2.1 Observations

The Galactic Plane Monitoring survey was performed using the Phase II ‘extended’ configuration of the MWA, under project code G0080. The GPM was conducted over three observing semesters, described in Table 1.

Table 1.

Observation details for the three conducted epochs of the GPM. Not all observations could be searched for transients on all timescales due to ionospheric phase scintillation at low elevations (see Section 2). $\tilde{N_\mathrm{tiles}}$ indicates the median number of functioning tiles in the array during each epoch.

Figure 2.

Pointing centres of the GPM over semesters 2022-A and 2024-A (blue), and 2024-B (lavender). Two pointings could not be scheduled and are shown in grey. The 20 $^{\circ}\times20^{\circ}$ boxes shown are at the $\sim$ 50% level of the MWA primary beam, relative to the maximum of each pointing centre. Detected transients described in Section 3 are overlaid with markers.

The MWA is capable of observing 30.72 MHz of instantaneous bandwidth placed within 72–300 MHz, with optimum instrumental sensitivity over 100–185 MHz. However, sky noise is dominant over the system temperature and is proportional to $\nu^{-2.7}$ (Rybicki & Lightman Reference Rybicki and Lightman1979), so similar noise levels are observed across 154–231 MHz when imaging (Hurley-Walker et al. Reference Hurley-Walker2017). The dispersive and scattering effects of the interstellar medium scale as $\nu^{-2}$ and $\nu^{\frac{-11}{3}}$ , respectively (Taylor & Cordes Reference Taylor and Cordes1993), and ionospheric phase shifts scale as $\nu^{-2}$ (Thompson et al. Reference Thompson, Moran and Swenson2001), with amplitude scintillation sometimes inducing extreme flux density fluctuations at the lowest frequencies (Waszewski et al. Reference Waszewski, Morgan and Jordan2022). These considerations might argue for as high an observing frequency as possible. However, the frequency range above 231 MHz is often contaminated with military satellites, and has a complex primary beam response with strong sidelobes (Cook et al. Reference Cook, Seymour and Sokolowski2021). To compromise between these competing effects, the GPM was performed using 185–215 MHz.

Galactic radio transients may produce time-variant emission on a range of timescales from less than a second (e.g. pulsars; Hewish et al. Reference Hewish, Bell, Pilkington, Scott and Collins1968) to many years (e.g. X-ray binaries; Markoff et al. Reference Markoff, Falcke and Fender2001). Detecting fast transients such as canonical pulsars and Fast Radio Bursts requires very high sampling rates such as provided by the Voltage Capture System (Tremblay et al. Reference Tremblay2015), as employed, for example, by the Southern-sky MWA Rapid Two-metre (SMART) pulsar survey (Bhat et al. Reference Bhat2023). High time and frequency resolution avoids smearing the short-duration ( $\lesssim10$ ms) pulses and enables coherent correction of dispersive effects. This is impossible with standard correlator data, although some fast ( $\sim$ 0.5 s) signals can be recovered, such as interplanetary scintillation of compact radio sources (Morgan et al. Reference Morgan2018). Processing such fast timescales induces a large computational and data-handling challenge. The goal of the GPM is to detect long-period radio transients and other Galactic radio transients, so to compromise between time-resolution and computational cost, while still retaining the highest-resolution data in case of a detection, we perform the observations at 0.5 s resolution, but average them to 4 s for onward processing.

Formed by an analogue beamformer, the MWA primary beam is fixed on the sky for each observation, and the sky drifts through it, slowly changing the apparent flux density of radio sources. To minimise this effect, while still making each observation useful for detecting sources with pulses 30–60 s across, we integrate for 296 s per observation. The MWA-X correlator upgrade has enabled seamless transition between observations such that there are no flagged time-steps between observations (Morrison et al. Reference Morrison2023).

Ten pointings were selected to cover the largest fraction of Galactic volume visible to the MWA at night-time in semesters 2022-A and 2024-A. The centres of the pointings are set to $b=0^{\circ}$ , and span $300^{\circ} \lt l \lt 57^{\circ}$ in 13 $^{\circ}$ intervals. Each pointing has an approximate full-width-half-maximum (FWHM) of 16–20 $^{\circ}$ , varying somewhat with pointing position. This yields an overall coverage of $|b|\lt15^{\circ}$ and $284^{\circ}\lt l \lt73^{\circ}$ , with reduced sensitivity toward the edges. At $l\gt50^{\circ}$ , imaging artefacts from Cygnus A result in decreased sensitivity; close to the Galactic centre, the sky temperature is higher, and at lower Declinations, the primary beam sensitivity is decreased. For 2024-B, we scanned the rest of the Southern Galactic plane (despite the lower Galactic volume). Here we attempted a closer longitude spacing of $7^{\circ}$ , but some pointings could not be included in the schedule, leading to somewhat unequal spacing of the 11 successful pointings. Figure 2 shows the layout of the pointings as well as the detected transients (Section 3).

We found that at low elevations, strong ionospheric phase scintillation induced a large number of false positive detections on short (Section 2.3) to medium (Section 2.4) timescales, either by causing sources themselves to appear and disappear, or by affecting bright sources nearby, causing their sidelobes to appear as transients. In combination with the scheduling requirement to fit as many pointings as possible in the night-time available, in a sequential order in Galactic longitude, this effectively placed further longitude limits on the searchability of each semester, shown in Table 1.

Figure 3 shows an overview of the main stages of the processing described in the subsequent sections.

Figure 3.

Diagram showing the data flow through the Galactic Plane Monitor transient search and deep mosaic creation. In each of the three epochs (2022-A, 2024-A, and 2024-B, there are M nights of observing (grey oblongs), each of which is internally composed of N snapshot observations (yellow boxes labelled ‘Obs’ in the diagram). M for each epoch is given in Table 1. Individual observations are searched for candidates (Section 2.3), catalogues from multiple observations pointed at a single location in a night are searched for transients (Section 2.4), and mosaics formed from a night of observations are also searched (Section 2.5). A subset of the highest-quality images from all three epochs is formed into a deep mosaic covering the entire GPM region (Section 2.6).

Figure 3. Long description

A diagram showing the data flow through the Galactic Plane Monitor transient search and deep mosaic creation. The diagram includes three epochs labeled as 2022-A, 2024-A, and 2024-B, each containing M nights of observing represented by grey oblongs. Each night is composed of N snapshot observations shown as yellow boxes labeled Obs. Individual observations are searched for candidates, and catalogues from multiple observations pointed at a single location in a night are searched for transients. Mosaics formed from a night of observations are also searched. A subset of the highest-quality images from all three epochs is formed into a deep mosaic covering the entire GPM region.

2.2 Imaging and calibration

The data were downloaded from the All-Sky Virtual ObservatoryFootnote ¹ without calibration solutions applied, with RFI flagging performed by AOFlagger (Offringa et al. Reference Offringa, van de Gronde and Roerdink2012). We performed (phase and amplitude) calibration on dedicated pointings toward Hercules A, which at 200 MHz is $\sim377$ Jy (Hurley-Walker et al. Reference Hurley-Walker2017), sufficient to dominate the visibilities and produce good calibration solutions. We used the GLEAM-X pipeline (Hurley-Walker et al. Reference Hurley-Walker2022b) which incorporates this and other sources in the field to produce a sky model, then MitchCal (Offringa et al. Reference Offringa2016) to form complex antenna gains for each antenna, frequency, and polarisation. These were applied to all data for each night. Imaging was also performed with the GLEAM-X pipeline, using WSClean (Offringa et al. Reference Offringa2014) to create a continuum image of each 5-min snapshot (top-right panel of Figure 4). For speed, we decreased the clean thresholds to 5- $\sigma$ (peak detection) and 3- $\sigma$ (masked clean), as well as reducing the maximum number of major iterations to 3. Subtracting the deep visibility model formed in this process, the data are re-imaged at 4-s time resolution to produce a series of near-empty snapshots in which only transient and variable sources are visible (Hurley-Walker et al. Reference Hurley-Walker2022b; Horváth et al. Reference Horváth, Hurley-Walker, McSweeney, Galvin and Morgan2025), which are thence combined into a cube in 16-bit HDF5 format for compact storage and onward analysis, as described in Appendix B of Morgan et al. (Reference Morgan2018). An example slice of such a cube is shown in Figure 4, demonstrating that diffuse emission and point sources are subtracted well by the initial imaging process, despite reasonably significant ionospheric blurring effects in the displayed observation. Figure 5 shows the typical RMS as a function of Galactic longitude.

Figure 4.

Examples of the four imaging products generated in this work, using the same region for reference. Top-left: a single slice from a continuum-subtracted time-step cube (Section 2.3); top-right: a single five-minute snapshot (Section 2.4); bottom-left: a night-long mosaic (effective integration time $\sim$ 30 min: see Section 2.5); bottom-right: the deep mosaic formed with the ‘lucky imaging” approach (Section 2.6). The RMS noise of each image is, respectively, 110, 26, 14, and 5 mJy beam $^{-1}$ . All images except the bottom-right were generated from observations on 2022-06-30 at $\sim$ 16h UTC. In each panel, a 15 $^{\prime}$ square region around the LPT GPM J1839 $-$ 10 is highlighted, and displayed in the bottom-left insets. Within these, the PSF of each image is shown as white ellipse in the bottom-left corner.

Figure 5.

Noise levels of the various data products formed in this work, as a function of Galactic longitude. The points show median values while the error bars indicate the 1st (lower) and 3rd (upper) quartiles. The formation of the cubes and snapshots is described in Section 2.2, and the mosaics in Section 2.5. The change in spacing of the observations over longitude is due to a combination of a change in pointing separation between 2022-A/2024-A ( $13^{\circ}$ ) and 2024-B ( $7^{\circ}$ ), with gaps resulting from unforeseen limitations in the scheduling (see Section 2.1 for details). The relatively high RMS of the mosaics at $l= 285^{\circ}$ is due to this region lying on the edges of both the GC and SP regions.

2.3 Intra-snapshot search

Individual observations are sensitive to transients on timescales between the imaged time resolution (4 s), to approximately half of the observation duration ( $0.5\times 5$ min). To extract transients on these timescales from within five-minute snapshots, we used filter methods developed by Horváth et al. (Reference Horváth, Hurley-Walker, McSweeney, Galvin and Morgan2025) in their search of a subset of GLEAM-X data. Briefly, these consist of:

1. 1. A sum of Gaussian matched-filters of 4, 8, 16, and 32 s widths;

2. 2. A search for any pixel that has brightness $\gt$ 7 $\times$ the local RMS of that voxel along the time axis;

3. 3. A search for any voxel that has an RMS along the time axis larger than $5\times$ the median value of other voxels in the image.

The ionosphere induces phase scintillation of radio sources over timescales of $\sim$ 30 s, and this can produce a large number of false positives. To combat this, we excluded any candidate with a peak brightness equal to or less than a known radio source in the GLEAM catalogues (Hurley-Walker et al. Reference Hurley-Walker2017; Hurley-Walker et al. Reference Hurley-Walker2019), cross-matched within a 2 $^{\prime}$ radius. The detection sensitivity of the latter catalogue, which covers $|b|\lt10^{\circ}$ , is similar to the per-snapshot sensitivity of the GPM data, and thus this cross-match performs efficiently.

Some observations still produced large volumes of candidates due to instrumental or calibration errors. Any observation that produced more than 10 candidate transients after the above ionospheric filtering was excluded from onward processing.

After selecting candidates that passed any of the three filters in acceptable observations, we produced light curves and animations of the surrounding $200\times200$ pixel ( $45'\times45'$ ) regions for visual inspection.

2.4 Inter-snapshot search

Each 30- or 45-min integration points within a few degrees of the same location in the Galactic plane (limited by the analogue beamformer), divided into 6–9 snapshot observations. Transients on timescales of the snapshot time (5 min) to the total integration time (30–45 min) can be detected by performing cross-matches of the source catalogues detected in these snapshots and performing some statistical tests.

We use Aegean v2.2.5Footnote ² (Hancock et al. Reference Hancock, Murphy, Gaensler, Hopkins and Curran2012; Reference Hancock, Trott and Hurley-Walker2018) and its companion tools such as the Background and Noise Estimator (BANE) to perform source-finding, with the default settings. We use the flux_warp Footnote ³ package (Duchesne et al. Reference Duchesne, Johnston-Hollitt, Zhu, Wayth and Line2020) applying a set of radial basis functions to bring the source flux densities into agreement with the GLEAM-X GP catalogue of Mantovanini et al. (Reference Mantovanini2025c), for $|b|\lt10^{\circ}$ and $233^{\circ} \lt l \lt 44^{\circ}$ , and GLEAM outside of this region (Hurley-Walker et al. Reference Hurley-Walker2022b). We use the fitted 200-MHz flux densities from these catalogues to best-match the observing frequency of the GPM.

We also apply fits_warp (Hurley-Walker & Hancock Reference Hurley-Walker and Hancock2018) to shift the radio source positions into agreement with higher-resolution radio catalogues. We used the Rapid ASKAP Continuum Survey (RACS; McConnell et al. Reference McConnell2020; Hale et al. Reference Hale2021) Low Data Release 1. These position shifts are typically $\sim$ 1–10 $^{\prime\prime}$ .

Using the Starlink Tables Infrastructure Library ToolSet (STILTS; Taylor Reference Taylor, Gabriel, Arviset, Ponz and Enrique2006), we perform a cross-match between all snapshot catalogues within a given pointing, using a 25 $^{\prime\prime}$ cross-matching radius and keeping all matches, which allows us to retain non-detections (i.e. transients). This relatively large cross-matching radius is necessary due to residual ionospheric errors caused by small ( $\sim$ degree-size patches) scale ionospheric defocussing, which cannot be corrected by fits_warp.

We then exclude from our catalogues:

• • sources with extent (integrated to peak flux density ratio) $\gt1.5$ in any image;

• • sources where Aegean failed to fit a shape in any image (usually artefacts or parts of extended structures such as supernova remnants);

• • sources more than $9.^{\circ} 5$ from the center of the combined pointings, where the coverage of the combined pointings is low and has high noise;

• • sources which are decomposed in any image into components lying within 2 $^{\prime}$ of each other.

Where a source is not detected in an image, we perform a forced photometric measurement at the location of the mean position obtained from images where the source was detected. We set the uncertainty on this flux density measurement to the local RMS noise measured by BANE at the same position.

Due to the limited number N of data points (6–9), we use the simplest transient search variability metrics: the significance of variability $\eta$ and the coefficient of variation V (McKay Reference McKay1932):

(1) \begin{equation} \eta = {\chi}^2_{N-1} = \frac{1}{N-1} \sum^N_{i=1} \frac{\left(S_{i}-\overline{S}^*\right)^2}{\sigma_i^2}\end{equation}

and

(2) \begin{equation} V = \frac{1}{\overline{S}} \sqrt{\frac{N}{N-1} \left(\overline{S^2} - \overline{S}^2\right)},\end{equation}

where $\overline{S}$ is the mean flux density, and $\overline{S}^*$ is the weighted average flux density, respectively. The standard error $\sigma_i$ on the ith flux density measurement $S_i$ is calculated as the quadrature sum of the local RMS noise and a 10 % flux density calibration error.

For sources which meet the following criteria:

• • $\eta\gt5$ ;

• • $V\gt0.5$ ;

• • mean $S/\sigma \gt 6$ , calculated using images in which the detection was made (i.e. not including forced measurements),

we generate a light curve and an animation of a $200\times200$ pixel ( $45'\times45'$ ) region around the source for visual inspection.

2.5 Mosaic search

Observing the same sky across multiple epochs also enables a search for radio transients that change between epochs, i.e. timescales of days to weeks. We use the flux-density calibrated primary-beam-corrected images where ionospheric position shifts have been removed, generated in Section 2.4. We then form weight images based on the inverse of the square of the RMS maps calculated by BANE. We noted that the primary beam correction had occasionally introduced artefacts into the edges of the images, so also cropped the outer 800 pixels (3.3 $^{\circ}$ ) of both the images and RMS maps. The RMS in these regions is typically $\sim100\times$ that of the centre part of the image, so very little useful data is lost.

We use the mosaicking software swarp (Bertin et al. Reference Bertin, Bohlender, Durand and Handley2002) to form a Cartesian FITS image in Galactic coordinates, weighting each 5-min snapshot by the weight images. The resulting images have RMS noise 7–25 mJy beam $^{-1}$ , largely depending on proximity to bright sources, around which calibration artefacts can be significant.

Night-time observations generally form good-quality images following this procedure, but daytime observations suffer significant direction-dependent time-dependent ionospheric effects, with the net effect producing significant smearing of the sources beyond the nominal point spread function. We therefore restricted our mosaicking to night-time observations. An example of the imaging quality obtained is shown in the bottom-left panel of Figure 4. Figure 5 shows the typical RMS of the mosaics within $|b|\lt10^{\circ}$ as a function of Galactic longitude.

To search for radio transients and variables, we performed a similar process to that described in Section 2.4. Several changes were made to the process to accommodate the large mosaics:

• • due to their similar sky coverage, we searched all data from semesters 2022-A and 2024-A in one analysis;

• • to account for point spread function (PSF) projection and ionospheric blurring without expending unnecessary effort in determining it exactly, we relaxed the extent constraint to $\gt3$ ;

• • instead of a pointing centre distance cut, we cut 150 pixels (36 $^{\prime}$ ) from the edges of mosaics and any areas where the local RMS noise exceeded 200 mJy beam $^{-1}$ (typically on the Northern edge where Cygnus A is a strong source in the primary beam sidelobes);

• • the source density in the mosaics is sufficiently high that imposing a 2 $^{\prime}$ ‘isolation’ cut (as performed for the snapshots) can result in the exclusion of real transients, so we adjusted this to best recover transients while not including too many false positives from sources decomposing into multiple objects:

  • − within the GC region, the calibration and imaging quality motivated a 112 $^{\prime\prime}$ radius and exclusion of groups size 2 or more;

  • − the SP region had higher imaging and calibration quality enabling a 60 $^{\prime\prime}$ radius and the exclusion of groups of size 3 or more;

• • as the image quality was generally higher in the mosaics than the snapshots, we reduced the cuts to select variable sources to:

  • − $\eta\gt2$ ;

  • − $V\gt0.3$ ;

  • − mean ‘on’ $S/\sigma \gt 5.5$ ,

then inspected the candidates visually using light curves and animations, as per Section 2.4.

2.6 Deep mosaic

By combining our accumulated $\sim19\,000$ 5-min snapshots with inverse-variance weighting, we can form a deep image of the Galactic plane at 200 MHz. Since in 2022-A and 2024-A a single calibration solution is applied for each night, and all individual snapshots are only relatively lightly-cleaned (Section 2.2), relative to the large amount of total integration time ( $\sim320$ h), the resulting mosaics become dominated not by thermal noise, but by a combination of sidelobe and classical confusion, as well as calibration and imaging artefacts. Ionospheric blurring can occasionally be significant, which would increase confusion and decrease image quality if strongly affected snapshots were included in a mosaic.

We can therefore create the cleanest and deepest mosaic by taking a ‘lucky imaging’ approach, selecting only the most high-quality observations to be included. We determine this quality by cross-matching the source-finding results of each snapshot to a catalogue of unresolved sources as determined from the SUMSS and NVSS catalogues; from this subset we calculate the mean integrated to peak flux density ratio; we retain only snapshots where this ratio is $\lt1.25$ . This yields 2 953 snapshots to be mosaicked ( $\sim$ 15% of the total). We then perform the cropping, weighting, and mosaicking steps identically to Section 2.5, with a template covering the full spatial range of data taken from 2022 to 2025. A section of this image is shown in the bottom-right panel of Figure 4.

The mosaic does not have a well-defined PSF as it is formed from such a large range of snapshots, each of which has a unique PSF. This is in contrast to the GLEAM and GLEAM-X surveys, which were deliberately conducted in drift scan modes so that the PSF would be well-defined, and any residual difference between the measured PSF and expected PSF could be attributed to ionospheric blurring, measured, and corrected.

Since the GPM visibilities lack short baselines, and there exists already a deep 200-MHz mosaic specifically produced to reconstruct the multiple spatial scales of the Galactic plane (Mantovanini et al. Reference Mantovanini2025c), the utility of the mosaic formed in this work is mainly in measuring a slightly ( $\sim2$ –3 $\times$ ) fainter point source population, as well as qualitatively revealing some of the fainter $\sim1'$ -scale filamentary structures in Galactic objects such as supernova remnants. This motivates a simplistic approach to characterising the PSF, wherein we simply measure it using unresolved sources, and do not attempt to correct for any ionospheric blurring. This is further supported by our ‘lucky imaging’ approach which should minimise the inclusion of ionospherically distorted data.

Following Hurley-Walker et al. (Reference Hurley-Walker2017) and (Reference Hurley-Walker2022b) we perform ( $5\sigma$ ) source-finding on this mosaic to produce an initial catalogue of 200-MHz sources. We use the RACS-Low 1 survey (Hale et al. Reference Hale2021) to determine which of these sources are unresolved and are therefore representative of the PSF. We concatenate the two Gaussian component RACS catalogues available from Vizier and select sources at $|b|\lt16^{\circ}$ to match our sky coverage. We then perform internal cross-match on this catalogue to remove all sources that have one or more neighbours within 45 $^{\prime\prime}$ (the resolution of the GPM). We also apply Hale et al.’s criteria to select unresolved sources (see their Section 5.2.1), furthermore selecting only sources that are well-fit by one Gaussian ( $Ng=1$ ), and brighter than 15 mJy (equivalent to 50 mJy at 200 MHz, for a typical spectral index $\alpha=-0.8$ ). We crossmatch this subset of the RACS catalogue with the initial GPM catalogue, further downselecting to sources in the GPM with $S_\mathrm{200\,MHz}/\sigma_\mathrm{200\,MHz}\gt10$ .

From this combined catalogue of isolated, compact sources, we create a PSF map following previously established methods (Hurley-Walker et al. Reference Hurley-Walker2017, Reference Hurley-Walker2019, Reference Hurley-Walker2022b). We average and interpolate the sources’ major axis $\theta_\mathrm{maj}$ , minor axis $\theta_\mathrm{min}$ , and position angle (PA) measurements, using Healpix (order $=5$ , i.e. pixels $\sim1 . \!\!^{\circ} 5$ on each side) as a natural frame in which to accumulate and average source measurements. These are projected on to a Galactic Cartesian grid for re-use. Figure 6 shows the final mosaic as well as the associated background, RMS noise, and PSF maps.

Figure 6.

Deep mosaic and its ancillary data. From top to bottom: the deep mosaic, the RMS sensitivity map, the background map, the PSF major and minor axes, and the PSF position angle.

We then use the PSF map as an input to a final source-find using Aegean, this time using a seed-clip of $4\sigma$ and a flood-clip of 3 $\sigma$ , subsequently discarding any source with a signal-to-noise of less than 4.5 $\sigma$ . We discuss the characteristics of the resulting catalogue in Section 3.4.