Earlier studies such as Potter and Chiew (2011) suggested the seasonality of rainfall as being one causative factor for lower Millennium Drought streamflows. As seen in Fig. 3a, precipitation reduced across the seasons, and to different degrees in different areas, but reductions in autumn (March–May) were the most severe in many areas. There were also reductions in winter rainfall (see also Pepler et al., 2021), although the percentage of reductions was not as great as those in autumn.

In this study area, autumn is the time when many catchments wet up ahead of the main flow-generating period in winter (June–August) and early spring. In terms of hydrological processes, tightly bound soil water, once seasonally depleted, likely receives priority in the seasonal re-wetting process (e.g. Brooks et al., 2010), since the mobility of water moving to the stream is restricted until after the soil pores have been replenished and the soil reaches field capacity. During the Millennium Drought, lower autumn rainfall may have extended the soil-pore-filling period later into the year, diminishing the period of flow generation, a concept supported by Wasko et al. (2020). The hypothesised process explanation (HPE) is thus summarised as follows: HPE01: streamflow was lower than expected because lower autumn rainfall led to delayed wetting up and muted flow response during winter (when runoff coefficients are usually highest).

As shown in Fig. 3b, during the Millennium Drought there was an increased proportion of time spent in spells with little or no rainfall in many locations. Although HPE01 and HPE02 are similar, HPE01 is focussed on the seasonal dynamics of refilling the soil moisture deficit accrued over the entire dry season, whereas HPE02 is concerned with shorter timescales and occurred in all seasons, albeit to different degrees (Fig. 3b). With longer periods between significant rainfall events, soil moisture deficits likely arose from continued evaporation and/or transpiration, which then needed to be filled prior to the resumption of streamflow (similar to HPE01). Because of this, a greater proportion of event rainfall would be used to fill this deficit rather than converted to streamflow. Although Verdon-Kidd and Kiem (2009) note that the increase in dry spells was not as high as during two other protracted droughts pre-1950, it is retained as a hypothesis for now as HPE02: streamflow was lower than expected because more time spent in spells without rain led to higher soil moisture deficits between rainfall events and thus less conversion to streamflow.

Generally, a small number of large rainfall events each year contribute disproportionately to annual flow volumes, and there is recent evidence for a reduction in the frequency of large rainfall events in many parts of Australia (Wasko et al., 2019; Dey et al., 2020). Heavier rainfall events are more conducive to preferential flow, but smaller or less intense events may be more readily absorbed by the soil and thus available to plants (e.g. Elliott et al., 2015). A reduction in the frequency (Fig. 3c) of high rainfall events (i.e. those above thresholds that produce runoff

Ideally Fig. 3c would plot the data against catchment-specific thresholds of runoff production, but this varies temporally and is difficult to quantify. Thus, a simple threshold is used instead, as marked in Fig. 3c.

Stephens et al. (2018) reported recent increases in pan evaporation within the area affected by the Millennium Drought (despite earlier declines), while Potter and Chiew (2011) cited higher evaporative demand as a potential driver of hydrological behaviour. Although southeastern Australia is typically water limited rather than energy limited, anomalies in evaporative demand may still be important seasonally and in wetter parts of the landscape. Increases in potential evapotranspiration (PET) when water is available (e.g. winter and spring are the wettest times in the southern part of the affected area) are more likely to impact the water balance. Figure 3d shows that evaporative demand – here quantified as Morton's (1986) wet environment evaporation – increased significantly at precisely this time of year and more so in the drier south and south west of the study area (in central and western Victoria; cf. Fig. 2). Conspicuously, this is also an area where almost all catchments exhibit hydrological shifts (Fig. 2c–d). The question of which environmental variable drives this increase is held over to Sect. 4 (see also HPE07). In summary, HPE04: streamflow was lower than expected because evaporative demand increased at a time of year when catchments are relatively less water limited and relatively more energy limited.

The Millennium Drought lasted 10, 13, or 16 years, depending on the location and choice of definition. By comparison, many earlier droughts, such as those in 1965–1968 and 1982–1983, were shorter. It is possible that the unexpectedly severe streamflow reductions during the Millennium Drought were caused by the activation of long-term or long-memory processes that were absent from these earlier droughts. The pre-1950 period included relatively longer droughts such as the Federation Drought (1895–1903) and the World War II Drought (1939–1945), and while comparison of meteorological conditions among these droughts is insightful (e.g. Verdon-Kidd and Kiem, 2009), few streamflow gauges were active during these periods (see also Sect. 4.1), which makes intercomparison of flow difficult. Thus HPE05: streamflow was lower than expected because expectations were conditioned by shorter droughts for which long-term or long-memory process responses were less apparent. This broad, overarching HPE is incomplete without further detail about which processes are in view, but this is held off to later HPEs (specifically HPEs 9, 15, 16, 17, and 18).

The climate drivers presented above – namely PET, rainfall seasonality, dry spells, and intensity – all relate to the Millennium Drought period and not post-drought. As many catchments did not recover post-drought, a test of these HPEs is whether the stated climatic driver continued post-drought or went back to pre-drought conditions. For example, did the changes in seasonality cease with the drought or continue post-drought? We leave such questions to Sect. 4 as part of the evaluation of the HPEs.

Atmospheric CO2 concentrations have been increasing over the instrumental period (IPCC, 2021), with their effect on vegetation water use and impact on catchment or regional water resources being the subject of considerable research (e.g. Peel et al., 2009; Cheng et al., 2017). Higher atmospheric CO2 concentrations generally increase plant water use efficiency (e.g. Ainsworth and Long, 2005; Gedney et al., 2006; Liu et al., 2020), which may manifest as a reduction in transpiration (e.g. Warren et al., 2011). However, the increase in water use efficiency can also manifest as increased growth, and in some cases, this can lead to increased transpiration (e.g. Piao et al., 2007; Uddling et al., 2008; Jaramillo et al., 2018). Observational studies from eastern Australia supporting the latter view include Ukkola et al. (2016), Ajami et al. (2017), and Trancoso et al. (2017). This suggests a potential cause for the reduced streamflow, but it is noted that responses may depend on land cover, e.g. forests and grasslands (Morgan et al., 2004, 2011; see next subsection), and may also vary under drought conditions (Yang et al., 2016). It should also be noted that, because CO2 fertilisation typically leads to a reduction in transpiration per unit leaf area, increased growth can often occur without any increase in transpiration (De Kauwe et al., 2021; Rifai et al., 2022). HPE06: streamflow was lower than expected because fertilisation by higher CO2 concentrations caused higher vegetation water use (relative to previous droughts).

This HPE spans vegetation behaviour, climatic forcing (HPE04), and atmospheric physics but is listed under vegetation because it is the characteristics of the vegetation itself which are key to the process and lead to spatial differences in this effect. Atmospheric coupling refers to the degree of mixing of the air between the canopy and the atmosphere above, which is related to the turbulent flow over and within the canopy (McNaughton and Jarvis, 1991). Small-stature canopies (e.g. grasslands) are more decoupled due to lower turbulence associated with homogeneous canopies (e.g. De Kauwe et al., 2017). This lower turbulence in grasslands allows a near-surface boundary layer to build up, which decouples the vapour pressure deficit (VPD) experienced by plants from the atmospheric VPD. Thus, transpiration in grasslands is relatively less responsive to atmospheric VPD and relatively more responsive to other factors such as solar radiation and temperature. Thus HPE07: streamflow was lower than expected because higher radiation and/or air temperature increased transpiration, particularly in grassland catchments, due to differences in atmospheric coupling between grasslands and forests. Discussion of whether increases in radiation actually occurred (and with a spatial distribution supporting this hypothesis) is held over to Sect. 4. Note that, in contrast to grasslands, transpiration from forests in southeastern Australia can be very sensitive to atmospheric VPD, and this is discussed further in the next subsection.

Under low moisture, plants may become water stressed, and leaves (or, more rarely, the whole plant) may die. In grasslands with pronounced dry seasons, loss of leaves due to desiccation may occur seasonally, leading to reduced water use, but the root system typically remains alive despite loss of foliage. Higher CO2 allows more efficient use of available water in grasslands (e.g. Hovenden et al., 2019) and may delay seasonal desiccation in C4 grasses under warmer conditions (Morgan et al., 2004; see also Morgan et al., 2011). Paradoxically, grasses that retain foliage for longer may transpire rainfall from mid-to-late summer storms (and/or cause greater evaporation via interception), thereby using more water than grass where leaves are already dead. This is a paradox because the higher water use efficiency, which enables the longevity of the C4 grass leaves, could actually result in greater annual water use. HPE08: streamflow was lower than expected because higher CO2 concentrations allowed C4 grasslands to delay seasonal desiccation and transpire water for longer (relative to previous droughts).

If vegetation dies entirely (e.g. Allen and Breshears, 1998; Fensham et al., 2005; Semple et al., 2010), then this could increase water use if the vegetation is replaced by species that are more opportunistic in their water use. HPE09: streamflow was lower than expected because vegetation mortality led to species replacement with plant types that are more opportunistic and thus use more water.

The next HPE is one of three (namely HPEs 10, 11, and 12) which are phrased differently. As noted earlier, forested catchments rarely shifted, whereas grasslands commonly shifted. The most straightforward way to explain this spatial pattern is to posit a hypothesis or hypotheses that act in grasslands catchments and not forested catchments. A less straightforward, but still possible, explanation is to posit that the mechanism causing the shifts acted on all catchments regardless of land cover, and then a second mechanism, specific to forested catchments, counteracted the first to produce the observed spatial pattern. HPEs 10, 11, and 12 are in this latter category, and this is why the HPE wording is preceded by the words “assuming some other mechanism for lower streamflow …”.

As noted, transpiration from forests can be very sensitive to atmospheric VPD (e.g. Gharun et al., 2013), with plants exercising strong stomatal control to limit moisture loss during periods of high VPD and low soil moisture (e.g. Renchon et al., 2018; Grossiord et al., 2020). Thus, if some other mechanism was acting to cause hydrological shifts across grasslands and forests alike, then stomatal control in forests may have counteracted this mechanism (in which case forest catchments would appear less shifted than grasslands). HPE10: assuming some other mechanism for lower streamflow, stomatal control by trees (responding to high VPD) made extra water available for streamflow, counteracting this mechanism in forests. Discussion of whether increases in VPD actually occurred (and with a spatial distribution supporting this hypothesis) is held over to Sect. 4.

Trees demonstrate additional water-saving strategies (e.g. McDowell et al., 2008) not related to stomatal control, with eucalypts being able to reduce leaf area (Nolan et al., 2021) or undergo xylem cavitation (sometimes temporarily), thus impairing water use (Blackman et al., 2019; Nolan et al., 2021). These are listed separately to the previous HPE because they may persist for longer than rapid stomatal dynamics and be less dependent on VPD. HPE11: assuming some other mechanism for lower streamflow, water use controls such as xylem cavitation or reduced leaf area made extra water available for streamflow, counteracting this mechanism in forests.

Wildfire recovery is known to affect water use, with impacts subject to competing factors. Tree mortality and morbidity reduce water use, whereas re-growth increases it, and this is species specific (Heath et al., 2014). Due to this combination of factors, yield impacts reported by Australian studies vary widely and include significant increases (Lane et al., 2006; Guo et al., 2021), decreases in the short–medium term (Nolan et al., 2015) and longer term (Kuczera, 1987), and negligible change in catchment yield (Feikema et al., 2013; Heath et al., 2014). Many eucalypts have post-fire resprouting mechanisms (e.g. epicormic buds and lignotubers) which can be activated, provided the fire does not kill the tree (e.g. Nolan et al., 2014). Thus, fire severity is an important factor, as is species mix (Nolan et al., 2015). A further consideration is changes to soil hydraulic properties and water repellency following fire (e.g. Nyman et al., 2010). Southeastern Australia was subject to multiple fire events during the Millennium Drought. For example, the 2003 bushfires were the first widespread fire events since the early 1980s, and subsequent major fires occurred in the summers of 2006–2007 and 2008–2009 and over different areas in each case. In general, the areas affected by wildfire were those that did not shift. The hypothesis is as follows: assuming a dip in vegetation water use after the wildfire (e.g. Lane et al., 2006; Guo et al., 2021), possibly coupled with runoff-induced changes in soil properties (e.g. Nyman et al., 2010), the wildfire caused an increase in streamflow that masked the underlying shift in behaviour. In other words, if not for wildfire, more catchments would have appeared to have shifted. HPE12: assuming some other mechanism for lower streamflow, the occurrence of forest wildfires made extra water available for streamflow, counteracting this mechanism in forests.

This HPE spans vegetation and groundwater processes. Prior to the Millennium Drought, multi-decadal wet conditions combined with historic deforestation led to high water tables and waterlogging in many areas, with naturally occurring soil salt brought to the surface (e.g. Cartwright et al., 2004). The waterlogged and saline topsoil reduced the health, and thus water use, of vegetation (Lambers, 2003). Waterlogged and salt-affected areas were planted with tolerant grasses or shrubs, while trees were added to the landscape in groundwater recharge areas or upslope of discharge areas (Schofield, 1992; Marcar, 2016). These revegetation efforts were focussed on the late 1980s and 1990s, and thus the timing of maturation of much of this vegetation would have coincided with the onset of the Millennium Drought (commencing in the late 1990s). Furthermore, the drought onset lowered water tables, removing waterlogging as a stressor of vegetation (although it is noted that legacy salinity may persist for many years and even intensify during drought due to evapoconcentration; see, e.g., Hrachowitz et al., 2015). HPE13: streamflow was lower than expected because vegetation recovered from prior waterlogging and because of the maturation of vegetation planted earlier to combat salinity.

The drier conditions during the Millennium Drought, together with potential greening, may have led to a greater accumulated dead plant matter in grasslands. Litter in grasslands has been shown to increase interception of precipitation (e.g. Naeth et al., 1991). HPE14: streamflow was lower than expected because accumulated decomposing litter in grasslands intercepted more precipitation, causing higher AET.

Peterson et al. (2021) suggest vegetation behaviour may explain the apparent appearance of hydrological shifts. Furthermore, they suggest that their findings (as summarised in Sect. 2) are “consistent with … watersheds having multiple stable states [of behaviour] and a finite resilience” (Peterson et al., 2021). In this view, the hydrological shift towards a lower rainfall–runoff relationship corresponds to a transition from one stable state to another (see Peterson et al., 2009). It is noted that the word “state” here has a different meaning to its common hydrological usage (e.g. state variable), as it refers to the behaviour of a system being organised into discrete preferential states (e.g. the wet and dry states of D'Odorico and Porporato, 2004), with intermediate conditions having low probability of occurrence. The Millennium Drought was an extreme disturbance that might have pushed several catchments from a wetter to a drier state, with the drier state otherwise seldom being apparent in a catchment's behaviour. Regarding vegetation, Peterson et al. (2021) say that “Evidence suggests that the vegetation responded to the drought by increasing the fraction of precipitation going to transpiration. … Like other natural systems with multiple stable states, this persistence may be caused by a biophysical adaptation to disturbances … that results in a positive feedback” (Peterson et al., 2021:4–5). The exact nature of this biophysical adaptation is yet to be determined and could be related to one or more of the preceding HPEs. However, it should be noted that none of the preceding HPEs inherently gives rise to multiple stable states, and thus such a hypothesis must be stated separately, which is why it is given its own HPE here. In any case, given that AET is the largest element (after rainfall) of the water balance in southeastern Australia, it seems reasonable to explain the hydrological shifts by hypothesising a direct change to the dynamics of AET by some cause. If it is accepted that the vegetation is the primary cause, then this explains both the streamflow shifts (since less water is available for runoff) and the water table decline described in the next subsection (since a higher fraction of drought rainfall is transpired, which dries out the soil, reducing recharge). HPE15: streamflow was lower than expected because biophysical adaptation, possibly associated with multiple stable states of catchment behaviour, caused an increase in the fraction of precipitation going to transpiration.

Multi-year water table declines in response to the Millennium Drought imply that the depleting fluxes of the groundwater water balance (see Fig. 4b) are less sensitive to the multi-year dry conditions than additive fluxes. This HPE explains the depletion of groundwater by sustained underground drainage from hillslopes to local low points in the landscape (e.g. base of slope and valley bottoms). These low points often contain deposits of alluvium (river deposits) and/or colluvium (unconsolidated sediments deposited at the base of hillslopes by gravity). Separate treatment of hillslope and alluvial aquifers has precedent in local groundwater–surface water studies such as the salinity-focussed rainfall–runoff modelling by Stenson et al. (2011), in addition to modelling in Africa undertaken by Hulsman et al. (2021a), albeit at a larger scale. According to this HPE, the alluvial/colluvial deposits receive the drained water, and then evaporation and transpiration within these deposits keep the water table below the surface and ensure that much of this drainage does not contribute to streamflow. This drainage is hypothesised to occur continuously throughout the pre-drought, drought, and post-drought periods. During pre-drought, the drainage was balanced by higher recharge, meaning that the water table in hillslopes was closer to the surface. This earlier shallow groundwater facilitated runoff processes in hillslopes that were subsequently diminished when the water table dropped, contributing to the hydrological shifts. During the drought, recharge declined, while, according to this HPE, the drainage was assumed to be more sustained, resulting in lower groundwater heads particularly in hillslopes. The plausibility of this HPE assumes that lateral aquifer transmissivity does not decline so much with lower head as to shut off the lateral drainage from the hillslopes. Thus, this HPE depends on the presence of (i) geological strata that support the sustained drainage (e.g. deeply fractured rock; Rempe and Dietrich, 2018) and (ii) colluvial/alluvial deposits (flat valley bottoms; not erosional/v-shaped). HPE16: streamflow was lower than expected because water slowly drained from hillslopes (contributing to water table decline and reduced GW–SW interaction) to nearby alluvial areas and was subsequently lost via evapotranspiration.

Whereas the previous HPE assumed drainage within a catchment, we can also hypothesise drainage crossing boundaries of surface water catchments as part of a regional flow system (e.g. Bouaziz et al., 2018; Condon et al., 2020; Hulsman et al., 2021b; Fowler et al., 2020a) and flowing inland or, alternatively, discharging to the ocean. Such regional systems are also known as mountain block recharge but in this case are associated with hills rather than mountains. Where such systems exist, it is likely that shallower and deeper groundwater flow paths co-exist together (e.g. Frisbee et al., 2012; Condon et al., 2020), associated with local and intermediate/regional groundwater systems, respectively. According to this HPE, regional drainage is hypothesised to occur continuously throughout the pre-drought, drought, and post-drought periods, with the drought bringing changes in the balance with other elements of the water balance in Fig. 4b. HPE17: streamflow was lower than expected because sustained drainage of groundwater across catchment boundaries contributed to declines in the water table and reduced GW–SW interaction. We note that, in the context of headwater catchments, the distinction between local and regional drainage is important because (i) regional discharge cannot support within-catchment wetting up, which would otherwise amplify event runoff, and (ii) at the point of discharge, groundwater discharge may be an important control on AET, so the distinction affects the appearance of a headwater catchment in remotely sensed data. Both (i) and (ii) affect how the catchment might be represented in simulation models.

A further distinction is that groundwater systems vary in their response times to disturbance. In southeastern Australia, groundwater systems which cover larger areas tend to have longer response times (Walker et al., 2003). Locations that are influenced by a combination of two or more systems may be subject to a mixture of dominant timescales – say, a 1–5-year timescale associated with a local system mixed with a 10–30-year timescale associated with a larger system. Likewise, drought recovery may be a question of the relative influence of shorter and longer timescale groundwater systems.

Even if rainfall–runoff relationships remained unchanged (i.e. no shift) within lower-order catchments, flow within higher-order stream reaches can be lost as it travels downstream, via seepage into the surrounding substrate. This is called a losing stream reach, in contrast to a gaining reach in which the stream receives water. The losing/gaining dynamic is driven by the stream level relative to the groundwater level in the surrounding substrate. This dynamic may be important in cases where the river flows over extensive alluvial deposits, since (i) these deposits can store significant amounts of water, and (ii) they commonly have relatively high hydraulic conductivity, which may assist in dispersing the water horizontally over the extent of the deposit, creating opportunities for transpiration or evaporation (e.g. Salama et al., 1993). Reaches can change from gaining to losing state, and vice versa (often associated with seasonal cycles), and they experience a net gain or loss while in a given state – a key point of difference with bank storage, which functions as a delaying process (instead of a source/sink term) as water flows into and then back out of stream banks as gradients switch back and forth episodically. Losing/gaining dynamics could have contributed to the observed shifts at stream gauges. The regional groundwater declines could have affected the spatiotemporal extent of losing reaches (i.e. greater total length losing; greater time spent losing rather than gaining), amplified water loss rates from streams, and increased the number of zero flow days. HPE18: streamflow was lower than expected because stream reaches lost more water (or gained less water) due to water table decline. Note that, while losses are sensitive to the water table, the reverse is also true, with losses from the stream causing the water table to decline less slowly than it would otherwise (subject to other fluxes such as evaporation and transpiration from the alluvium). The strength of this feedback/coupling increases for smaller alluvial stores.

Regardless of the cause of the water table decline, the result was a thicker vadose (unsaturated) zone as the groundwater subsided, leaving behind newly unsaturated layers. This HPE is concerned with the rate at which these layers drain from saturation to field capacity. Well-drained near-surface soils are generally considered to drain quickly, of the order of days (e.g. Cassel and Nielsen, 1986). However, drainage may be slower for deeper layers (particularly those composed primarily of bedrock), possibly associated with spatially discontinuous preferential flow mechanisms including funnelled flow (e.g. Nimmo, 2021) and unstable flow (e.g. Jury et al., 2003). These deeper unsaturated layers might temporarily absorb some recharge, delaying its passage downwards to saturated zones and spreading the recharge signal over time. Whereas a rapid recharge event might trigger rapid groundwater discharge to the stream, the dampening of the recharge signal might increase the probability of water being diverted to transpiration (particularly by vegetation close to water courses and drainage lines) that would have otherwise contributed to streamflow, thus increasing transpiration per unit precipitation. HPE19: streamflow was lower than expected because delaying of recharge by the enlarged unsaturated zone increased opportunities for transpiration.

Soils may change properties with time, and some drought-related changes may decrease streamflow. While many soil properties and temporal changes are driven by humans (discussed below in HPE21), there are other mechanisms that are not explicitly human-driven, although they may interact with land management. For example, soil cracking may lead to increased infiltration, more evaporation, and less surface runoff (e.g. Arnold et al., 2005), provided the infiltrated water remains in the soil rather than recharging groundwater. There is also evidence that soil porosity may increase, as the climate becomes drier, on relatively short timescales of years to decades (Hirmas et al., 2018; Caplan et al., 2019). HPE20: streamflow was lower than expected because drought-induced changes in soil properties such as higher porosity and/or cracking led to higher infiltration and evaporation. On the other hand, soil water repellency can build up during dry periods (e.g. Filipović, 2018), which would tend to increase streamflow, which is counter to observations; likewise, soil cracking may provide an avenue for preferential flow, thus increasing runoff (e.g. Beven and Germann, 1982).

The last 50 years have seen changes in land management practices, and the hydrological impacts of these changes may be important. Soil loss around Australia in the 1960s and 1970s prompted a nationwide evaluation of land degradation and soil conservation strategies in the 1970s, in addition to a National Soil Conservation Program, established in 1983 (Hannam, 2003). As a result, soil conservation measures such as no-till farming (e.g. Cornish et al., 2020) were more prevalent by the 1990s than in previous decades, which may have caused hydrological differences (see e.g. Huang and Zhang, 2004) between the Millennium Drought and previous droughts. In general, poor management practices lead to degraded soils that may amplify event runoff, with partial soil recovery (towards, e.g., higher soil carbon) possibly causing trends towards higher infiltration. Techniques such as contour banks may also retain water in the landscape. Last, soil properties were impacted by trends in livestock stocking rates driven by both market forces and government policy. HPE21: streamflow was lower than expected because better management of agricultural land, including soil conservation techniques, caused differences in soil hydraulic properties (compared to previous droughts).

Land use change is commonly reported to alter hydrology (e.g. Zhang et al., 2001; Brown et al., 2005; Zhang and Schilling, 2006; Farley et al., 2005). In southeastern Australia, a key example is new plantation forestry, with the higher water demand of plantation trees impacting groundwater (Benyon and Doody, 2004; Lane et al., 2005; Webb and Kathuria, 2012; Brown et al., 2015; Dresel et al., 2018) and presumably streamflow. Plantation data (Legg et al., 2021) show a peak in nationwide plantations in the late 1990s, with approximately 100 000 ha yr-1 added for 1998–2000 and much of it within the area impacted by the Millennium Drought (Legg et al., 2021; Table 1). HPE22: streamflow was lower than expected because increased plantation forestry led to increased water use. The 1970s–1990s also saw planting for salinity mitigation (see HPE13).

Farm dams, also known as farm ponds, are small private dams located on hillsides, drainage lines, or small streams. Individual dams typically harvest only small volumes, but collectively, the impact can be large (Habets et al., 2018; Morden et al., 2021). Thus, the total water storage capacity of all dams in a watershed may be an important index, particularly when relativised by precipitation or streamflow to account for aridity; that is, a given storage size may have a higher relative impact in drier areas (e.g. Liu et al., 2014). Volumetrically, water harvesting by farm dams is greatest during moderate and high flows, but if expressed in relative terms, impacts are greatest during the dry season and during dry years (Morden et al., 2021). Flow impacts occur due to both extracted water (for stock, domestic, or irrigation use) and evaporated water being unavailable to flow downstream. Farm dam development has increased with time (with a higher rate of increase before 2000; Malerba et al., 2021), so there were more dams during the Millennium Drought than previous droughts. HPE23: streamflow was lower than expected because farm dam harvesting of water increased due to trends in farm dam development (i.e. more dams than prior droughts).

Regionally, the bulk of agricultural water use occurs in broad, flat plains using water captured in large reservoirs. This activity occurs downstream of the headwater catchments in focus here. However, minor modes of extraction are relevant in the headwater catchments, including extractions from groundwater and diversions from streams by individual landholders. For the set of catchments used by Peterson et al. (2021) and shown in Fig. 2, individual diversions from streams are uncommon because they were considered to be a criterion in catchment selection, and where they do exist, they are strictly controlled. Extractions from groundwater can be relatively larger than those from streams within some of these catchments and are of interest, given the strong groundwater trends reported above. Most extraction bores are for stock and domestic use. However, licenses for irrigation, while rarer, typically correspond to larger volumes and thus may be significant. HPE24: streamflow was lower than expected because groundwater extraction caused declines in groundwater levels and reduced GW–SW interaction.