Habitat modelling of key submergent macrophytes within the Lower Lakes,
South Australia
Matthew Linn Student ID: 211294695
School of Life and Environmental Sciences Deakin University
Submitted in partial fulfilment of the degree of Bachelor of Environmental Science (Honours)
October 2014
Statement of Responsibility
This thesis is submitted in accordance with the regulations of Deakin University in partial fulfilment of the requirements of the degree of Environmental Science Honours. I, Matthew Linn, hereby certify that the information presented in this thesis is the result of my own research, except where otherwise acknowledged or referenced, and that none of the material has been presented for any degree at another university or institution.
Date: 27/10/2014
Front cover: Site photo Point Sturt (site 6)
ii Abstract
The Murray-Darling Basin in south-eastern Australia is subject to the compounding effects of river regulation and extraction. The recent decade-long Millennium Drought saw large-scale changes in environmental conditions, degrading ecological communities and reducing species occurrence. With limited recovery of many communities post-drought, predictive habitat models were developed and field-validated to investigate the relationship between two key submergent macrophytes (Myriophyllum salsugineum and Vallisneria australis) and the environmental variables influencing their occurrence, using the Lower Lakes as a case study.
Telemetered records of logged environmental data were paired with vegetation monitoring data to develop non-parametric multiplicative regression models. The influence of the intra- seasonal variation in conductivity and water temperature from these telemetered records in conjunction with water pH from field surveys was found to define the habitat envelope for those species, and therefore, potentially limiting species occurrence post-drought. These findings provide managers with regional predictions of species responses that can be incorporated into management decisions to ensure submergent macrophyte assemblages remain viable into the future, while providing a proof of concept for a modelling approach that can be undertaken to describe similar relationships for other key taxa within the Murray- Darling Basin and abroad.
Key words: Non-parametric multiplicative regression, habitat modelling, post-drought, macrophytes
iii Acknowledgements
I wish to thank my two tireless supervisors Dr Rebecca Lester and Dr Jan Barton sincerely, for their unwavering encouragement and support throughout the entire honours year. Their wealth of knowledge and readiness to discuss, share and develop thoughts and ideas has made what was a challenging year, a really enjoyable and rewarding experience. I would also like to kindly acknowledge Dr Jason Nicol (with a special mention for field assistance and plant identification) and the South Australian Research and Development Institute, as well as Jason Higham, Peta Hansen and the Department of Environment Water and Natural
Resources, South Australia, for their generous contribution of time, effort, expertise, along with the provision of data. Additionally, I wish to thank Michael Diplock, Laurie Rankine Junior and the Ngarrindjeri Nation for their enthusiastic involvement in the present study. I also acknowledge the Elliott Newspaper Group, The Murray-Darling Foundation and The Murray-Darling Freshwater Research Centre for their generous contribution to the project through the awarding of the Elliot Newspaper Group scholarship. It is an honour to receive this prize and to have the study recognised for its role in enhancing the sustainability of river systems within the Murray-Darling Basin. David Dodemaide is greatly acknowledged and thanked for his ongoing support during the field work as well as throughout the entire process. Alex Pearse is also thanked for his assistance with lab work. I wish to acknowledge Deakin University and the Warrnambool School of Life and Environmental Sciences staff who have, fostered a study environment full of passion and desire to share knowledge that I am yet to see matched. I also acknowledge the funding provided by Deakin to support my study. Finally, I wish to acknowledge the support of friends, family and the 2014 honours cohort who have been supportive, compassionate and understanding throughout the year.
iv Table of Contents
Statement of Responsibility ... i
Abstract ... ii
Acknowledgements ... iii
List of Abbreviations ... vi
Introduction ... 1
Materials and Methods ... 7
Study area... 7
Vegetation monitoring data... 9
Telemetered environmental data ... 9
Developing preliminary habitat envelope models ... 12
Likelihood of occurrence ... 13
Vegetation assessment ... 16
Site characteristics ... 18
Statistical analyses ... 21
Refined models ... 21
Results ... 23
Preliminary habitat envelope model ... 23
Models predicting likelihood of occurrence ... 25
Myriophyllum salsugineum and Vallisneria australis abundance ... 27
Testing the preliminary models ... 28
v
Vegetation assemblages ... 29
Water quality and other site characteristics ... 33
Refined models ... 34
Discussion ... 36
Modelling approach and proof of concept ... 36
Pairing data from telemetered sites and monitoring to enable key species modelling ... 40
Influence of ranges to capture intra-seasonal variation ... 41
Consistency of predictors ... 44
Conductivity and water temperature ... 46
Water pH ... 47
Management within the system ... 48
Conclusion ... 50
References ... 51
Appendices ... i
Marine & Freshwater Research example publication http://www.publish.csiro.au/?paper=MF13163
Marine & Freshwater Research Author guidelines http://www.publish.csiro.au/nid/129/aid/434.htm
vi List of Abbreviations
NPMR - Non parametric multiplicative regression
SARDI - South Australian Research and Development Institute DEWNR - Department of Environment Water and Natural Resources EC - Electrical conductivity (μS cm-1 @ 25 °C)
AHD - Australian height datum (approximates mean sea level)
1 Introduction
Freshwater environments are dynamic and complex ecosystems with energy and nutrient pathways linking aquatic systems with their terrestrial surrounds (Likens and Bormann, 1974).
These aquatic systems provide important ecosystem services depending on the health and function of these systems (Rapport and Costanza 1998). However freshwater systems around the globe have been altered, with an estimated two thirds of the world’s flowing fresh water obstructed by dams on route to the sea (Nilsson et al. 2005). Regulating the flows of rivers with dams, weirs, locks and barrages has created reliable water resources for power production, agriculture and domestic use out of otherwise variable freshwater systems, shaping surrounding communities and industries (Grey and Sadoff 2007).
The environmental cost of regulation, and often over-extraction from, freshwater ecosystems has included shifts in diversity and abundance in biota. Natural lake level fluctuation plays a crucial role in freshwater aquatic ecosystem structure and function (Leira and Cantonati 2008), with native taxa having life histories adapted to a variable hydrological regimes. Water regulation has greatly affected fish, aquatic vegetation and invertebrate communities (Blanch et al. 1999; Ning et al. 2013; Bice et al. 2014). In addition, it has also affected water quality, influencing conductivity, turbidity and nutrient loads within these systems (Ahearn et al. 2005;
Mosley et al. 2012).
In drought-prone aquatic ecosystems, native taxa have adaptive life histories to persist and recover post-drought (Ning et al. 2013). Within regulated systems, these drought events can be exacerbated by water extraction and the natural resilience of taxa can be altered by deviations from natural conditions pre-drought (Lake et al. 2011). Drought places taxa under increasing stress via a progressive loss of resources, further reduced water quality and increased biotic interactions within a diminishing habitat, having deleterious consequences on many taxa (Bond
2 et al. 2008). With the natural cycle of drought linked with flood and subsequent recovery, recovery in regulated systems post-drought has been seen to vary among taxa (Boix et al. 2010) and in time frame (Rapport and Whitford 1999). With loss of natural variability, the synergistic effects of drought and subsequent loss in resilience of these systems, natural resource managers around the globe are dealt the challenging responsibility of ensuring ecosystem health into the future, whilst managing the demands of water-dependent industries.
Managers have recognised the value of controlled releases of additional water for environmental purposes (environmental watering) as a management tool to enhance ecosystem health and resilience in altered freshwater systems (Lind et al. 2007). Large water releases over time have been used to mimic natural cycles and pre-regulation condition, whilst other environmental watering programs have focused on targeted outcomes such as improving physicochemical characteristics (e.g. electrical conductivity), clearing periphyton off substrate in stream beds, clearing sediment deposits, reducing algal blooms, induce fish spawning, connecting fragmented habitat, inundating floodplains or restoring littoral vegetation and ecological communities (Kashaigili et al. 2005). As the targets of environmental flows have varied, so has the success of these actions (Tonkin and Death, 2014). Justifying further watering, particularly in regions with multiple competing interests for water resources, requires knowledge of the likely outcomes of watering actions (King 2006).
Large-scale flow regulation and over-extraction for irrigation has been occurring in the Murray-Darling Basin for well over a century (Murray Darling Basin Authority 2013). The River Murray is Australia’s largest river, draining 14 % of the total land mass of the continent (CSIRO 2008) but, when compared internationally to other rivers with similar catchment areas, the natural flow characteristics are low and sporadic, typical of semi-arid zone rivers (Puckridge et al. 1998). As a result of regulation, the River Murray has seen a loss in hydrological variability, seasonality and a reduction in annual flow volumes (Maheshwari et
3 al. 1995). This shift in hydrology from its natural state has been in favour of a productive and extensive system of water-intensive irrigation agriculture (Quiggin 2001). The Lower Murray system ranges from freshwater lakes to saline and hypersaline lagoons with diverse fringing wetlands, with a complex conductivity gradient balanced by freshwater inflows from the River Murray channel. This reliance on upstream flows leaves the Lower Murray system vulnerable to the effects of regulation and over-extraction (Mosley et al. 2012). The Lower Murray region contains a number of Ramsar-listed wetlands of international importance which support populations of nationally- and internationally-significant flora and fauna. However many macroinvertebrates, water birds, fish and aquatic vegetation within the region are in decline (Frahn et al. 2013; Wedderburn et al. 2012; Paton et al. 2009). The two Lower Lakes, Lakes Alexandrina and Albert, mark the freshwater extent of the River Murray, with barrages in the south of the system preventing salt water intrusion. Prior to regulation, the conductivity gradient in the Lower Lakes was naturally maintained with flushing river flows (Mosley et al.
2012). During the Millennium Drought (1997-2009), the Lower Lakes saw water levels drop, conductivities rise and acid sulphate soils exposed along lakeshores and surrounding channels (Mosley et al. 2012).
Research examining the response of a range of taxa to environmental variables within the Lower Murray system has identified conductivity and water regime as the two major variables influencing species occurrence and abundance (Wedderburn et al. 2007; Kefford et al. 2007;
Ning et al. 2013). Increasing conductivity, reduced flows and a changed water regime are symptoms of regulation and water management exacerbated by drought, therefore understanding how each affect a wide range of taxa within the system is crucial to its future management. In order to undertake effective management, natural resource managers require evidence-based and location-specific tools for assessing and accurately inferring ecosystem responses (Pace 2001). Researchers from Australian, and more recently international,
4 universities have been working with state and national government authorities to develop such tools for this region (Lester et al. 2013). This collaborative work has focused on developing tools to assess the large-scale ecosystem responses to changing environmental conditions.
Accurately forecasting the ecosystem response of a complex river system requires detailed knowledge of how organisms respond, at a community and species level.
Habitat encapsulates the biotic and abiotic features of an environment within which a species can occur. Species have definable habitat requirements or a habitat niche (Whittaker et al. 1973) and organisms are limited by the availability of resources and their tolerances to environmental conditions (Grinnell 1917). Changes in the suitability of habitat can result in changes in the distribution or occurrence of species, as is the case with the changes evident in the Lower Lakes in recent years. Habitat modelling uses environmental and associated biological data to determine what environmental characteristics define a given species habitat (Soberón 2010).
Depending on spatial scale of these models, they can be used to predict the extent of species distributions geographically, looking at the large-scale biogeographic characteristics that limit their distribution. At a finer scale, habitat models have been used to predict species occurrence using readily measurable and changing environmental variables such as physicochemical variables (Wedderburn et al. 2007). Using a combination of historical data and field validations, the spatial and temporal ranges of taxa within the Lower Murray system can be identified. This allows ecologists to not only assess the size and quality of habitat, but to further predict how this provision of habitat will change as these variables change through time.
Developing habitat models to identify the spatial and temporal distribution of key aquatic taxa within the Lower Lakes will assist managers to infer the future response of the taxa in question to predicted environmental conditions. Such knowledge can inform decisions regarding water allocation and environmental flows.
5 Non-parametric regression analysis (NPMR) is a flexible modelling approach that has many advantages over more traditional methods. For example, the predictors are not predetermined as they are in parametric regression but are constructed according to information derived from within the analysed data. Furthermore, the variables are able to be applied multiplicatively as opposed to additively (McCune 2006) which is beneficial because the effect of each variable can interact with others. Based on the understanding that species respond to multiple interacting factors within their environment, this method allows these interactions to be modelled with their natural complexity intact (McCune 2006).
NPMR has been widely used to model the response of a range of taxa to various environmental gradients at differing spatial scales and with differing outcomes. It has been used to quantify geographic ranges of bird communities post fire disturbance, Grundel and Pavlovic (2007) and finding an optimal fire frequency to be incorporated as management targets. DeBano et al.
(2010) examined the outbreak of insect pests in crops in Northwest United States, suggesting that warmer seasonal temperatures and lower elevations intensified outbreaks in crops, therefore enabling targeted mitigation measures to be developed. Habitat distribution modelling using NPMR has been undertaken within the Lower Murray system, providing insight into how species respond within their changed and changing environments.
Wedderburn et al. (2007) utilised NPMR modelling to determine differences in the distribution of two similar small-bodied Arthenidae fish species within the River Murray that share similar geographic ranges yet are rarely found to coexist. Rogers and Paton (2009) used the same approach to examine the decline in the range of Ruppia tuberosa, a key submergent macrophyte within the South Lagoon of the River Murray estuary. NPMR modelling was successfully used to determine what variable, or combination of variables, best explained the distribution of the target species in both these studies. Wedderburn et al. (2007) found conductivity to be the most influential predictor of species occurrence for both of the species in question, with the
6 separation in range described by differing predicted responses to a conductivity gradients within the system. For Ruppia tuberosa, increased conductivity was identified as the driver for the observed diminishing range Rogers and Paton (2009). Such studies highlight the importance of establishing not only what drives occurrence at a species level, but to also understand the type of response that each shows to measurable environmental gradients, so as to accurately predict habitat use.
This study focused on the submerged macrophytes Myriophyllum salsugineum and Vallisneria australis which were previously common within the system (Frahn et al. 2013). The two species were studied as individual taxa and also as potential surrogates for submerged macrophyte assemblages as a whole. Submerged macrophytes form important littoral habitat for invertebrate and fish species (Bice et al. 2014) and provide food and foraging sites for aquatic birds (Paton and Rogers 2009). They are also important in nutrient cycling within aquatic ecosystems (Carpenter and Lodge, 1986). As sediment-rooted species, M. salsugineum and V. australis, are limited by water depth and are subject to desiccation and the effects of water quality and turbidity (Middelboe and Markager 1997; Deegan et al. 2007). Despite recruiting in the system post-drought, neither species has returned to historic abundance levels (Frahn et al. 2013).
The factors limiting the resurgence of submergent macrophytes post drought are currently unknown. This study aimed to identify the environmental drivers limiting M. salsugineum and V. australis within the Lower Murray system by quantifying the environmental requirements for each species using predictive models. A secondary aim of the project was to field-test these models to assess the degree of success of each in predicting species occurrence in the Lower Lakes. The developed and validated models will provide natural resource managers in the region with decision-making tools to assist in the development of targeted management actions to achieving specific positive ecological outcomes within the system.
7 Materials and Methods
Study area
The study area comprised Lakes Alexandrina and Albert and the adjoining Goolwa Channel in the River Murray catchment, South Australia (Fig. 1). The three connected freshwater bodies are shallow, turbid and eutrophic (Mosley et al. 2012) and can be collectively described as the Lower Lakes system (Fig. 1)
Fig. 1. Map showing the sixteen lakeshore vegetation monitoring sites monitored by SARDI (2008-2014) from within Lake Alexandrina, Lake Albert and the Goolwa Channel (south-eastern Australia) used to develop predictive models.
All vegetation sites shown were used in the development of models, the selection of sites sampled during field testing of models shown in green, sites not sampled shown in blue. Major influences of hydrological disconnection include the barrages (permanent) and temporary regulators in operation during focal time period (2008-2014).
8 The Lower Lakes system is located above the Murray Mouth and the Coorong estuary, at the terminal end of the Murray-Darling Basin. The delineation between the freshwater Lower Lakes and the estuarine and saline Coorong is artificially maintained by a series of barrages constructed along the southern margins of Lake Alexandrina and the Goolwa Channel (Fig. 1).
In the Lower Lakes system, the three focal water bodies differ in hydrology, geomorphology and management. Lake Alexandrina is the largest (662 km2) and deepest (2.8 m average depth) and receives inflow directly from the River Murray to the north of the lake (Fig. 1). It also receives inflow from the Eastern Mount Lofty Ranges via the Angas and Bremer Rivers near Milang (Fig. 1) and can receive input from the Goolwa Channel. To the east, the shallower (1.7 m average depth) and smaller (177 km2) Lake Albert is connected to Lake Alexandrina by a narrow (200-300 m) channel (named the Narrows). The Goolwa Channel is a narrow (600-900 m) western extension of Lake Alexandrina and the junction of the Finniss River and Currency Creek tributaries, to the south the Goolwa channel is separated from the estuarine Coorong lagoon by the Goolwa Barrage (Fig. 1)
The focal time period for my study (2008-2014) encompassed the final years of the decade- long Millennium Drought in the Murray-Darling Basin which resulted in large-scale changes in water levels, conductivities, turbidity and water chemistry within the Lower Lakes system (Mosley et al. 2012). The focal time period also includes drought-breaking flooding that occurred in 2010 (Mosley et al. 2012). During the Millennium Drought, record low lake levels exposed acid sulphate soils along the lake margins of the three water bodies of the Lower Lakes system. In response, lake level regulators were constructed disconnecting the Goolwa Channel (Clayton Regulator, constructed in 2009) and Lake Albert (Narrung Bund, constructed in 2008) from Lake Alexandrina to enable the water level in each section to be managed independently and minimise the impact of acidification. The water bodies were later reconnected as the Lake Alexandrina rose with high flow levels in 2010 (Mosley et al. 2012).
9 Vegetation monitoring data
The South Australian Research and Development Institute (SARDI) monitor macrophyte species abundance at 16 lakeshore sites within the Lower Lakes. For my study, vegetation data were available from spring 2008 to autumn 2013, covering ten sampling periods. These biannual vegetation assessments re-survey the same sites at set elevations at each sampling period, providing species cover abundance scores for each elevation. The monitoring program includes 106 species and results have shown that submergent vegetation has not returned in comparable abundances and diversity since the Millennium Drought (Frahn et al. 2013).
Through consultation with the SARDI Plant Ecology Sub-program Leader (pers. comm. J.
Nicol, 2014), two species were selected for initial modelling, Myriophyllum salsugineum and Vallisneria australis, as being representative of the submerged vegetation assemblage of the region. They were selected as the target of modelling so as to identify the factor or factors limiting their recovery within the system which was of interest to the management agency responsible for biodiversity in the region (the South Australian Department of Environment and Natural Resources; DEWNR).
Telemetered environmental data
Vegetation assessment sites were matched with corresponding available telemetered environmental sites. Environmental data were provided by DEWNR from 26 telemetered real- time water monitoring stations within the Lower Lakes system. These moored telemetry stations vary with respect to the variables that are recorded and duration of records available.
In total, data were available for up to six variables, recorded from the telemetry stations within the Lower Lakes between January 2006 and March 2014. Of these, three variables were sufficiently spatially and temporally complete for use in modelling: Electrical conductivity (EC, μS cm-1 @ 25 °C), water temperature (oC) and lake level (m above the Australian Height
10 Datum [AHD], which approximates mean sea level) (Appendix 1). These three variables were recorded across the focal time period at eight of the 26 telemetry stations throughout the Lower Lakes. Six of these eight stations had recorded all three variables whilst two of the eight did not record lake level and so lake level was aliased from the nearest lake level recorder, taking into account any potential hydrological disconnection when deciding on pairs (Fig. 1). The six complete stations and two stations with aliased lake level provided eight telemetered sites covering the focal period. These were then matched to 16 vegetation sites by proximity, again taking into account any hydrological disconnection that may exist within the water bodies (Fig.
2; Table. 1).
Table 1. Paired telemetry sites matched vegetation sites from within the Lower Lakes used to model habitat characteristics for two key macrophytes, Myriophyllum salsugenium and Vallisneria australis
Site number
Telemetric site name
Vegetation
site number Vegetation monitoring site names
T1 b20/23GC 1,2,3 Goolwa South, Hindmarsh Island Bridge 1 and Hindmarsh Island Bridge 2
T2 Beacon 65 4 Clayton Bay
T3 b78/Point McLeay 5 Clayton Upstream of Regulator T4 Beacon 97 6,15,16 Point Sturt & Narrung, Terrengie T5 Milang Jetty 7,8 Milang & Bremer Mouth
T6 Near Mulgundawa 9 Lake Reserve Road
T7 Near Waltowa 10,11,14 Browns Beach 1, Browns Beach 2 and Nurra Nurra
T8 Warringee Point 12,13 Warrengie 1 and Warrengie 2
11 The archived data from these eight sites were initially checked against site status codes provided by DEWNR. Status codes provided dates of faults with either unknown or known causes that resulted in values outside of expected ranges or errors with logging of data. Site data were compared among all eight sites for each recorded variable of focus (EC, water temperature and lake level) and data listed as containing errors were excluded. Daily averages for each variable were plotted against their date of recording. Deviation from trends were observed and compared to the recent history and management of the region. For the Goolwa Channel and Lake Albert, the management interventions of the construction of the Clayton Regulator and Narrung Bund explained differences from trend in lake levels and conductivities observed during this time period and these data were included in the modelling. Seasonal
Fig. 2. Map of the Lower Lakes, showing telemetered sites (blue triangles) matched with lakeshore vegetation monitoring sites (green circles).Connections between the telemetered sites and the lakeshore vegetation monitoring sites shown as solid red lines. The aliasing of lake level data from nearby lake level recorder used in the absence of lake level recording at the telemetered site to allowing for complete records is indicated by dashed red lines.
12 means, maxima, minima and ranges were calculated for each variable from each telemetered site (summer, autumn, winter and spring). Seasons were only included when they had a minimum of 60 records to minimise bias due to unverified records or faults in data recording.
These eight telemetered sites were found to have sufficient data to calculate seasonal statistics across the focal time period.
Developing preliminary habitat envelope models
Non-parametric Multiplicative Regression (NPMR) was undertaken using HyperNiche (ver.
1.0, McCune, 2006) to model the interactions between the response variables of species occurrence for each of the two focal species and the environmental predictor variables of EC, water temperature and lake level, from the eight paired telemetry stations (Fig. 2). A Gaussian weighting function with a local mean estimator was used, because a Gaussian distribution assumes a smooth and continuous response to predictor variables (McCune 2006) which matched the expected response of submerged macrophytes within the system (Gehrig and Nicol 2010) and had been used for a similar study of a submergent species within the region (Rogers and Paton 2009). Models were first developed using environmental data from a selection of 13 of the available 16 vegetation sites, corresponding with the most complete telemetered records, allowing for the inclusion of up to two lag years and eight lag seasons as predictor variables.
An additional series of four separate models were developed with differing lengths of record and numbers of sites to encompass all available data. This series of models were constructed including up to all 16 sites, however was limited to the inclusion of one lag season (the season previous), with lag years excluded. The most spatially-inclusive model of this series contained all 16 available sites covering eight seasons (spring 2010- spring 2012) while the most temporally inclusive model included 13 sites and 18 seasons (autumn 09 – spring 2013) (Appendix 2).
13 Model evaluation was undertaken by determining model fit to the response variables, defined by cross-validated R-squared values (xR²). Unlike traditional R-squared values, xR² values are calculated using a leave-one-out-cross-validated approach, whereby the data point used in the estimate of the response from that point is left out. This consequently means that the residual sum of squares can exceed the total sum of squares resulting in a negative xR² value for weak models and is undertaken to reduce over fitting (McCune 2006). The predictors used within each of the best-fitting models were evaluated using sensitivity and tolerance analysis.
Sensitivity analysis undertaken during the model evaluation stage used small nudges (±5 % of predictor value) to the value against observed changes in the response variable. Evaluating the relative importance of each predictor, higher sensitivities indicate a higher influence within the model (McCune 2006). The tolerance of continuous predictors used in the models is inversely related to the importance of a variable, with variables that have high tolerance using a larger neighbourhood of data points to make a prediction, suggesting a wider tolerance to that predictor variable (McCune 2006). The model’s ability to explain a significant proportion of the variation within the response variables was tested using Monte Carlo random runs.
Likelihood of occurrence
To form predictions of occurrence for the two target species, four categories were calculated describing the likelihood of occurrence based on the best models for each species (high, moderate, low and very low). The models were used to generate an estimate of expected abundance per site and per time period (season). These predictions of abundances were averaged for the last available year (2013) to be applied to the sites selected for field surveys in 2014. The predicted annual averages were divided into the four categories using the quartiles of the estimates of predicted abundance. Due to a small number of observations, categories were compiled into two broader categories for statistical testing using Fishers Exact test. A
14 Fishers Exact test allowed a test of the relationship between predictions and observations, using likely present (high and moderate) and likely absent (low and very low) as the potential outcomes and the recorded observations from field survey in June 2014 as the observed outcomes.
Field validation of model predictions
Field surveys were undertaken in the Lower Lakes to test the quantitative predications of the preliminary models, comparing submerged macrophyte coverage with predictions of occurrence based on point-in-time environmental conditions. Field surveys were undertaken in July 2014 at nine existing vegetation monitoring sites used in modelling (as described below;
Fig. 3). Of these nine sites, six sites were selected because M. salsugineum or V. australis had been recorded at those sites over the monitoring period. The remaining three out of a possible 10 sites, where the target species were less likely to be present, were selected at random. Two of the nine sites, Clayton Upstream and Bremer Mouth, were selected for their recorded presence of M. salsugineum and V. australis during the monitoring program. They had not been included in the most complete modelled datasets due to gaps in the paired telemetered dataset not allowing for the inclusion of lagged season and years. So, these two sites were not categorised by likelihood of occurrence and did not provide a test for the predictive models.
They did, however, allow for the variation of lakeshore sites to be assessed, although not formally analysed or used in the refined models.
15
Fig. 3. Map showing predicted likelihood of occurrence categories for the nine lakeshore vegetation monitoring sites resampled in July 2014 testing the accuracy of the best developed models. Species depicted as circles Myriophyllum salsugineum (M) and Vallisneria australis (V) with categories likely present (green), likely absent (red) and not assessed (black)
16 Vegetation assessment
In the field, detailed habitat assessments were conducted at each site to test the influence of a wider range of site characteristics than had been surveyed by SARDI, which had focused predominantly on vegetation. Sites were located using GPS and vegetation coverage assessments were conducted in accordance with the methods described in Frahn et al. (2013), for consistency with the monitoring data used in modelling. Verification of consistency with the SARDI vegetation assessment methods and assistance with identification of species was provided by the SARDI Plant Ecology Subprogram Leader, Dr Jason Nicol on the first day of sampling. At each site, vegetation assessments involved sampling three replicate transects running perpendicular to the shoreline separated by 1-m intervals. At each transect, five quadrats (1 x 3 m) were established at set elevations of +0.8, +0.6, +0.4, +0.2, and 0.0 m AHD (Fig. 4), with elevations kept consistent across sites.
Real-time telemetered data from the closest lake level station were accessed and, using lake level as a reference point, the elevations were established during the field survey for each site.
A Leica Geo-systems NA720 automatic level and surveyor’s staff were used to determine elevations above the water level, whilst depth of water was used to determine the elevations below that level. Macrophyte species were identified in the field (Sainty and Jacobs 2003) or photographed for later identification, whilst percent coverage was estimated for species present within each quadrat. The logistical constraints of water depth at the lowest elevation (-0.5 m AHD) prevented vegetation and habitat assessments and this elevation was excluded from analysis.
17
Fig. 4. Vegetation sampling protocol for lakeshore sites undertaken by SARDI and repeated for this study. The plan view shows the placements of quadrats relative to the shoreline. The shoreline is shown at 0.75m AHD at the current (2014) lake level target.
18 Site characteristics
Site characteristics were recorded at the time of vegetation assessments to be used as additional predictor variables in the secondary modelling. Fourteen additional variables were selected for their potential influence on macrophytes assemblages (Table 2). Water quality measurements were taken from mid-water depth (10-40 cm depth), standing at the shore line with aid of a sampling pole, with replicates at each of the three vegetation transects (Table 2). All measurements were taken prior to entering the water to avoid disturbing the water column.
Dissolved oxygen, EC, pH and temperature were all measured using an YSI Pro Plus Multi- Parameter Water Quality Meter calibrated before each field day. Water clarity was measured using a secchi disc, measuring the depth at which the quadrants on the disc were no longer visible.
A rapid VISOCOLOUR® ECO Photometer nutrient test kit was used to measure concentrations of nitrate (NO3-), ammonium (NH4+), nitrite (NO2-) and phosphate (PO43-). For each replicate transect, one 50-ml water sample was taken (a total of three per site) using a sampling pole to avoid disturbance of the water column. Samples were placed in rinsed collection vials before being stored on ice. Samples were stored for a maximum of 10 hours prior to analysis. Upon analysis, samples were homogenised and 10-ml sub-samples were taken and filtered (with a pore size of 0.2 μm) to remove particulates before for each test.
The slope of the shoreline was calculated based on measurements taken between the distances between each sampled elevation using a surveyor’s tape held parallel to the water’s surface.
Exposure to wind was estimated using a clinometer, adding the measured angles of inclination of the horizon taken at each of eight compass points. Three measurements of sediment pH, sediment temperature and redox were measured from each site, one quadrat per replicate transect that was selected at random. Sediment pH was measured using an Inoculo soil pH kit.
Sediment redox and temperature were measured with a HANNAH HI98120 ORP/Temperature
19 meter submersed in the sediment for two minutes, allowing the reading to stabilise before the measurement was recorded. Sediment particle size estimates were undertaken by visual and physical examination of surface sediments from each quadrat.
20
Table 2. Quantitative site characteristics recorded during field surveys from nine sites within the Lower Lakes in June 2014, indicating the unit for each variable measured, the level of replication (transect, site, quadrat, elevation, core/quadrat) ,the number of observations taken and inclusion as predictor variables in development of refined models (yes/no)
Water quality Units
Replicate level
Replicates per site
Total replicates
Included in modelling
Temperature °C Transect 3 27 Yes
Dissolved O₂ % Transect 3 27 No
pH - Transect 3 27 Yes
EC @ 25°C μS cm-1 Transect 3 27 Yes
Water clarity (secchi depth) cm Transect 3 27 Yes
Nutrients
Total Phosphorous mg L-1 Site 1 9 No
Nitrate mg L-1 Site 1 9 No
Nitrite mg L-1 Site 1 9 No
Ammonia mg L-1 Site 1 9 No
Physical characteristics
Water height above quadrat cm Quadrat 15 135 Yes
Slope of bank (distance between elevations) m Transect 3 27 Yes
Wind exposure (angle of inclination of horizon) degrees Site 1 9 Yes
Sediment characteristics
Oxidation reduction potential (redox) mV Transect 3 27 Yes
Particle size estimate % Elevation 5 45 Yes
21 Statistical analyses
Prior to secondary modelling, a statistical analysis of the field-assessed variables was undertaken to complement NPMR modelling. Multivariate analyses were performed using PRIMER v. 6.0 (Clarke et al. 2006) with the PERMANOVA + add on (Anderson 2007). Non- metric multidimensional scaling (MDS) plots allowed for differences and variability among sites and likelihood of occurrence categories (likely present and likely absent) to be visualised.
This was undertaken for each of the site characteristics, water quality, sediment characteristics, macrophyte assemblages and macrophyte species coverage, with permutation-based multivariate analysis of variance (PERMANOVA) used to test these differences. A two-factor PERMANOVA design was used with sites (random factor) nested within likelihood of occurrence categories (fixed factor). Environmental data were normalised to account for the differing scales of measurement and then a Euclidian distance similarity matrix was constructed before a MDS plot was used to visualise patterns. Coverage data underwent no pre- treatment and a Bray-Curtis resemblance matrix with dummy variable (of 1) was constructed, with the dummy variables added to account for the influence of zero abundance values within the dataset.
Refined models
To refine the models, the same telemetered and vegetation monitoring datasets (i.e. including lag seasons and years) were used, with the addition of measured site characteristics from the field surveys. The site characteristics that were included as potential predictor variables in the refined models excluded dissolved oxygen and soil temperature (Table 2), as changes in these variables were likely to be influenced by the presence of macrophytes (Carpenter & Lodge 1986) as opposed to those variables influencing macrophyte occurrence. The eleven site characteristics included were water pH, measured EC, water clarity, average water depth, wind
22 exposure, bank slope, sediment pH, sediment redox and the percentage contribution of sand, gravel, cobble and boulders to the sediment (Table 2).
These additional eleven predictor variables were applied to the seven resampled sites (from the best models for both species) as site averages expressed over all time periods. The models again were developed using HyperNiche, using a Gaussian weighting function with a local mean estimator. The models developed were then evaluated based on their fits to model dataset defined by the xR2 value and best fitting models were selected, as described for the preliminary model development above. Sensitivity and tolerance analyses were again undertaken to investigate the relative importance of each selected predictor variable within these best selected models. Validation of the predictions of these models was not undertaken as was undertaken for the preliminary models, as these models were derived from observation from the field data and a new round of sampling would be required to quantify the accuracy of predictions, which was outside the scope of this project.
23 Results
Preliminary habitat envelope model
All sixteen vegetation sites were included in modelling, using the eight telemetered sites to identify habitat envelopes for the two target vegetation species. The best models for M.
salsugineum and V. australis were developed from models including 13 of these sites and lagged seasons and up to two lagged years. These models showed a moderate ability to represent vegetation distributions with xR² values ranging from 0.30 to 0.66 (Table 3). The best model for M. salsugineum (xR² = 0.66) used four predictor variables: Average EC for the season sampled; Range of EC for the previous season, Average lake level for two seasons previous; and Range of EC for three seasons previous (Table 3). For V. australis (xR² = 0.30), the best model included Average EC for the season sampled, Range of EC for the season sampled, Range of lake level for the season sampled and Average lake level for two years previous (Table 3). Based on Monte Carlo random runs, the model for M. salsugineum explained a significant proportion of the variation recorded in M. salsugineum abundance (P(MC) = 0.047), as did the model for V. australis (P(MC) = 0.047).
24
Table 3. The best preliminary models derived from 13 vegetation sites that were matched with the most temporally-complete telemetered records allowing for the inclusion of lagged seasons (e.g. -1 season) and the inclusion of lagged years (e.g. -2 years). xR2 describes the model fit and p(MC) denoting the result of Monte Carlo tests (α = 0.05). Predictor variables included in best models are each listed in order of the sensitivity, representing the influence of each predictor within the modelled data set. The tolerance shown for each predictor included in the best models for each species shows the impact of each predictor on the response variable.
When a series of models were developed using sub-sets of the available sites and years to maximise the data included (Appendix 2), all had lower correlations than those described for the overall model (i.e. xR² < 0.2) but the predictor variables of average EC and average lake level that were common among these models were similar to the predictors identified from the best single models selected for M. salsugineum and V. australis, as described above.
Sensitivity analysis for the best model for M. salsugineum showed that the Range of EC for one season previous (sensitivity = 0.759) was the most important predictor within the model (Table 3). Higher sensitivities indicate that larger changes in the response variable as a result of changes in the predictor variable. Tolerance values are not directly comparable among the predictor variables like sensitivity analysis, as data were not normalised, but they represent the species range of tolerance to predictor variable with higher tolerances equating to wider species tolerance to variables. For V. australis, the most important predictor for describing abundance
Model Predictor Sensitivity Tolerance
Myriophyllum Range of EC, - 1 season 0.759 633.750 salsugineum Average Lake level, -2 season 0.160 0.061
xR²=0.656 Range of EC, - 3 season 0.092 2108.700
P(MC)=0.047 Average EC, current 0.079 1637.045
Vallisneria Range of EC, current 0.169 1449.400 australis Range in Lake level, current 0.138 0.077
xR²=0.302 Average lake level, - 2 years 0.111 0.229
P(MC)=0.047 Average EC, current 0.043 777.532
25 was Range of EC of the current season (sensitivity = 0.169), as defined by sensitivity, with this predictor variable also having the highest tolerance value within the model.
Models predicting likelihood of occurrence
Three of the sixteen modelled sites were not categorised by these best models, derived from the 13 sites with the most temporally complete telemetered records. Likelihood of occurrence, based on the best models for M. salsugineum and V. australis, predicted the same four categories for each of the thirteen sites for the two species. However, the predicted abundances for each site from the best models varied between the two species. M. salsugineum was predicted in the highest abundance at Clayton Bay (0.30 % coverage averaged across site) with the highest abundance of V. australis predicted at Lake Reserve Road (0.16 % coverage averaged across site). Quartiles of predicted abundances were used to derive the categories resulting in two sites that were predicted to have high abundances for both species, three moderate, two low and six where abundance was predicted to be very low or absent (Table 4).
26 Table 4. Predicted abundances used to categorise likelihood of occurrence for Myriophyllum salsugineum and Vallisneria australis among the 16 lakeshore vegetation monitoring sites derived by the best preliminary models. Likelihood of occurrences were derived from quartiles of predicted abundances predicted the same likelihood categories (High, Moderate, Low, Very low) for both species at all sites. Sites not included in the best preliminary models (i.e. 13-site model with lagged seasons and lagged years) were unable to be categorised (Bremer Mouth, Clayton upstream and Milang). *denotes sites resampled in July 2014.
Likelihood of occurrence
Site
number Site
Predicted abundance % cover of M. salsugineum
Predicted abundance % cover of V. australis
High 4 *Clayton Bay 0.299 0.0192
9 *Lake Reserve Road 0.051 0.1612
Moderate 15 Narrung 0.047 0.0045
6 *Point Sturt 0.041 0.0045
16 Terrengie 0.047 0.0045
Low 13 Warringie 1 0.013 0.0004
12 *Warringie 2 0.013 0.0004
Very low 14 *Nurra Nurra 0.010 0.0003
10 Brown Beach 1 0.010 0.0003
11 Brown Beach 1 0.010 0.0003
1 *Goolwa South 0.000 0.0000
3 Hindmarsh Island Bridge 1 0.000 0.0000
2 *Hindmarsh Island Bridge 2 0.000 0.0000
Not categorised 8 *Bremer Mouth No prediction No prediction
5 *Clayton upstream No prediction No prediction
7 Milang No prediction No prediction
27 Myriophyllum salsugineum and Vallisneria australis abundance
From the nine resampled sites the target species, M. salsugineum and V. australis were found within Lake Alexandrina and the Goolwa Channel (Fig. 5). No submergent macrophyte species were recorded at sites within Lake Albert (Nurra Nurra & Warrengie). The four sites at which M. salsugineum were recorded at were Clayton Bay (predicted to have a high likelihood of occurrence), Goolwa South (predicted low likelihood), Bremer Mouth (not categorised) and Clayton Upstream (not categorised) (Fig.5). V. australis were recorded at three sites: Point Sturt (predicted moderate likelihood); Clayton Upstream (not categorised) and Bremer Mouth (not categorised).
Fig. 5. Map showing predicted likelihood of occurrence categories for the nine lakeshore vegetation monitoring sites that were sampled in July 2014 to test the accuracy of the best developed models. Species are depicted as circles labelled Myriophyllum salsugineum (M) and Vallisneria australis (V) with likelihood categories shown as likely present (green), likely absent (red) and not assessed (black). The four correct predictions for Myriophyllum salsugineum are shown as ticks and the three incorrect predictions are shown as crosses. Site 1 (Goolwa South) was the only site predicted likely absent that was found to have either species present. For Vallisneria australis, again the five sites predicted correctly are shown with ticks and the two incorrect predictions are shown as crosses.
28 The percent coverage of M. salsugineum and V. australis averaged over the 15 quadrats at these sites ranged between 1.0 and 2.4 % for M. salsugineum and 0.4 to 3.3 % for V. australis, much higher than the predictions of the models. There was little difference in the elevation and water depth at which these two species were recorded. M. salsugineum was recorded at four of the five sampled elevations (+0.8, +0.6, +0.2 and 0.0 m AHD) in water depths up to 76 cm at 0.0 m AHD. The shallowest occurrence was at 0.8 m AHD, at the water’s edge (i.e. 5 cm). V.
australis was recorded at the same range of elevations (+0.8, +0.6, +0.4 m and 0.0 m AHD) with a similar range of water depths, of wetted edge at 0.8m AHD up to 79 cm deep at 0.0 m AHD. Three other submergent macrophyte species, Potamageton crispus, P. pectinatus and Ceratophyllum demersum were identified during sampling. P. crispus was found at Bremer Mouth (not categorised). C. demersum was found at four sites, including Clayton Upstream (not categorised), Bremer Mouth (not categorised), Clayton Bay (predicted high likelihood for the two target species) and Hindmarsh Island Bridge 2 (predicted very low likelihood for the two target species), and P. pectinatus was found at Point Sturt (predicted moderate likelihood for the two target species).
Testing the preliminary models
The best preliminary model for M. salsugineum (Table 5) correctly predicted the occurrence at site which had the highest abundance, Clayton Bay (predicted high likelihood), but was poor at predicting occurrences at the other six categorised sites. In total, M. salsugineum was recorded at only one site that was categorised as high likelihood (Clayton Bay) and one that had been categorised as very low likelihood (Goolwa South). V. australis was recorded at one site that had been categorised as a moderate likelihood site (Point Sturt) but was absent at all other categorised sites (Table 5).
29
Table 5. Comparison of the likelihood of occurrence categories (High, Moderate, Low and Very low) and the observed presence or absence of Myriophyllum salsugineum and Vallisnria australis during field surveys in June 2014.
Contingency tables of the likely occurrence categories for both species were tested using a Fishers Exact Test, grouping the high and moderate likelihood categories (likely present) and low and very low categories (likely absent). The ability of the model to accurately predict the likelihood of occurrence (based on the proportion of the predicted present or predicted absence categories that were correct) was not statistically significant for either species (M. salsugineum, P = 1.00 and V. australis, P = 0.43).
Vegetation assemblages
Field vegetation assessments identified 34 macrophyte species in total (Appendix 3) comprising of seven functional groups (Frahn et al. 2013). The most frequent and abundant species among sites was Phragmites australis, recorded at all sites and covering an average 12
± 21% of all of the area sampled (Fig. 6). Emergent species were the most abundant functional group at all sites, covering an average 32 ± 30% of the sampled area at sites. Terrestrial plants
Likelihood of occurrence
Site
number Site M. salsugineum V. australis
High 4 Clayton Bay Present Absent
9 Lake Reserve Road Absent Absent
Moderate 6 Point Sturt Absent Present
Low 12 Warringie 2 Absent Absent
Very low 14 Nurra Nurra Absent Absent
1 Goolwa South Present Absent
3 Hindmarsh Island Bridge 2 Absent Absent
30 made up the most diverse functional group, with 8 species recorded covering an average of 3
± 6 % at sites sampled (Fig. 6).
Submergent K-selected Floating
Amphibious fluctuation responder plastic
Amphibious Fluctuation responder tolerator emergent Emergent
Terrestrial damp Terrestrial dry
Fig. 6 (a) The variation of average percent coverage of macrophyte functional groups among the nine resampled sites and (b) the percentage contribution of macrophyte functional group coverage at each site observed during field surveys in June 2014.
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Lake reserve Rd
Pt Sturt Clayton Bay
Goolwa South
H. Island bridge 2
Nurra Nurra
Warrengie Bremer Mouth
Clayton Upstream
% contribution
Site
(b)
0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0
Lake reserve Rd
Pt Sturt Clayton Bay
Goolwa South
H. Island bridge 2
Nurra Nurra
Warrengie Bremer Mouth
Clayton Upstream
% cover
(a)
31 When considering whether sites were predicted to be likely present or likely absent, vegetation assemblages at the seven resampled categorised sites were not significantly different between the two categories for species percent coverage (pseudo-F1, 5 = 0.94 P = 0.46; Table 5) or the seven functional groups (Fig. 6) (pseudo-F1, 5 = 0.83, P = 0.505; Appendix 4). In contrast, there were significant differences among sites nested within the likelihood categories for species percent coverage (pseudo-F5, 28 = 2.83, P = 0.001; Appendix 4) and functional groups (pseudo- F5, 28 = 2.00, P = 0.007; Appendix 4), indicating that there were high levels of variability at small spatial scales that was not explained by the preliminary model predictions.
Water quality and other site characteristics
EC was the most variable water quality parameter recorded, ranging from 858 μs cm-1 at Lake Reserve Road (predicted high likelihood of occurrence) to 8866 μs cm-1 near the Coorong estuary at Goolwa South (predicted low likelihood), with an average among sites of 2900 ± 5966 μs cm-1 (Table 6). Dissolved oxygen and water clarity also had wide ranges, of 10.8 to 103.2 % and 4 to 46 cm, respectively. Water pH ranged between 6.9 and 8.4, in contrast, temperature which was the most consistent variable among sites, with a range of 8.6 to 10.3
°C. Nitrogen and phosphorus, as measured using the test kit, were below detection limits for 23 of the 36 nutrient tests, with total nitrogen recorded as <2.0 mg L-1 and total phosphorus was above detection limits >15.0 mg L-1. Given these results, nutrient concentrations were not included in any further analysis.
32
Table. 6 Site characteristics recorded from field surveys in June 2014 that were then included in the development of refined models. The unit for each variable measured is indicated, with measuremnets reported as site means. * denotes sites that were not included in the refined models.
Site Name Units *Clayton
Upstream
Goolwa South
Hindmarsh Island Bridge 2
Nurra
Nurra Warrengie Pt Sturt *Bremer Mouth
Lake Reserve Road
Clayton Bay Water Quality
Temperature °C 8.8 11.7 10.5 9.2 9.6 10.5 9.6 11.9 9.7
pH - 7.0 7.6 7.5 8.3 7.6 8.2 7.1 8.2 7.3
Electric Conductivity μS cm-1 1239.3 8841.7 4714.7 2575.3 3820 1207.7 1748.7 858.7 1470.7
Water clarity cm 27.7 32.3 44.7 5.3 9.3 7.7 34.3 11.7 25
Physical Characteristics
Water depth cm 36.3 38.7 32.0 38.0 35.0 30.0 26.0 32.0 30.0
Slope of bank m 43.6 17.0 34.6 39.0 44.8 33.6 21.4 3.1 38.5
Wind exposure degrees 48 73 22 19 57 46 29 18 71
Sediment Characteristics
Soil redox mV 30.7 -37.0 16.7 44.3 -83.7 120.3 -85.3 -90.7 1.7
Sediment grain size % contribution
Clay/silt 97.9 81.4 52.1 8.7 74.2 2.7 53.4 97.9 97.9
Sand 0.5 17.3 34.3 89.8 24.3 82.0 41.3 0.5 0.5
Gravel 0.5 0.4 2.4 0.5 0.5 1.6 4.4 0.5 0.5
Cobble 0.5 0.4 10.6 0.5 0.5 12.1 0.5 0.5 0.5
Boulders 0.5 0.4 0.5 0.5 0.5 1.6 0.5 0.5 0.5
33 Sediments at all sites largely consisted of clay/silt (<0.06 mm) and sand (0.06-2 mm). Gravel (2-64 mm) was predominant at Bremer Mouth, Hindmarsh Island Bridge 2 and Point Sturt.
Cobble (65-256 mm) and boulders (>256 mm) were found at Hindmarsh Island Bridge 2 and Pt Sturt. Wind exposure (angle of inclination of the horizon) varied within predicted likelihood occurrence categories, with Lake Reserve Road (predicted high likelihood) and Nurra Nurra (predicted very low likelihood) being the most exposed sites. In contrast, Goolwa South (predicted very low likelihood) and Clayton Bay (predicted high likelihood) were the most sheltered sites. Slope of bank varied among the sites, the steepest sloping site with a 0.170 m gradient being Lake Reserve Road (predicted high likelihood) and the shallowest gradient with 0.007 m at Goolwa South (predicted very low likelihood) (Appendix 5). Soil pH was consistent with water pH, Clayton Bay (pH 7.0) and Clayton upstream (pH 6.5) were the most acidic recorded and Nurra Nurra (pH 8.5) and Warrengie 2 (pH 8) the most alkaline. Soil redox varied among sites, with the highest being Point Sturt (120 mV) and the lowest being Lake Reserve Road (-90 mV).
There were no significant differences between the predicted likelihood of occurrence categories of M. salsugineum and V. australis when examining water quality (pseudo-F1, 5 = 0.86, P(MC)
= 0.47; Appendix 4) or physical site characteristics (pseudo-F1, 5 = 0.75, P(MC) = 0.589;
Appendix 4). However, similar to the tests of vegetation assemblages, there were significant differences among sites nested within categories for both water quality (pseudo-F5, 14 = 81.49, P = 0.03; Appendix 4) and physical characteristics (pseudo-F5, 28 = 1.6259, P = 0.001;
Appendix 4). Again, this suggests that the variation among recorded site variables were not described by the categories of likelihood predicted by the best preliminary vegetation models.
34 Refined models
Following the field validation of the original models, refined models for both M. salsugineum and V. australis were developed. The refined models were developed using the seven categorised sites that were surveyed in the field, the telemetry records used in the development of the previous best preliminary models and water quality and site characteristics from the field survey. The best refined models, again defined by the xR2 value, both used three predictor variables and had increased fits to the previous best preliminary models. For M. salsugineum, the refined model had an excellent fit (xR2 = 0.936), as did the model for V. australis (xR2 = 0.9743) and both were found to describe a significant proportion of the variation in the modelled data (M. salsugineum, P = 0.046 and V. australis, P = 0.046). Akin to the preliminary models, the refined models did not vary in predictor variables used between the two target vegetation species, however the relative importance of each did vary within the model, as did the period of influence of the predictor. For M. salsugineum, the most important variable (again defined by sensitivity) was Range in EC for one previous season (sensitivity = 0.20), followed by Range in water temperature for the current season (sensitivity = 0.10) and Water pH from the resampling event in June 2014 (sensitivity = 0.02). The most influential predictor of V.
australis abundance was Range in water temperature two years previous (sensitivity = 0.08), with Water pH having similar influence within the model (sensitivity = 0.07) and finally, Range in EC for the current season (sensitivity = 0.02).
35
Table 7. The best refined models derived from 13 vegetation sites that were matched with the most temporally-complete telemetered records, allowing for the inclusion of lagged seasons (e.g. -1 season) and the inclusion of lagged years (e.g. -2 years). xR2 describes the model fit, p(MC) denoting the result of Monte Carlo tests (α = 0.05).
Predictor variables included in best models are listed in order of the sensitivity of each, representing the influence of each predictor within the modelled data set. The tolerance shown for each predictor included in the best model for each species, showing the impact of each predictor on the response variable.
Model Predictor Sensitivity Tolerance
Myriophyllum salsugineum Range in EC, -1 season 0.199 2112.500
xR²=0.956 Range WT 0.099 1.396
P(MC) = 0.046 Water pH 0.021 0.124
Vallisneria australis Range in water temp, -2 years 0.082 0.794
xR²=0.9743 Water pH 0.066 0.249
P(MC) = 0.046 Range in EC 0.018 9058.750
36 Discussion
The current state of reduced diversity and coverage of submergent macrophytes within the Lower Lakes represents a loss of complex and crucial littoral habitat for birds, macroinvertebrates, small-bodied and larval fish (Ning et al. 2013; Paton and Rogers 2009;
Wedderburn et al. 2012). Thus, there is a need to identify the factors that are limiting the recovery of submergent macrophytes within the Lower Lakes post-drought (Frahn et al. 2013), so as to inform the future recovery of this region, but also other large lake ecosystems that may be affected by drought around the world. This study was able to identify and quantify the drivers of occurrence for two key submergent species using a non-parametric multiplicative regression (NPMR) modelling technique (McCune 2006). The preliminary individual models developed for each of the target species were then field-tested and refined to include the site- specific variables measured during the field surveys. This project was able to make a significant contribution to the management capability within the Lower Lakes system by quantifying a strong relationship between conductivity, temperature and pH and the occurrence of two key macrophyte species. This study has therefore provided the managers of the Lower Lakes system with a tool enabling them to predict the response of these species under potential future scenarios. Furthermore, this method can now be further applied across a range of taxa within the system and abroad, to enable similar predictive capacity for other key species.
Modelling approach & proof of concept
The NPMR approach was successful in enabling the development of preliminary predictive models for Myriophyllum salsugineum and Vallisneria australis, using only three telemetered predictor variables: conductivity; water temperature; and lake level. The models for M.
salsugineum and for V. australis both had good modelled fits to the species coverage data sets available (M. salsugineum xR2 = 0.656, V. australis xR2 = 0.330). The ability to fit models
