HIGH-VELOCITY CLOUDS IN THE GALACTIC ALL SKY SURVEY. I. CATALOG

To identify HVCs, we use an automated source finding routine which adopts the flood fill algorithm. The flood fill algorithm searches for emission connected from a peak down to a user-specified threshold, and identifies this as an object or a source. We specifically adopt the algorithm developed and used by Murphy et al. (2007), but extended to a three-dimensional cube. The flood fill algorithm works in the following way.

• 1.  
  Identify all peaks > the threshold peak level (T_pk).
• 2.  
  Extend from the peak in each dimension (in our case: α, δ, v), amalgamating pixels into one source until the cutoff level (T_min) is reached.
• 3.  
  Record this as a candidate source.

We give examples of the kinds of clouds identified by our source finder in Figure 1. In order to reduce the impact of noise and artifacts, we searched on cubes binned by averaging five channels in the spectral axis (see Section 2.3 for more details about binning). We set a minimum source width of 3 pixels in both spatial directions, which is close to the size of the GASS beam in the spatial axes, and to three-binned spectral channels (minimum total width of ∼12 km s⁻¹). For our flood fill parameters, we choose a peak of T_pk = 4σ and T_min = 2σ, extrapolating from the GASS noise of 57 mK to the binned cubes. As stated above, we adopted the rms noise of the first release of GASS to guarantee a consistent catalog sensitivity across the sky and near the Galactic plane, although the noise of GASS II is lower at higher Galactic latitudes. The nominal column density limit for a single channel at this noise of 57 mK is ∼8.5 × 10¹⁶ cm⁻².

Zoom In Zoom Out Reset image size

We examined the effect of binning using three cubes that contain predominantly noise, calculating the change in rms with increasing number of bins. The results are shown in Figure 2. We find that the GASS noise closely follows the theoretically expected trend with increasing bin number (based on Poisson statistics), although it is slightly higher than the theoretical noise. The difference across these three samples of noise peaks at a value of at most 3 mK in the case of binning by three bins and occurs at ∼2.6 mK for the case of five bins. This is a very small difference overall and simply results in probing further into the noise by assuming the theoretical limit. We thus assume the N-binned rms, σ_N, to be ∼σN^−1/2, and calculate σ₅ ∼ 25 mK, which gives 2σ₅ ∼ 50 mK and 4σ₅ ∼ 100 mK. We use these values as our T_min and T_pk respectively. For comparison to the data in our noise cubes, we found a 4σ five-binned rms of 101, 102, and 106 mK in each which is very close to the theoretical 100 mK value. Although a T_pk of 4σ is lower than most automated source finding, the combination of this with the three channel minimum in each axis increases the significance of the detections.

Zoom In Zoom Out Reset image size

One of the challenges of identifying HVCs close to Galactic velocities is distinguishing them from emission that is due solely to Galactic rotation near the center of the Milky Way. We use the parameter of deviation velocity, which indicates how much the object's velocity deviates from a simple model of Galactic rotation where v_dev = v_LSR − v_model. As HVCs can be considered to require a deviation velocity v_dev of at least 50 km s⁻¹ (Wakker 1991; Wakker & van Woerden 1997), we use this definition to help separate HVCs from high-velocity Milky Way gas. To determine v_dev, we construct a model of the Galactic H i distribution. We have chosen to use the model of de Heij et al. (2002) to facilitate comparison with past studies of HVCs and intermediate-velocity clouds (IVCs). This model is composed of a thin disk with a constant central density in H i, n₀ = 0.35 cm⁻³, and kinetic temperature, T_K = 100 K, within Galactocentric radius R = 11.5 kpc. Beyond 11.5 kpc, the midplane density in H i is

$$\begin{equation} n =n_{0} \exp \left({-\frac{R-11.5\,{\rm kpc}}{3\,{\rm kpc}}}\right). \end{equation} \tag{ 1 }$$

In the vertical direction, the H i density follows a Gaussian distribution with

$$\begin{equation} n =n_0 \exp \left({-\frac{(z-z_0)^2}{2\sigma _z^2}}\right), \end{equation} \tag{ 2 }$$

where z is the scale height of the Milky Way in pc. For R < 11.5 kpc, σ_z = 180 pc, and z₀ = 0 pc. For R > 11.5 kpc, σ_z = 180 + 80(R − 11.5 kpc) pc, and z₀ = (R − 11.5)/6sin ϕ + (R − 11.5/6)²(1 − 2cos ϕ) pc. ϕ is the cylindrical Galactic azimuth (increasing in the direction of Galactic rotation) centered on the Galactic center where ϕ = 180° is the direction toward the Sun. The variation of z₀ and σ_z beyond R = 11.5 kpc provides for the warp and flare of the Galactic disk. The Milky Way is assumed to have a flat rotation curve with R₀ = 8.5 kpc, v_rot = 220 km s⁻¹, and σ_v = 20 km s⁻¹.

We constructed a model using a 50 pc grid out to a Galactocentric radius of 50 kpc where

$$\begin{equation} T_B(v)=T_K (1-e^{-\tau _v}), \end{equation} \tag{ 3 }$$

and where

$$\begin{equation} \tau _v = \frac{33.52 n_0}{T_K \sigma _v} \exp \left({-\frac{v-v_r}{2 \sigma _v^2}}\right), \end{equation} \tag{ 4 }$$

with all temperatures in K, velocities in km s⁻¹ and densities in cm⁻³.

Using this model, we identify the most positive and negative v_LSR for H i in the model with T_B ⩾ 0.5 K at every position in the sky; these velocities are the basis for calculating v_dev. The collective anomalous velocity gas deviating from Galactic rotation in the Milky Way is typically categorized as low-velocity gas, intermediate-velocity gas or high velocity gas depending on v_LSR, v_dev or a combination of the two. This had led to the existence and studies of such objects as LVCs and IVCs as well as HVCs, which makes the physical distinction between each class of AVC more important as there is a significant degree of overlap between the three groups of anomalous velocity gas. In order to guarantee detection of cloud emission peaking at deviation velocities of 50 km s⁻¹ in accordance with the deviation velocity cutoff proposed by Wakker (1991) and in this context of AVCs, we mask out all emission in our data that has a deviation velocity of <30 km s⁻¹. We refer to any GASS clouds which do not meet the HVC velocity criteria (see Section 3) as AVCs because they exist on the blurry boundary between IVC and HVC, and as such we refrain from classifying them in absence of further investigation into their physical properties.

Due to the masking process, as well as the effects of frequency switching, we found it necessary to manually remove any sources which appeared to either be poorly masked Galactic emission, artifacts near regions of "ghosting" where some of the frequency-shifted emission is incompletely subtracted or regions of generally higher noise. The initial source finding routine resulted in ∼6000 candidate sources, which were then inspected to ensure the reliability of the final catalog. This inspection reduced the total number of clouds to ∼2000. The decision on each source was made in the context of all information available, including measured properties, spectra, moment maps and proximity to artifact regions.

To investigate the effect of the spectral binning on our ability to find narrow spectral line (narrow-line) clouds, we simulated the binning process on narrow-line clouds of varying signal-to-noise. We did this by generating model clouds of varying FWHM and varying peak brightness temperature embedded in Gaussian noise of 57 mK. We then performed binning on the spectrum of each cloud and measured the resulting properties of the binned spectrum. We considered both the best case scenario (a narrow cloud peaks within the edges of a spectral bin) and the worst case scenario (a narrow cloud peaks on the edge of a spectral bin) in order to understand at what level we are sensitive to narrow-line clouds when binning the data by five channels. The effect of a cloud's position within a spectral bin or on a bin edge is only really apparent for low brightness, very narrow clouds. This is shown in Figure 3 for a 3σ cloud with a FWHM of 3 km s⁻¹, where the red dashed line represents the binned spectrum. The detrimental effect on peak flux of falling on the edge of a bin is very clear for this type of cloud.

Zoom In Zoom Out Reset image size

We generated model clouds over the peak brightness temperature range of [1, 15] σ in steps of 1σ and over the FWHM range of [1, 20] km s⁻¹ in steps of 1 km s⁻¹ both within a spectral bin and on a bin edge. We define a cloud as detectable if, in the binned spectrum, the peak is greater than our required source finding peak (100 mK) and if the positions on either side of the peak are greater than our required source finding threshold (50 mK). We simulated each cloud at each FWHM and σ 1000 times to average out the effect of statistical noise, and kept track of how many clouds in each case are "detected." From this we generated a detectability map for clouds. This is shown in Figure 4. We see that we are likely to miss clouds of width less than ∼8 km s⁻¹ even at the maximum simulated brightness temperature of 15σ in our binned search.

Zoom In Zoom Out Reset image size

We compared this result to two-component clouds, where narrow-line components are embedded within a broad component of FWHM comparable to that of the median HVC population (∼20 km s⁻¹; Bekhti et al. 2006; Kalberla & Haud 2006). We performed the same simulations as before, except this time we added a broad component at the same position as the narrow-line component with one-third of its intensity such that the total brightness is at the simulated σ level. The detectability plots are shown in Figure 5. We see that the embedding of a narrow-line component within a broad component, even if it is faint, makes a considerable difference to the detectability of the source in total. We are likely to be complete down to the narrowest FWHM simulated for all clouds with a total peak brightness temperature greater than ∼12σ, and complete at the same FWHM limit as without a broad component (∼8 km s⁻¹) for clouds greater than 8σ.

Zoom In Zoom Out Reset image size

The completeness is much higher in this case than in the isolated narrow cloud case due to the contribution of flux from the broad component during the binning process ensuring that the narrow cloud was not lost among the noise. In the GASS catalog of HVCs, we have identified two-component clouds (either featuring a narrow-line component within a broad component or co-located components) with a T in the flag column (Section 3).

To find any isolated narrow clouds (those without a co-located broad component), we performed a targeted search for narrow clouds on the unbinned GASS II data, with a peak cutoff of 6σ (0.33 K) and a threshold of 2σ (0.11 K). This resulted in almost 3500 candidate clouds. We crossmatched these sources with those found by our main source finding method in order to remove any that are already accounted for and removed any with spectral extents greater than 15 channels (the minimum width after binning), which reduced the total number to 532 candidate narrow clouds. The distribution of all of these revealed a high concentration of artifacts at both positive and negative velocities near the southern Galactic pole and along the Galactic plane, as well as some associated with small regions of incompletely flagged RFI. We removed these sources from the candidate list, leaving 54 candidates to be investigated.

Almost all of these candidates appear at low deviation velocity, with many showing an artificial narrowness due to their proximity to the masked data. Only 14 candidates were kept, which were those showing a well-defined maximum in their spectrum indicating that most of the source is seen. Of the rest, 31 were removed due to their spectra being cutoff by the deviation velocity mask and 9 were removed as noise. We added the 14 sources to our final catalog, with a flag of N identifying them as separately found narrow-line sources. While this procedure was necessary to ensure our main source finding had not missed a significant population of narrow-line clouds due to the effect of binning the data, we conclude that there were very few narrow clouds missed.

We have summarized the key statistical values for each cloud property in Table 3, with histograms for properties of the GASS HVCs shown in Figures 8 and 9. In each histogram plot, we have separated the HVCs (shown in black) from the AVCs (shown in red). The very largest GASS sources (e.g., the two large positive-velocity sources which appear to be unmasked Galactic plane emission seen in Figure 6) are a unique population due to their large area on the sky and were not included in the statistics or histograms. There are six of these large regions in total, covering the Magellanic Clouds, Magellanic Stream, Magellanic Leading Arm, Complex L and the Galactic center region.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

There is no obvious trend in right ascension or declination for either of the populations, although a slightly increased number of sources near right ascension of 0° is likely due to the influence of the Magellanic Stream. There is an evident gap between 100° and 150° in the AVC distribution due to the effect of deviation velocity masking near the Galactic plane (which occupies most of this angle range). The declination distribution tails off at declinations close to 0° because of the increased noise in the GASS data at declinations that were observed at low elevation by the Parkes telescope. This effect is far less exaggerated for the AVCs, which may be a product of the lower number of AVCs as well as the influence of the Magellanic system on the HVC distribution at declinations ∼ − 40°. Similarly, there are no clear trends in Galactic latitude or longitude due to the limitations of a southern sky only survey.

All velocity properties are shown in Figure 8. The distribution in LSR velocity and deviation velocity is dominated by positive clouds due partly to the influence of the Magellanic Clouds and Stream, and partly because the southern view of the Galaxy is also dominated by gas at positive velocities. The gap in the middle of both velocity histograms is due to the effect of searching only at deviation velocities higher than 30 km s⁻¹. The AVC population occupies the low end of both deviation velocity and LSR velocity by definition, although it is worth noting that the AVC population is dominated by sources which have both low deviation velocities and low LSR velocities, which is not necessarily to be expected.

We see a preference toward negative velocities with respect to the GSR in both HVCs and AVCs, although this is evident as an asymmetric distribution for HVCs and as a shifted distribution for AVCs. This bias toward negative velocities is consistent with previous findings in both neutral hydrogen (Stark et al. 1992) and ionized hydrogen (Haffner et al. 2003).

The distribution in FWHM of the HVC population is fairly even around the median value of 19 km s⁻¹, with a slight tail toward higher FWHMs. The AVC population is clearly shifted toward lower FWHMs than the HVC population. Based on inspection of the spectra of AVCs, this shift appears to be genuine but we cannot exclude the possible effect of measuring the FWHM on partially cutoff spectra (which is the case for 213 of the 295 AVCs) leading to increased uncertainty and artificial narrowing. We find our distribution is radically different to the HIPASS HVCs, although this is not surprising given the much coarser velocity resolution of HIPASS. In the case of HIPASS, the majority of their population was not completely resolved at the large velocity resolution of 26.4 km s⁻¹. A comparison between the two catalogs is shown in Figure 10, with the respective velocity resolutions plotted as dashed lines. We show both the measured FWHMs of the HVC population in each survey and the FWHM distribution corrected for velocity resolution, using

$$\begin{equation} \hbox{FWHM}_{{\rm observed}} = \hbox{FWHM}_{{\rm actual}}^2 - \Delta v^2, \end{equation} \tag{ 7 }$$

which reveals good agreement between the cloud populations of GASS and HIPASS. We include in this comparison only those HIPASS clouds which are comparable with the GASS population, namely with declinations <0° and excluding galaxies.

Zoom In Zoom Out Reset image size

It is worth noting that difference between our distribution centered on a FWHM of 19 km s⁻¹ is somewhat different to that found in the northern hemisphere Leiden/Dwingeloo Survey or the later LAB survey data which found a distribution centered on ∼25–30 km s⁻¹ (de Heij et al. 2002; Kalberla & Haud 2006), and to the distribution found by the 140 foot telescope in Green Bank which centered on 23 km s⁻¹ (Cram & Giovanelli 1976). However we note that both Cram & Giovanelli (1976) and Kalberla & Haud (2006) chose to separate velocity structure into multiple spectral components, including spatial data across their sources. To better investigate our catalog's trend toward lower FWHMs and closer match the methods of previous studies, we extracted all spectra containing at least a 5σ signal from the spatial regions of all GASS HVCs with a total area <5 deg². This resulted in a total of 13,977 spectra, of which 12,180 were used as we removed any with fit LSR velocities of <90 km s⁻¹.

To determine whether a spectrum was better fit with a single component or two components, we performed fitting for both cases and chose a two-component spectrum fit if the standard deviation of the single fit residual was larger by 10% or more than the standard deviation of the two-component fit residual. This gave a total of 13,593 individual spectral components within these 12,180 spectra, of which 10,767 were identified as single-component spectra. The results are shown in Figure 11. In this case of spectral components distributed spatially across GASS HVCs, we see a median FWHM of 21 km s⁻¹ (σ = 9 km s⁻¹). Compared with the result of Kalberla & Haud (2006), we find that our spectral component distribution is much narrower and concentrated on a significantly lower σ. Although there are similarities in both distributions, our narrow components, although they appear in the total distribution, are not as pronounced and we see evidence for two-component structure in the distribution of spectra identified to be single component. The incongruity is likely due both to the differing angular resolution of GASS and LAB and to the differences between their targeted population of the HVC complexes identified by Wakker & van Woerden (1991) and our all-southern sky population identified from GASS.

Zoom In Zoom Out Reset image size

We further investigated the lower median FWHM of the GASS HVCs by measuring their properties as found in the LAB data (Kalberla et al. 2005). We extracted the LAB spectrum closest to the best-fit position of the GASS HVCs with fitted brightness temperatures >5σ (to aid detection) and performed the same fitting routine on each cloud. We show the results for all 5σ and 10σ clouds in Figure 12. In the case of the measured FWHMs in the LAB data, we see fairly good agreement with the GASS FWHM, although the entire distribution is positively biased. We find a median FWHM for our 5σ clouds measured in the LAB data of ∼24 km s⁻¹ and a mean FWHM of ∼28 km s⁻¹. Based on these results, it appears possible that our lower median FWHM of 19 km s⁻¹ compared to the previous studies of ∼28 km s⁻¹ is a combination of the effects of slightly coarser spectral resolution (1.3 km s⁻¹) and lower angular resolution (36') in the LAB data rather than any significant difference from the HVC populations studied previously. We show the two distributions together in Figure 13.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

In fitted peak brightness temperature (Figure 9), we see a distribution approximately following a power law, with the dashed lines on the histogram corresponding to 3σ and 5σ. Many sources have fitted peak brightness temperatures below this cutoff because their actual peak (which was targeted during the source finding procedure) is affected by positive noise bias. This effectively increases the sensitivity of the catalog below our target of 4σ, though the completeness is likely to be limited at lower brightness temperatures. A linear fit for both HVC and AVC distributions gives slopes of −1.4 and −0.9 respectively, with good approximation by a power law.

The column density distributions of HVCs and AVCs show an approximate power law distribution that turns over between 10¹⁸ and 10¹⁹ cm⁻². This suggests that we are particularly incomplete below densities of ∼5×10¹⁸ cm⁻². If we adopt the median FWHM of 19 km s⁻¹, we see that this corresponds to an average brightness temperature of ∼0.15 K, which is just below the 3σ value of 0.165 K. Based on this, we expect our catalog to be complete to 4σ and to closely approximate the true distribution of sources down to the 3σ level as well. A linear fit for both HVC and AVC distributions gives slopes of −1.3 and −1.1 respectively, with good approximation by a power law. In both cases, our fitted slopes for brightness temperature and column density are shallower than those found by Putman et al. (2002), although closer in agreement to those of Wakker & van Woerden (1991). As suggested by Putman et al. (2002), this is likely due to the fact that the GASS clouds are more similar to the complexes of clouds identified by Wakker & van Woerden (1991), resulting in the dominance of high brightness temperature and column density end of the distribution over the low end which becomes absorbed into the complex structure.

There is a significant spread in area among both AVCs and HVCs, with a few extremely large sources among a majority of sources below ∼1 deg² in size. The slight difference between the size of the clouds in the HVC population and the AVC population may indicate that the AVC population is dominated by angularly resolved sources and thus may be closer to us than those in the HVC population, although this depends considerably on the characteristic physical size of clouds and their likely distance.

We first checked for the containment of each HIPASS source within the data cube of each GASS cloud. We used two possible positions to define the location of a HIPASS HVC: (1) the intensity-weighted position as given in the HIPASS catalog and (2) the position of the actual peak brightness temperature in a GASS-extracted cube of the same HIPASS HVC. We included consideration of the second position because the cataloged brightness temperature and FWHM of each HIPASS HVC are measured through the peak of each source (not the intensity-weighted position). We then crossmatched both the cataloged intensity-weighted position and the position of the maximum pixel against the GASS clouds to check containment. The difference between the containment results for the two positions was two positive velocity sources and three negative velocity sources, so they were in agreement for the majority of sources.

Based on containment of the intensity-weighted and peak positions of the HIPASS sources, our cataloged clouds contain 1193 positive velocity HIPASS HVCs and 646 negative velocity HIPASS HVCs, which leaves 79 sources uncontained within the region of a GASS HVC or AVC. The distribution of uncontained sources is shown on the left (in red) in Figure 14, with a clear limitation due to sensitivity. This method results in a containment level of ∼95%, which is likely to be skewed positively due to the possibility of chance containment of sources that are not necessarily detectable. This is discussed when we crossmatch the uncontained sources against those sources undetected in Section 4.2. Because the HIPASS HVC catalog targeted compact, seemingly isolated clouds (due to the reduced sensitivity to large-scale structure), we might expect that multiple HIPASS clouds may be contained within a single GASS source rather than a one-to-one correlation. We checked this by calculating how many GASS clouds contain a HIPASS HVC, which resulted in 1230 GASS clouds containing a total of 1839 HIPASS HVCs.

Zoom In Zoom Out Reset image size

Another way of approaching the question of the completeness of GASS is to perform a targeted search for all HIPASS HVCs. To do this, we used the same regional definition of a HIPASS HVC as outlined in Section 4.1. From each cube, we extracted two spectra. The first spectrum is taken at the reported HIPASS position, which is the intensity-weighted position of the cloud. However, because the cataloged brightness temperature and FWHM of each HIPASS HVC are measured through the peak of each source, we also extracted a spectrum at the position of the maximum pixel within the cloud.

We used the Kolmogorov–Smirnov (K-S) two-sample test (Smirnov 1944; Bol'Shev 1963) to determine the detectability of a HIPASS HVC, with incorporated matched filtering as outlined in Jurek (2012). The K-S two-sample test allows comparison of representative samples from two distributions and tests the null hypothesis that the samples were drawn from the same distribution. It returns a p-value which gives the probability of the null hypothesis being true, and as such the lower the p-value, the more different the two samples may be.

We first extracted a section of each of the two spectra of a total width that is six times the FWHM of the source (to ensure that we contained the source as well as local noise on either side), and then run a blind matched filtering K-S test on this spectrum. We started with a window of five channels wide and scanned across the extracted spectrum, comparing the window with the rest of the spectrum via the K-S two-sample test. We stored the position of the center of the window as well as the window width only if the returned p-value was lower than any previous scan through the spectrum at a different window width. This then guarantees the lowest p-value for each position, with corresponding window width. The window width was widened until a maximum width of half the extracted spectrum. We selected the position with the lowest p-value (hence the greatest difference to the rest of the spectrum) as representative of the most probable location and width of any source within that spectrum. Due to the presence of other emission in the same spectrum, we adjusted the extracted spectrum width from the default 6 × FWHM for 87 sources.

A source was considered to be detected if the position in either spectrum of the lowest p-value (as found by the matched filtering process) fell within 1.5 times the HIPASS spectral FWHM, or within the inner 75% of the HVC in the spectral dimension. For the 1918 sources searched we detect a total of 1864 clouds, which corresponds to a detection level of just over 97%. The undetected clouds are shown on the right (in red) in Figure 14, with again the key limitation being sensitivity. When comparing these results to the results obtained in Section 4.1, we find that we have slightly higher completeness using this method, which makes sense in the context of being able to detect a source given a known position versus ability to source-find blindly on data.

When we crossmatch the undetected clouds against the uncontained clouds, we find only 6 clouds overlap in both samples. This result suggests that the HIPASS sources that are uncontained by GASS HVCs are not uncontained because of undetectability. The fact that they are detected by the K-S test and yet do not have their intensity-weighted or peak positions contained by any GASS HVCs indicates that their peaks are still too low compared with our source finding criteria, but their signatures are still present in the data however faintly. This is not entirely surprising given that HIPASS is more sensitive than GASS. Conversely, some undetected HIPASS clouds are likely to be contained by chance coincidence within GASS HVCs which is why we see some undetected clouds that have been classified as contained.

In the context of the distinction between which HIPASS sources are successfully "found" within the GASS HVCs versus those which are detectable when we specifically target their region, we note that the K-S test method of searching is not an unbiased, blind search for sources. For this case, we targeted known sources with known positions, and as such are source finding with prior knowledge that there should be a signal present. This prior knowledge is key to why we detect more sources that we contain, because it gives us the ability to differentiate between regions of increased noise, artifacts, interference and actual source signal. Because of this we do not expect to have actually found all of the HIPASS sources (even those that are detectable) with our blind source finder. It is unlikely that a different source finder would do significantly better because source finding is limited by differentiation between other aspects of the data that look like source signal but are in fact spurious.

4.2.1. Summary of Completeness and Reliability