Queensland, the second largest state in Australia, covers a surface area of 1.85 million km2 and has a population of 4.75 million people. Land use across the state is dominated by agriculture with over 80 % of the total surface area utilised for grazing cattle or irrigated cropping (Fig. 2a; QLUMP, 2018). The Queensland agriculture sector contributes more than AUD 13 billion per year to the state economy and includes 15 million cattle and sheep as well as 4526 km2 of land under irrigation (ABS, 2018). The climate is subtropical or tropical with mean annual temperatures ranging from 27.5 ∘C in the state's north to 15.8 ∘C in the southern interior. There are large gradients in rainfall across the state ranging from a mean annual rainfall of over 3000 mm in the coastal north-east to less than 100 mm in the arid western regions (Fig. 2b). Rainfall has a distinct annual pattern with up to 80 % falling during the summer months from November to April and is subject to decadal drought and flood cycles (Klingaman et al., 2013). The economic importance of agriculture coupled with the need to provide a year-round water supply for these activities and the lack of predictable rainfall has resulted in the proliferation of artificial water bodies across the state (Fig. A1 in the Appendix). However, the number and surface area of ponds in Queensland is relatively unknown as there is no legal requirement to refer ponds to the state registry due to their small size. Under current state law only dam walls in excess of 10 m and volumes above 750 ML (megalitres) are referable (DEWS, 2017) and the maximum reported volume for ponds in Queensland is 3 times less than the referable volume (< 250 ML) (SKM, 2012). This study has assumed ponds are less than 100 000 m2 as this is recognised globally as the major area of uncertainty in surface area assessments (Lehner and Döll, 2004; Downing, 2010) and has been identified as a threshold in global lake inventories (Downing et al., 2006; Verpoorter et al., 2014).

To determine the number and relative surface area of ponds across Queensland, three state government GIS databases of artificial water bodies were utilised. However, these databases required additional processing to extract comparable pond data as there were inconsistencies in the format and nomenclature of feature types. The primary database used was the most recent official assessment of land use from March 2018 (QLUMP, 2018), and within the primary land use classification of “Water” there is a secondary category of artificial “Reservoirs/dam” divided further into “Reservoirs, Water storage and Evaporation basin”. The individual water body surface area is provided and all ponds (< 105 m2) were extracted from the database. Evaporation basins were excluded, as these are commonly used for salt extraction. These ponds were then compared against two state government databases from a high-resolution assessment of artificial water bodies across the state published in 2014 and 2015. Both databases are derived from aerial (10 to 60 cm orthophotography) and satellite (0.5 to 2.5 m resolution) imagery captured between 2010 and 2014. One database contains water bodies greater than 625 m2 at full supply (Reservoirs – Queensland; http://qldspatial.information.qld.gov.au/catalogue/, last access: 28 November 2017) and for water bodies less than 625 m2 a second database was used (Water Storage Points – Queensland; http://qldspatial.information.qld.gov.au/catalogue/, last access: 28 November 2017).

Water bodies larger than 625 m2 contained individual polygons where water body surface area was provided and all water bodies less than 105 m2 were extracted from the database. The database of water bodies smaller than 625 m2 contained point data providing only the location of water bodies and no information on their dimensions (Fig. A1b, c). To estimate the surface area of these systems, 100 water bodies were randomly selected using the Subset Features tool in the Geostatistical Analyst toolbox in ArcGIS (Version 10.3, ESRI Inc., Redlands, California, USA). The surface area of selected water bodies was then quantified using high-resolution aerial imagery (Nearmap; https://www.nearmap.com.au/, last access: 15 May 2018). Typical pixel resolution of 7 cm greatly improves edge detection of ponds as it can be very challenging to separate the water edge from riparian vegetation stands with coarser-scale data. Pond edges were mapped following the methodology of Albert et al. (2016) where imagery was georeferenced and the water edge was manually traced to create individual polygons for each pond. The mean surface area of all 100 polygons was then assumed to approximate the surface area of all individual ponds within this database and the total surface area was calculated by multiplying this mean surface area by the total number of ponds.

To ensure only one water body was reported from each location, all databases were first screened to remove repeated detections of water bodies. All remaining water bodies were then summed together to calculate the total surface area of ponds and this was compared to larger reservoirs to determine their relative surface area. To undertake regional scaling of pond emissions, individual ponds were sorted using two different size class classifications. Firstly, we categorised sites into the three smallest size classes (102 to 103; 103 to 104; and 104 to 105 m2) in the Global Reservoir and Dam (GRanD) assessment (Lehner et al., 2011). Secondly, we divided sites into water bodies less than 3500 m2 (primarily stock dams) and larger water bodies (primarily irrigation dams and urban lakes), following the findings of Lowe et al. (2005) and SKM (2012).

To quantify the range of emission rates from ponds, a monitoring program was undertaken from August to December 2017 across a wide spectrum of ponds including: farm dams (irrigation and stock watering), urban lakes, small weir systems (i.e. small dams leading to widening and slowing of river flows) and rural residential water supplies (Fig. 1). Stock dams, irrigation dams and urban lakes account for the vast majority of ponds across Queensland and ponds within each category were selected to represent the regional size class distribution (Fig. A2). The majority of sites were located in coastal catchments in south-east Queensland, Australia, as well as one urban lake and three stock dams in central Queensland (Fig. 2c).

There are a number of commonly used methods to assess methane emissions from water bodies depending on the pathway of interest. For the diffusive emission pathway, rates may be modelled using the thin boundary methods or directly measured using manual or automatic floating chambers (St. Louis et al., 2000). For ebullition pathways, rates can be directly measured using acoustic surveys or funnel traps (DelSontro et al., 2011). Thin boundary layer models cannot be used to quantify the ebullition pathway and acoustic surveys or funnel traps cannot be used effectively in ponds as the water depth is often too shallow (< 1 m). We chose to use floating chambers to capture both ebullition and diffusive fluxes. CH4 emission rates were measured by deploying between 3 and 16 floating chambers per water body, covering both peripheral and central zones (Fig. A3). Chamber design followed the recommendations of Bastviken et al. (2015), as these lightweight chambers (diameter 40 cm, 12 L headspace volume and 0.7 kg total weight) were ideally suited to deployment in ponds where both site access and on-water deployments can be challenging (Fig. A4). The floating chambers used were designed to yield negligible bias on the gas exchange and compare well with non-invasive approaches (Cole et al., 2010; Gålfalk et al., 2013; Lorke et al., 2015).

Where possible, 24 h measurements were undertaken; however, in three water bodies this was not possible (Appendix Table A1) and here measurements lasted between 6 and 8 h. The 24 h deployment time was chosen to increase the likelihood of capturing ebullition, which is episodic in nature, and of incorporating diel variability in diffusive emissions which can be up to a 2-fold bias (Bastviken et al., 2004, 2010; Natchimuthu et al., 2014). The use of long-term deployments may underestimate diffusive fluxes, which decrease as the chamber headspace approaches equilibrium with the water. However, in contrast to CO2, CH4 has a long equilibration time and it has been shown that a 24 h deployment of these types of flux chambers on lakes underestimate diffusive fluxes by less than 10 % (Bastviken et al., 2010). An initial gas sample was collected at chamber deployment and a final chamber headspace gas sample after 24 h following the Exetainer method described in Sturm et al. (2015). CH4 emission rates were calculated from the change in headspace concentration over time and normalised to areal units (Grinham et al., 2011).

The effect of inundation status on emission rates was tested on a stock dam (Gatton 4) where measurements were undertaken on peripheral areas during periods of inundation and no inundation. This pond was selected as stock dams generally experience accelerated rates of water level change due to their relatively small size compared to other pond types (Fig. A2). In addition, the construction of this pond is typical for stock dams (a shallow pit is dug out and the soil used to construct the wall and spillway) and the surface area (1893 m2) closely matched the median for all farm dams (1586 m2; Fig. A2). Emission measurements for the inundated period followed the methodology outlined above for the water body emissions survey. Three weeks later water levels within the ponds had dropped and emission measurements were repeated at the same sites which were now exposed. For these emission measurements five chambers (90 mm diameter, 150 mm length) were carefully inserted 50 mm into the ground and care was taken to minimise disturbance to the soil surface. The headspace of each chamber was flushed with ambient air to remove headspace contamination due to chamber insertion, then the sampling port of each chamber was sealed. After the deployment period, a gas headspace sample was collected and CH4 concentration was analysed.

Emissions rates and surface area data were analysed using a series of one-way analyses of variance (ANOVAs) with the software program Statistica 13 (Dell Inc., 2016). Analysis of emissions rates collected during the monthly monitoring study and the inundation study used sampling month or inundation status as the categorical predictors and chamber emission rates as the continuous variable. Emission rates from individual water bodies collected during the broad survey were first pooled into four primary use categories (irrigation, stock, urban and weirs) or three different GRanD size classes and these categories were used as the categorical predictors. The primary use of each pond was provided by pond owners or managers; in the case of urban lakes that had both aesthetic and storm water functions these were classified as urban (Table A1). A total of 22 ponds were included in this survey with 4 irrigation ponds, 9 stock watering ponds, 7 urban ponds and 2 weirs. Changes in water surface area (as a percentage of AFSL) from individual water bodies were pooled into three GRanD size classes and these categories used as the categorical predictors. Where necessary, continuous variable data were log transformed to ensure normality of distribution and homogeneity of variance (Levene's test) with post hoc tests performed using Fisher's LSD (least significant difference) test (Zar, 1984). Tests for normality were conducted using the Shapiro–Wilks test as recommended by Ruxton et al. (2015). The non-parametric Kruskal–Wallis (KW) test was used for continuous data which failed to satisfy the assumptions of normality and homogeneity of variance even after transformation. Statistical results were reported as follows: test applied (Fisher's LSD or Kruskal–Wallis test), the test statistic (F or H) value and associated degrees of freedom with p value.

Emissions were scaled to water body size classes following two different approaches. Firstly, emissions were grouped according to their respective GRanD size class. These match the size class of water bodies used in the emissions monitoring of this study, and the GRanD database was used in the most recent global synthesis of greenhouse gas emissions from reservoirs (Deemer et al., 2016). Secondly, water bodies less than 3500 m2 in area were assumed to be primarily stock dams and larger water bodies primarily irrigation dams (Lowe et al., 2005). To extrapolate pond emission rates to regional scales, an appropriate measure of centrality should be used. Three common measures, arithmetic mean, geometric mean and median values, were calculated for each water body category and size class. To assess the most suitable measure of centrality for water body emissions, normal probability plots of raw and log-transformed emissions data were generated and tested using the Shapiro–Wilks test (Fig. A5). The emissions data from all replicate measurements fitted a log-normal (p=0.081) but not a normal distribution (p=0.0000) and, therefore, the geometric mean would provide the most appropriate measure of centrality for this data (Ott, 1994; Limpert et al., 2001). Fluxes were scaled to annual rates using the cumulative surface area of water bodies and the respective emissions rate for each size class using the geometric means. The variability in geometric mean was given by the exponential of the 95 % confidence interval range of log-transformed data. Emissions for water bodies less than 3500 m2 were scaled using stock dam rates and larger water bodies (3500 to 105 m2) using rates obtained from irrigation dams and urban lakes. Total fluxes from respective size classes were then combined to provide regional estimates. Annual fluxes of CH4 were converted to CO2 equivalents assuming a 100-year global warming potential of 34 (IPCC, 2013).

The statewide land use assessment identified 13 046 ponds across Queensland, occupying a total surface area of approximately 467 km2 (Fig. 2c). However, with the inclusion of the additional Reservoir and Water Storage Point datasets the number of ponds increased over 20 times to a total of 293 346, and the surface area more than doubled to 1087 km2. The official land use assessment of Queensland underestimates the surface area of ponds by 57 % and the total number of water bodies by more than 1 order of magnitude. The revised total surface area of all artificial water bodies across Queensland increased by 24 % to just over 3248 km2 (Table A2).

Ponds were widely distributed across the state, but over 78 % of ponds were located on grazing land, suggesting that stock dams represent the primary water body type (Fig. 2a). Over two-thirds of ponds were confined to regions of the state where rainfall isohyets were above 600 mm (Fig. 2b) and heavily concentrated in cropping and residential areas in the central and south-eastern parts of the state (Fig. 2c). These findings highlight the importance of striving to incorporate all artificial water bodies into flooded land emission assessments; omitting water bodies below a size threshold can lead to a dramatic underestimation of the total number of water bodies present and a considerable underestimate of the available surface area for CH4 emissions.

All 22 water bodies monitored in this study were shown to be emitters of CH4, and emission rates ranged from a minimum of 1 mg m-2 d-1 to a maximum of 5425 mg m-2 d-1 (Table A1). Only one water body (Mt Larcom 3) had a maximum rate below the reported upper range (50 mg CH4 m-2 d-1) for diffusive fluxes found in larger water bodies in this region (Grinham et al., 2011; Musenze et al., 2014). Mean flux rates of only four individual water bodies were below 50 mg m-2 d-1 (Table A1), suggesting ebullition to be the dominant emission pathway in these systems.

Grouping ponds according to their primary use resulted in no significant differences in emissions rates between irrigation dams, stock dams and urban lakes; however, weirs were significantly higher (F(3,121)=6.43, p < 0.001) than all other categories (Fig. 3a). Mean emission rates were, however, higher in stock water bodies (168 mg m-2 d-1) compared with irrigation and urban bodies (84 and 129 mg m-2 d-1, respectively). Weir water bodies had mean emission rates of 730 mg m-2 d-1, which is more than 4 times higher than those of any other category (Fig. 3a). Grouping ponds according to their GRanD size classes resulted in significantly higher emission rates (KW H(2,121)=7.354, p < 0.05) from ponds in the 102 to 103 m2 size class compared to 104 to 105 m2 (Fig. 3b). Overall, mean emissions decreased with increasing size class. Note that all weir sites fell into the smallest size category.

The water surface area of a single stock dam ranged from 395 to 2808 m2 over a 40-month period (Fig. 7a) with an outer band of 580 m2 undergoing frequent inundation cycles (May 2016 to December 2017 – Fig. 7a). Emissions rates from peripheral areas during an inundated period were significantly higher (more than 1 order of magnitude) compared with emissions when not inundated (KW H(1,10)=6.818, p < 0.001; Fig. 7b). In contrast emissions from central areas were over 100 mg m-2 d-1, which is more than double the peripheral area emission rates (Table A1). This modifier of rates will primarily impact emissions from smaller size classes which have greater variability in water surface area (Fig. 6). An additional implication is in the importance of designing monitoring studies where emissions rates are quantified from both peripheral and central areas for each system. Rates monitored only in peripheral areas will likely bias towards lower emissions, particularly if these have undergone recent inundation.

The findings of this study demonstrate ponds are an underreported and important CH4 emission source in Queensland and likely also globally. These findings highlight the importance of striving to incorporate all artificial water bodies into flooded land emission assessments; omitting water bodies below a size threshold can lead to a substantial underestimation of the total number of water bodies present and a considerable underestimate of the available surface area for CH4 emissions. Mean annual CH4 fluxes from ponds for the state of Queensland ranged between 1.7 and 1.9 million t CO2eq. (Table 1) depending on the scaling approach. Given that ponds represent 33.5 % of the total flooded lands surface area in Queensland and emission rates are equivalent to larger water bodies in the region (Musenze et al., 2014; Sturm et al., 2014), ponds represent one-third of total emissions from flooded lands in Queensland. Remarkably, mean total emissions from ponds represent approximately 10 % of Queensland's land use, land use change and forestry sector (NGERS, 2015) emissions using either scaling approach.

Future regional and global emissions estimates would be greatly improved with the inclusion of ponds, as their proliferation has been noted in five continents. In the continental United States ponds have been shown to cover 20 % of the total artificial water body surface area (Smith et al., 2002); in South Africa there are an estimated 500 000 ponds (Mantel et al., 2010); in Czechoslovakia ponds make up over 30 % of the total artificial water body surface area (Vacek, 1983); and in India ponds are estimated to comprise 6238 km2, or over 25 % of India's artificial water body surface area (Panneer Selvam et al., 2014).

Emissions rates from ponds observed in this study are consistent with ebullition being the dominant pathway. Diffusive emissions from studies of three larger water bodies in the region found the upper limit for diffusive emission was 50 mg m-2 d-1 (Grinham et al., 2011; Musenze et al., 2014) and only five ponds had emission rates below this level. Ebullition was observed at all ponds with maximum rates all in excess of 50 mg m-2 d-1, with the exception of only one stock dam (Mt Larcom 3) where the maximum rate was 19 mg m-2 d-1. This is a consistent finding with larger water bodies in the region where ebullition has been shown to dominate total emissions (Grinham et al., 2011; Sturm et al., 2014). The relatively higher emissions from smaller pond size classes is consistent with previous observations of increased ebullition activity in shallow zones, particularly water depths less than 5 m (Keller and Stallard, 1994; Joyce and Jewell, 2003; Sturm et al., 2014). Virtually all ponds within the smaller size classes would be less than 5 m deep. In addition, ponds trap large quantities of sediment and organic material (Neil and Mazari, 1993; Verstraeten and Prosser, 2008) and these deposition zones have been identified as methane ebullition hotspots in larger water bodies (Sobek et al., 2012; Maeck et al., 2013). The patterns in emissions from the intensive spatial study in an urban lake, where shallow areas adjacent storm water inflows were shown to be ebullition hotspots, have also been observed in larger water bodies were ebullition activity was highest adjacent to catchment inflows (DelSontro et al., 2011; Grinham et al., 2017; de Mello et al., 2017). The emissions from small weirs were clearly dominated by ebullition, which is consistent with emissions from three larger weirs where rates ranged from 1000 to over 6000 mg m-2 d-1 (Bednařík et al., 2017). Weirs intercept the primary streamflow pathways and will likely cause large quantities of catchment-derived organic matter to deposit within the weir body which, coupled to the shallow nature, results in high rates of ebullition. Overall, the rates observed for all categories, except irrigation dams, were in the upper range of reservoir areal flux rates reported in global reviews (St. Louis et al., 2000; Bastviken et al., 2011; Deemer et al., 2016), reflecting the dominance of the ebullition pathway in ponds. An additional consideration for future studies of ebullition patterns in ponds stems from recent studies of reservoirs which found significant changes in ebullition intensity and ebullition distribution as water levels decrease (Beaulieu et al., 2018; Hilgert et al., 2019). Under decreasing water levels, deeper zones of ponds may begin bubbling or increase the intensity of bubbling; this could potentially offset the reduction in surface available for emissions and total emissions would remain relatively constant.

Efforts to develop flooded land emission inventories rely heavily on the emission rate used to scale the surface area of water bodies within selected categories. Given the high variability in emission rates within and between individual ponds and relatively low replication, it is critical to select an appropriate measure of centrality (arithmetic mean, geometric mean or median) in order to scale regionally and globally (Downing, 2010). For rice paddies, septic tanks, peatlands and natural waters (Aselmann and Crutzen, 1989; Dise et al., 1993; Diaz-Valbuena et al., 2011; Bridgham et al., 2006), the geometric mean has been applied. Likewise, in this study the log-normal distribution of emissions data indicated the geometric mean as the most appropriate measure and the total emission rates using this measure fell within the reported range from larger artificial water bodies in the region (Grinham et al., 2011; Sturm et al., 2014). However, the geometric means for all water body categories and size classes were less than half of their respective arithmetic mean values (Fig. A6). For irrigation, stock and urban water bodies, geometric mean values were actually outside of 95 % confidence interval limit for the arithmetic mean (Fig. A6a, b). Geometric mean and median values were similar across all water body categories and size classes, and these measures, therefore, represent conservative emissions rates from ponds. This raises an important issue with scaling ebullition-dominated water bodies as there is always going to be a high likelihood of detecting a small number of very high rates which will invariably give rise to log-normal data distributions. Future studies will focus on determining whether the conservative estimates generated through the use of geometric means approximate the true emissions from ponds.