Discussion

Summarising, this study used multiple approaches to confirm that in general streamflow in south-eastern Australia is declining based on the 41 years of data, and that this decline tends to be greater than the associated decline in rainfall (Tables 4 and 6). In other words, there is significant “amplification” of the observed rainfall declines. However, the overall declines, for small forested catchments and considering 41 years of data, is (still) very small in percentage and mm year-1 terms, once variability of rainfall and temperature is accounted for. In addition, using the LTPMK approach, which accounts for the autocorrelation in the data and test for significance of the Hurst coefficient, most of the approaches indicated non-significant trends. This indicates that while there is a clear local trend in the 41 years of some of the data, we cannot (yet) say with confidence whether this represents a long-term declining trend in streamflow.

Comparison to other studies

In comparison, an analysis of Australian streamflow records for the latter half of the 20th Century using a Mann-Kendall test by Chiew and McMahon (1993) found significant declines at 7 of the 30 studied stations. That study points out that due to the high inter-annual variability of Australian streamflow, a greater change might be required to identify a statistically significant trend (Chiew & McMahon, 1993). A similar observation can be made in this study considering the significance under the LTPMK test. Significant larger declines in streamflow have been observed by Petrone et al. (2010), for catchments in southwest Western Australia between 1950-2008. A more recent paper identified significant trends in 20 of 199 catchments, concentrated in the South-East of Australia (Ajami et al., 2017). An analysis of global data (Milly, Dunne, & Vecchia, 2005) similarly indicated decreases in streamflow in Australia, but the database from that study overlaps with the other studies.
In general, decreasing trends in rainfall in the south east of Australia and increasing trends in temperature have been widely identified (Cai & Cowan, 2012; Chowdhury, Beecham, Boland, & Piantadosi, 2015; Murphy & Timbal, 2008). In contrast, positive trends in rainfall were found in the 20th Century in earlier studies (Collins & Della-Marta, 1999; Smith, 2004), although the data period was suggested to be an issue (Smith, 2004). Some other studies have found much larger decreases in annual rainfall (Gallant, Hennesy, & Risbey, 2007; Murphy & Timbal, 2008) in the order of 1 mm per year and greater decreases in streamflow (Petrone et al., 2010). The differences are most likely related to the difference in aggregation time scale (annual anomalies versus weekly data). Another complicating factor could have been that many of these earlier studies had timeseries to 2008, which was at the end of the millennium drought (van Dijk et al., 2013), while our study included the drought breaking year 2010. This “local trend” phenomena can also explain why many of our LTPMK tests showed no significant trends. This highlights once again the need for careful consideration of the trend tests applied to the data.
The low rainfall elasticities calculated from the rainfall runoff simulations (Figure 4) differ from earlier research (Chiew, 2006), which suggested much higher amplifications from their SimHyd modelling. There are however differences with the original work (Chiew, 2006), which could explain the outcomes. Our study simulates the streamflow on a daily scale and scales this to annual data, while the original work calibrated on monthly data and scaled to annual. As larger time scales are smoother and show more significant trends than weekly and daily data (see the monthly and annual Mann Kendall results in the supplementary documents, document 3) this could have also affected the amplification calculations.
There could be differences in the magnitude of the analysed trends in this study compared to other studies, due to the differences in time scale. A reduction in variance occurs when the data is summarised to annual and monthly data (Cai & Cowan, 2008; Petrone et al., 2010). In addition, there could be differences due to the translation of remotely sensed (Ukkola et al., 2015), filtered (Trancoso et al., 2017) or modelled data (Chiew et al., 2009) to streamflow declines. Several other studies investigated mainly the difference between decades or specific periods (Chiew et al., 2014; Potter et al., 2010; Saft et al., 2015).
Furthermore, our work concentrates on 13 single land cover (forested) small catchments, while several older studies did not control for land cover, but mainly selected “unimpaired” catchments (meaning no water resource development). The data based analyses (triangles in Figure 4), and the GAMM modelling supports the magnitude of the amplification (Chiew, 2006) for the few catchments that showed significant trends. However, the non-parametric rainfall elasticities (grey triangles in Figure 4) could be biased (Fu et al., 2007) as this does not take into account the full interaction between rainfall and temperature.

Remaining trends

What the remaining trends in the streamflow (after removing rainfall and temperature effects) exactly represent is still difficult to say. Amplification has been mainly explained as an additional drying effect (Chiew, 2006), mostly driven by the increases in temperature. Given that in most catchments the remaining trend (Table 6) represents up to 80% of the original trend in the streamflow (Table 4), this would correspond to amplifications around 4 (the data suggests 0.98 - 4.5).
There is no indication that streamflow in these forested catchments has decreased by a large amount (24 - 28%) due to increases in evapotranspiration as a result of CO2 fertilisation (Ukkola et al., 2015; van Dijk et al., 2013). Because we selected homogeneous land cover (forestry), the larger declines found in earlier studies (Ukkola et al., 2015; van Dijk et al., 2013) might be due to trends in land management and land cover, such as changes in cropping patterns, introduction of conservation tillage, or terracing and crop variety effects. Furthermore, changes in the vegetation water use efficiencies and changes in the tree physiology (Franks et al., 2013; Ukkola et al., 2015) can also explain part of the remaining trends. Conceptually, if trees become more water use efficient and produce more leaf mass, then this could affect evapotranspiration and streamflow. In addition, larger leaf area could also result in increased interception, specifically for smaller rainfall events (Stoof et al., 2012). This clearly highlights that more work in the area of vegetation effects on streamflow would be useful, given the dominance of transpiration in the global water balance (Jasechko et al., 2013), to support investigations into climate change effects.

Distribution changes

Changes in the timing and distribution of rainfall could impact the amplification of the rainfall signal in the streamflow. Shifts in the main rainfall season from winter towards summer, or changes in the average storm depth and timing between rainfall events can all affect the runoff process (Ishak, Rahman, Westra, Sharma, & Kuczera, 2013; Westra et al., 2012) and shift the balance between surface runoff and groundwater flow. For example, Trancoso et al. (2017) recently found declines of about 1 mm/year focussing on baseflow. We compared streamflow and rainfall distribution curves across the four decades in the data (supplementary data, document 4) but found no obvious or clear patterns. But this analysis was essentially qualitative and a more detailed analysis of changes in the distribution of flows and rainfall is warranted.

Limitations

However, there are also some limitations to the present study. Small catchments might have smaller temperature effects on streamflow (as there is less time for evaporation to work on the catchment storage). However, the advantages of controlling for routing and land cover in smaller catchments were important for this study.
Some of the reasons for the low performance of the GAMM models could be that the weekly data is too coarse to pick up some of the non-linear rainfall responses, or that the AR1 model is not complex enough to model the remaining storage delays in the weekly data despite the small catchments. However, computational difficulties and missing data make working with daily data in GAMM challenging.
The low impact of rainfall and ET on explaining variation in streamflow contrasts the earlier work in this area using mainly linear regression and with annual and monthly data (Cai & Cowan, 2008; Potter & Chiew, 2011). We believe that this difference has to do with the differences in aggregation time scale and the non-linearity of the processes.
The value of using observed data to analyse trends is that the model bias (Figure 5) is not affecting the results. However, finding the right method of analysis is tricky (Fu et al., 2007). We chose to use statistical modelling to remove the need to make the assumptions related to mechanistic rainfall-runoff models. This also removes the need for water balance closure, which can affect the parameter calibration (Vervoort, Miechels, van Ogtrop, & Guillaume, 2014).
This study has concentrated on a relatively small group of Australian catchments as an example, because we controlled for catchment size and land cover. We identified about 5 more catchments that could be included, but these were either very small or would skew the spatial distribution of the catchments. However, future research could be to take advantage of other (global) long term databases of unimpaired and forested catchments, such as the Critical Zone Observatories in the US (Troch et al., 2015) or the well-publicized French catchment database (Oudin, Andréassian, Perrin, Michel, & Le Moine, 2008) and other European catchments (Euser et al., 2013; Fenicia et al., 2014; Kirchner, 2009). This would identify whether the trends in streamflow are unique to Australia, or link in with global changes in the hydrological cycle (Huntington, 2006). Another avenue of future research would be to run the same analysis on a comparison catchment dataset which is strictly agricultural. Based on this research, we would hypothesise that these catchments would demonstrate greater changes in streamflow and larger elasticities.