II, hybrid, or unclassifiable, closely following the original morphological definitions.

The classification logic is consistent with the classic Fanaroff-Riley division summarized in the related FRII${\sl CAT}$ discussion, where the original FR-I/FR-II division is defined by

$$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$

with FR-I corresponding to $R_{FR} < 0.5$ and FR-II to $R_{FR} > 0.5$ (Capetti et al., 2017). In MiraBest, however, the operational emphasis is on manual visual labelling rather than on a strict quantitative $R_{FR}$ measurement. This suggests that the dataset prioritizes expert morphological judgement as the ground truth for downstream machine-learning use.

2. Sample construction and selection constraints

The sources classified in MiraBest are drawn from the Miraghaei & Best (2017) catalogue. The parent list is derived from cross-matching Sloan Digital Sky Survey (SDSS) DR7 galaxy catalogues, NVSS, and FIRST; both NVSS and FIRST are 1.4 GHz VLA surveys (Porter et al., 2023). This survey combination parallels the infrastructure used in FRII${\sl CAT}$, which was also based on a sample formed by combining observations from the NVSS, FIRST, and SDSS surveys (Capetti et al., 2017).

MiraBest applies a multi-stage filtering procedure. The redshift interval is

$0.03 < z < 0.1,$

with the lower limit excluding very nearby sources whose large angular extent could introduce SDSS parameter errors, and the upper limit ensuring spectroscopic classification in SDSS is available (Porter et al., 2023). A flux density cut is imposed at

$S_{1.4\,\mathrm{GHz}} > 40\,\mathrm{mJy},$

to ensure that sources are bright enough for extended emission detection. Sources with angular size greater than ${\sl CAT}$0 arcseconds are excluded because images are processed to ${\sl CAT}$1 pixels, corresponding to approximately ${\sl CAT}$2, so the full source must fit within the image frame.

Additional exclusions remove sources not morphologically classifiable as FRI/FRII, sources outside the FIRST imaging coverage, and any subclass represented by only a single source for train/test splitting (Porter et al., 2023). After filtering, 1256 sources are included in MiraBest.

A direct comparison with FRII${\sl CAT}$3 clarifies the distinct design objective. FRII${\sl CAT}$4 selected only sources with ${\sl CAT}$5, edge-brightened radio morphology, and at least one emission peak located at radius ${\sl CAT}$6 larger than 30 kpc from the center of the host, with inclusion requiring agreement by at least two out of three classifiers and exclusion of Wide Angle Tail sources (Capetti et al., 2017). By contrast, MiraBest is framed explicitly as a machine-learning dataset and retains a broader morphological taxonomy, including hybrids and uncertainty annotations. The data further states that only ${\sl CAT}$7 overlap exists in either FR-I or FR-II classes within ${\sl CAT}$8–0.15, mainly because MiraBest required multi-component FIRST objects, whereas FRII${\sl CAT}$9 included single- or two-component sources and prioritized morphological purity (Capetti et al., 2017). This suggests that differences in sample construction materially affect catalogue overlap even when the underlying survey basis is similar.

3. Manual labelling and subclass structure

MiraBest uses visual inspection of NVSS or FIRST images, with human experts determining FR class by applying the classic Fanaroff-Riley definition: core-dominated for FRI and edge-brightened for FRII, based on where the brightest flux occurs versus the source center (Porter et al., 2023). The dataset records both class and confidence information.

Each object is labelled using a three-digit scheme: ${\sl CAT}$0 Within this encoding, FR class is 1 for FRI, 2 for FRII, 3 for hybrid, and 4 for unclassifiable; confidence is 0 for confident and 1 for uncertain; and the final digit specifies the morphological subclass, for example standard, double-double, wide-angle tail, head-tail, or diffuse (Porter et al., 2023). If the morphology is ambiguous or uncertain, the source is marked accordingly. Hybrid and unclassifiable sources are also identified, and the information is captured in the source filename and dataset metadata.

The retained class composition is reported explicitly. The total sample contains 591 FRI sources, 631 FRII sources, and 34 hybrids, while unclassifiable sources are removed from the main dataset (Porter et al., 2023). The detailed breakdown is as follows.

Category	Count
FRI total	591
FRII total	631
Hybrid total	34

Subclass	Count
FRI standard, confident	339
FRI wide-angle tail, confident	49
FRI head-tail, confident	9
FRI standard, uncertain	191
FRI wide-angle tail, uncertain	3
FRII standard, confident	432
FRII double-double, confident	4
FRII standard, uncertain	195
Hybrid confident	19
Hybrid uncertain	14

This subclass structure is one of MiraBest’s defining technical features. The data characterizes it as the only dataset with clearly labeled non-standard subclasses such as double-double and wide-angle tail, and as the only one with labelled confidence and detailed morphological breakdown (Porter et al., 2023). A plausible implication is that MiraBest supports both coarse binary classification and more fine-grained error analysis tied to morphological ambiguity.

4. Pre-processing, normalization, and machine-learning packaging

MiraBest was constructed to be directly compatible with deep learning frameworks such as PyTorch and Keras (Porter et al., 2023). FIRST survey images are retrieved as FITS files using the SkyView Virtual Observatory. The initial image size is ${\sl CAT}$1 pixels, corresponding to ${\sl CAT}$2 arcseconds on the sky.

The pre-processing pipeline applies sigma-clipping noise reduction using Astropy’s sigma_clipped_stats, with all pixels below ${\sl CAT}$3 of the mean set to zero, repeated up to 5 times (Porter et al., 2023). After cleaning, images are cropped to ${\sl CAT}$4 pixels centered on the source, corresponding to ${\sl CAT}$5. A fixed 150-pixel diameter circular mask is then applied to exclude pixels at large radii and remove non-central sky or background.

Normalization rescales each image to occupy the full range of PNG grey levels ${\sl CAT}$6–${\sl CAT}$7 according to

${\sl CAT}$8

The stated rationale is that this ensures machine-learning models cannot trivially learn the FR class from total flux or brightness, which is physically correlated with FR type (Porter et al., 2023). Final images are stored as single-channel greyscale PNG files to minimize storage and facilitate use in ML libraries.

For batching, the dataset is split into 8 batches of 157 images each, with 7 training batches and 1 test batch (Porter et al., 2023). The batching is done with care to balance class and subclass representation per batch, especially to avoid rare subclasses or classes being absent from test data. A Python class is provided for easy loading, inheriting from the PyTorch Dataset class; it supports both “labels” for binary FRI/FRII use and “fine labels” for detailed subclassification, and can also be used for directly downloading the dataset.

For downstream normalization in ML workflows, mean and standard deviation values are provided for the full MiraBest set: ${\sl CAT}$9 (Porter et al., 2023).

5. Dataset variants, metadata, and access

The full dataset contains 1256 sources, but several derived subsets are identified for different learning settings (Porter et al., 2023). These subsets merge or filter labels according to confidence and class scope.

Dataset	Images	Description
MiraBest	1256	Full set, all subclasses, Confident/Uncertain/Hybrid labels
MBFRFull	1222	All FRI/FRII, subclasses merged
MBFRConfident	833	Only confidently labeled
MBFRUncertain	389	Only uncertain labelled

Every source includes an associated processed PNG image. The filename encodes the 3-digit class, right ascension, declination, redshift, and angular size, and tabular metadata are also available (Porter et al., 2023). The data is publicly available via Zenodo, and the accompanying documentation provides preprocessing instructions intended to preserve compatibility when extending or merging the dataset.

Within the literature summarized here, MiraBest is described as a publicly available batched dataset and as compatible with standard deep learning libraries (Porter et al., 2023). This combination of release format, structured metadata, and loader support explains why it is commonly used as a benchmark rather than only as a static image archive.

6. Comparison with FRII$$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$0, FRI$$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$1, and the low-power FR dichotomy

MiraBest is explicitly compared with catalogue-based morphology studies derived from the same broader radio-optical survey ecosystem. FRII$$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$2 is a catalog of 122 FR II radio galaxies, selected from a published sample obtained by combining observations from the NVSS, FIRST, and SDSS surveys, with redshift $$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$3, edge-brightened radio morphology, and at least one emission peak at radius $$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$4 kpc from the host center (Capetti et al., 2017). Its radio luminosity range at 1.4 GHz is $$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$5 erg/s, and the catalog contains 90% low excitation galaxies and 10% high excitation galaxies.

The comparison is significant because FRII$$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$6 and FRI$$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$7 were constructed using similar selection and analysis techniques, enabling direct comparison of FR-I and FR-II hosts at low radio power (Capetti et al., 2017). According to the supplied data, FRII$$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$8 LEGs are mostly luminous, red early-type galaxies with black hole masses in the range $$R_{FR} = \frac{\text{distance between brightness peaks on opposite sides of the host}}{\text{total source extent}},$$9, and they are essentially indistinguishable from the FR Is belonging to FRI$R_{FR} < 0.5$0. HEG FR IIs are associated with optically bluer and mid-IR redder hosts than LEG FR IIs, and with galaxies and black holes that are smaller on average by a factor of $R_{FR} < 0.5$1. FR IIs have a factor of $R_{FR} < 0.5$2 higher average radio luminosity than FR Is, but most selected FR IIs have radio power lower, by as much as a factor of $R_{FR} < 0.5$3, than the transition value between FR Is and FR IIs found in the 3C sample (Capetti et al., 2017).

Within this context, MiraBest and FRII$R_{FR} < 0.5$4 illuminate different aspects of the same morphological problem. MiraBest is optimized for standardized machine-learning experiments, whereas FRII$R_{FR} < 0.5$5 is optimized for homogeneous astrophysical population analysis. The supplied comparison further states that despite limited overlap, the main result is robust: host properties are homogeneous at low redshift and low radio power regardless of FR morphology, whereas spectroscopic class is more discriminating (Capetti et al., 2017). This suggests that a binary FR label alone is not a sufficient summary of the underlying physical population.

7. Use in machine learning and scientific significance

The MiraBest paper reviews existing applications that utilise the dataset and situates it among other radio-galaxy classification resources (Porter et al., 2023). The data states that MiraBest has been used in attention-gated CNNs, group-equivariant CNNs, dataset-shift studies combining MiraBest with Radio Galaxy Zoo unlabelled data, Bayesian uncertainty quantification, and benchmarking against other datasets such as FR-DEEP. It is also compared with AT17 and with CRUMB, a cross-matched combination of MiraBest, FR-DEEP, AT17, and hybrid samples comprising 2100 unique sources (Porter et al., 2023).

The significance attributed to MiraBest in the supplied material is threefold. First, it is described as the largest public ML-ready image dataset for FR classification to date, with over 1200 unique sources (Porter et al., 2023). Second, it contains a breadth of morphologies, including clearly labeled non-standard subclasses and confidence annotations. Third, its processing, normalization, and batching explicitly target deep learning research, with standardized data access and batching resembling MNIST and CIFAR-10.

In astrophysical terms, the broader relevance of such a dataset is sharpened by the low-power FR-II results summarized for FRII$R_{FR} < 0.5$6. There, the correspondence between morphological classification of FR-I and FR-II and separation in radio power is reported to disappear when including sources selected at low radio flux thresholds (Capetti et al., 2017). The conclusion given is that a radio source produced by a low power jet can be edge brightened or edge darkened, and that the outcome is not related to differences in the optical properties of the host galaxy. MiraBest does not itself formulate that claim, but as a standardized morphology dataset built from expert labels, it provides an operational substrate for testing automated methods on exactly the kinds of classifications whose physical interpretation has become more nuanced. A plausible implication is that MiraBest is valuable not only for classifier benchmarking, but also for probing where morphological taxonomies align, or fail to align, with other observables in radio-loud AGN populations.