Catchments In Southeast Australia Research Articles

Development of a convolutional neural network based regional flood frequency analysis model for South-east Australia

Flood is one of the worst natural disasters, which causes significant damage to economy and society. Flood risk assessment helps to reduce flood damage by managing flood risk in flood affected areas. For ungauged catchments, regional flood frequency analysis (RFFA) is generally used for design flood estimation. This study develops a Convolutional Neural Network (CNN) based RFFA technique using data from 201 catchments in south-east Australia. The CNN based RFFA technique is compared with multiple linear regression (MLR), support vector machine (SVM), and decision tree (DT) based RFFA models. Based on a split-sample validation using several statistical indices such as relative error, bias and root mean squared error, it is found that the CNN model performs best for annual exceedance probabilities (AEPs) in the range of 1 in 5 to 1 in 100, with median relative error values in the range of 29–44%. The DT model shows the best performance for 1 in 2 AEP, with a median relative error of 24%. The CNN model outperforms the currently recommended RFFA technique in Australian Rainfall and Runoff (ARR) guideline. The findings of this study will assist to upgrade RFFA techniques in ARR guideline in near future.

On the existence of multiple states of low flows in catchments in southeast Australia

This study investigates whether low flow events in a catchment exhibit a possibility of switching regimes. If this happens, it can result in a catchment experiencing a flow anomaly that is even more extreme than what was historically considered a conventional low flow event. A Hidden Markov Modeling approach was used to identify shifts in the intensity of low flow events in catchments in southeast Australia. We found very strong evidence of low flow intensity exhibiting two distinct states (viz., a normal and an intensified state) for several catchments, providing convincing reasons to believe that low flows can and have undergone regime changes. The periods when the catchments exist in an intensified low flow state were associated with higher maximum and average monthly temperature (Δ≈+2.1,+1°C, respectively) and lower precipitation (Δ≈−25mm, a decline of 30%) compared to periods when these exist in their normal state.

Comparing Performance of ANN and SVM Methods for Regional Flood Frequency Analysis in South-East Australia

Design flood estimations at ungauged catchments are a challenging task in hydrology. Regional flood frequency analysis (RFFA) is widely used for this purpose. This paper develops artificial intelligence (AI)-based RFFA models (artificial neural networks (ANN) and support vector machine (SVM)) using data from 181 gauged catchments in South-East Australia. Based on an independent testing, it is found that the ANN method outperforms the SVM (the relative error values for the ANN model range 33–54% as compared to 37–64% for the SVM). The ANN and SVM models generate more accurate flood quantiles for smaller return periods; however, for higher return periods, both the methods present a higher estimation error. The results of this study will help to recommend new AI-based RFFA methods in Australia.

Statistical analysis of attributions of climatic characteristics to nonstationary rainfall‐streamflow relationship

This study demonstrates the benefits of using a simple statistical analysis to investigate the attributions of climate variables to the nonstationary rainfall-streamflow relationships. Rainfall and streamflow data from 65 catchments in southeast Australia are used to quantify the rainfall-streamflow relationship prior to, during, and after, the 1997–2009 Millennium Drought. The results showed that: (1) the annual streamflow is strongly correlated to the annual rainfall, however the relationship is different in the different hydroclimate periods; (2) the number of rainfall days and the mean length of wet-spell (MeWS) are the two most “stable variables” that can best predict annual streamflow in the different hydroclimate periods, followed by the maximum and 99th percentile of 90-day accumulated total rainfall (D90Max and D90P99); (3) the combination of MeWS with several other rainfall characteristics can explain the nonstationary rainfall-streamflow relationship through the different hydroclimate conditions in the eastern part of the study region; and (4) the relative differences in the key rainfall characteristics can help explain the nonstationary rainfall-streamflow relationship. The results here can help identify the key rainfall and climate characteristics to be considered in assessing the impacts of climate variability and change on streamflow, and potentially for developing hydrological models that can be better extrapolated to predict streamflow under a changing climate.

Conceptual Model Modification and the Millennium Drought of Southeastern Australia

Long-term droughts observed in southern Australia have changed relationships between annual rainfall and runoff and tested some of the assumptions implicit in rainfall–runoff models used in these areas. Predictive confidence across these periods is when low using the more commonly used rainfall–runoff models. Here we modified the GR4J model to better represent surface water–groundwater connection and its role in runoff generation. The modified model (GR7J) was tested in 137 catchments in south-east Australia. Models were calibrated during “wetter” periods and simulation across drought periods was assessed against observations. GR7J performed better than GR4J in evaluation during drought periods where bias was significantly lower and showed improved fit across the flow duration curve especially at low flows. The largest improvements in predictive performance were for catchments where there were larger changes in the annual rainfall–runoff relationship. The predictive performance of the GR7J model was more sensitive to objective function used than GR4J. The use of an objective function that combined daily and annual error produced a better goodness of fit when measured against 80, 50 and 20 percent excedance flow quantiles and reduced evaluation bias, especially for the GR7J model.

Evaluation of rainfall–runoff model performance under non-stationary hydroclimatic conditions

ABSTRACT Understanding of rainfall–runoff model performance under non-stationary hydroclimatic conditions is limited. This study compared lumped (IHACRES), semi-distributed (HEC-HMS) and fully-distributed (SWATgrid) hydrological models to determine which most realistically simulates runoff in catchments where non-stationarity in rainfall–runoff relationships exists. The models were calibrated and validated under different hydroclimatic conditions (Average, Wet and Dry) for two heterogeneous catchments in southeast Australia (SEA). SWATgrid realistically simulates runoff in the smaller catchment under most hydroclimatic conditions but fails when the model is calibrated in Dry conditions and validated in Wet. All three models perform poorly in the larger catchment irrespective of hydroclimatic conditions. This highlights the need for more research aimed at improving the ability of hydrological models to realistically incorporate the physical processes causing non-stationarity in rainfall–runoff relationships. Although the study is focussed on SEA, the insights gained are useful for all regions which experience large hydroclimatic variability and multi-year/decadal droughts.

A linked surface water-groundwater modelling approach to more realistically simulate rainfall-runoff non-stationarity in semi-arid regions

Interactions between surface water (SW) and groundwater (GW) have been identified as a major contributing factor to non-stationarity in rainfall-runoff relationships. However, because existing rainfall-runoff models do not realistically account for these SW-GW interactions, they fail to robustly simulate runoff during multi-year droughts, especially in arid and semi-arid catchments that are relatively flat. Therefore, this study introduces a linked SW-GW modelling approach and tests it under different climatic conditions (Average, Dry and Wet periods) in two heterogeneous, semi-arid catchments in southeast Australia (SEA). The linked SW-GW modelling approach includes a fully-distributed SW (i.e. rainfall-runoff) model (SWATgrid) and a three-dimensional finite-difference GW model (MODFLOW). The results show that the linked SW-GW modelling approach produces highly improved runoff simulations for the study catchments, especially during dry conditions. These findings demonstrate the importance of accounting for SW-GW interactions when conducting rainfall-runoff modelling and highlights that failing to do so will result in overestimation of runoff during droughts – this is salient given projections for hotter and drier futures in many semi-arid regions. Although the study is focused on SEA, the insights gained are highly applicable for water managers in any region that experience high hydroclimatic variability or change.

Mechanisms influencing non-stationarity in rainfall-runoff relationships in southeast Australia

Arid and semi-arid regions experience large hydroclimatic variability and decadal-scale droughts. During droughts, the reduced water availability alters the biophysical properties of the catchments which leads to non-stationarity in rainfall-runoff (r-r) relationships. Uncertainty exists regarding the mechanisms causing r-r non-stationarity, especially at seasonal scales. This study analyses endogenous and exogenous catchment mechanisms which could possibly influence r-r relationships in two case study catchments in southeast Australia (SEA). The case study catchments are chosen such that they are representative of arid to semi-arid catchments that experience large climatic variability and decadal-scale droughts. The results demonstrate that endogenous catchment mechanisms (i.e. groundwater table, baseflow and vegetation) have stronger influences on non-stationarity in r-r relationships than exogenous catchment mechanisms like temperature and potential evapotranspiration. These findings highlight the importance of understanding changes in r-r relationships in order to develop more realistic and reliable hydrological models that are needed to quantify and adapt to the impacts of changing catchment or hydroclimatic conditions. While this study is focussed on SEA the findings are relevant to all catchments that experience large climatic variability and decadal-scale droughts.

Uncertainties of statistical downscaling from predictor selection: Equifinality and transferability

The nonhomogeneous hidden Markov model (NHMM) statistical downscaling model, 38 catchments in southeast Australia and 19 general circulation models (GCMs) were used in this study to demonstrate statistical downscaling uncertainties caused by equifinality to and transferability. That is to say, there could be multiple sets of predictors that give similar daily rainfall simulation results for both calibration and validation periods, but project different amounts (or even directions of change) of rainfall changing in the future. Results indicated that two sets of predictors (Set 1 with predictors of sea level pressure north-south gradient, u-wind at 700hPa, v-wind at 700hPa, and specific humidity at 700hPa and Set 2 with predictors of sea level pressure north-south gradient, u-wind at 700hPa, v-wind at 700hPa, and dewpoint temperature depression at 850hPa) as inputs to the NHMM produced satisfactory results of seasonal rainfall in comparison with observations. For example, during the model calibration period, the relative errors across the 38 catchments ranged from 0.48 to 1.76% with a mean value of 1.09% for the predictor Set 1, and from 0.22 to 2.24% with a mean value of 1.16% for the predictor Set 2. However, the changes of future rainfall from NHMM projections based on 19 GCMs produced projections with a different sign for these two different sets of predictors: Set 1 predictors project an increase of future rainfall with magnitudes depending on future time periods and emission scenarios, but Set 2 predictors project a decline of future rainfall. Such divergent projections may present a significant challenge for applications of statistical downscaling as well as climate change impact studies, and could potentially imply caveats in many existing studies in the literature.

The frequency of major flooding in coastal southeast Australia has significantly increased since the late 19th century

Millions of Australians live in a 1500 km stretch of coastal catchments in south-east Australia. Major flooding in this region causes death, economic loss and major disruptions to the lives of its inhabitants. Concerns have been raised that anthropogenic climate change might lead, or has already led, to an increased risk of extreme rainfall and associated flooding. Images of flooding commonly appear in the media, fuelling perceptions that flood frequency has already in-creased. Here we use a new dataset that allows us to estimate reliable trends over much longer periods than has previously been the case. The statistical significance of the trends is assessed using a method that is suitable for the non-Gaussian, serially correlated flood frequency data. We identify a statistically significant, increasing trend in the frequency of major floods since the late 19th century, which contributes to a 50% increase in frequency. While possible reasons for the increase are discussed (e.g. land use change, anthropogenic climate change, natural climate variability), further research is needed to clarify the relative importance of possible contributors.

The frequency of major flooding in coastal southeast Australia has significantly increased since the late 19th century

Millions of Australians live in a 1500 km stretch of coastal catchments in south-east Australia. Major flooding in this region causes death, economic loss and major disruptions to the lives of its inhabitants. Concerns have been raised that anthropogenic climate change might lead, or has already led, to an increased risk of extreme rainfall and associated flooding. Images of flooding commonly appear in the media, fuelling perceptions that flood frequency has already in-creased. Here we use a new dataset that allows us to estimate reliable trends over much longer periods than has previously been the case. The statistical significance of the trends is assessed using a method that is suitable for the non-Gaussian, serially correlated flood frequency data. We identify a statistically significant, increasing trend in the frequency of major floods since the late 19th century, which contributes to a 50% increase in frequency. While possible reasons for the increase are discussed (e.g. land use change, anthropogenic climate change, natural climate variability), further research is needed to clarify the relative importance of possible contributors.

Impact of bushfire and climate variability on streamflow from forested catchments in southeast Australia

This paper quantifies the impacts of bushfire and climate variability on streamflow from three southeast Australian catchments where bushfires occurred in February 1983. Three hydrological models (AWRA-L, Xinanjiang and GR4J) were first calibrated against streamflow data from the pre-bushfire period and then used to simulate runoff for the post-bushfire period with the calibrated parameters. The difference in simulated streamflow between pre- and post-bushfire periods provides an estimate of the impact of climate variability on streamflow. The impact of bushfire on streamflow is quantified by removing the climate variability impact from the difference in mean annual observed streamflow between post- and pre-bushfire periods. For the first 15 years after the 1983 bushfires, the results from hydrological models for the three catchments indicate that there is a substantial increase in streamflow; this is attributed to initial decreases in evapotranspiration and soil infiltration rates resulting from the fires, followed by logging activity. After 15 years, streamflow dynamics are more heavily influenced by climate effects, although some impact from fire and logging regeneration may still occur. The results show that hydrological models provide reasonably consistent estimates of bushfire and climate impacts on streamflow for the three catchments. The models can be used to quantify relative contributions of forest disturbance (bushfire, logging and other forest management) and climate variability. The results presented can also help forest managers understand the relationship between bushfire and climate variability impacts on water yield in the context of climate variability.

Short-Term Forecasting of Water Yield from Forested Catchments after Bushfire: A Case Study from Southeast Australia

Forested catchments in southeast Australia play an important role in supplying water to major cities. Over the past decades, vegetation cover in this area has been affected by major bushfires that in return influence water yield. This study tests methods for forecasting water yield after bushfire, in a forested catchment in southeast Australia. Precipitation and remotely sensed Normalized Difference Vegetation Index (NDVI) were selected as the main predictor variables. Cross-correlation results show that water yield with time lag equal to 1 can be used as an additional predictor variable. Input variables and water yield observations were set based on 16-day time series, from 20 January 2003 to 20 January 2012. Four data-driven models namely Non-Linear Multivariate Regression (NLMR), K-Nearest Neighbor (KNN), non-linear Autoregressive with External Input based Artificial Neural Networks (NARX-ANN), and Symbolic Regression (SR) were employed for this study. Results showed that NARX-ANN outperforms other models across all goodness-of-fit criteria. The Nash-Sutcliffe efficiency (NSE) of 0.90 and correlation coefficient of 0.96 at the training-validation stage, as well as NSE of 0.89 and correlation coefficient of 0.95 at the testing stage, are indicative of potentials of this model for capturing ecological dynamics in predicting catchment hydrology, at an operational level.

How does bias correction of regional climate model precipitation affect modelled runoff?

Abstract. Many studies bias correct daily precipitation from climate models to match the observed precipitation statistics, and the bias corrected data are then used for various modelling applications. This paper presents a review of recent methods used to bias correct precipitation from regional climate models (RCMs). The paper then assesses four bias correction methods applied to the weather research and forecasting (WRF) model simulated precipitation, and the follow-on impact on modelled runoff for eight catchments in southeast Australia. Overall, the best results are produced by either quantile mapping or a newly proposed two-state gamma distribution mapping method. However, the differences between the methods are small in the modelling experiments here (and as reported in the literature), mainly due to the substantial corrections required and inconsistent errors over time (non-stationarity). The errors in bias corrected precipitation are typically amplified in modelled runoff. The tested methods cannot overcome limitations of the RCM in simulating precipitation sequence, which affects runoff generation. Results further show that whereas bias correction does not seem to alter change signals in precipitation means, it can introduce additional uncertainty to change signals in high precipitation amounts and, consequently, in runoff. Future climate change impact studies need to take this into account when deciding whether to use raw or bias corrected RCM results. Nevertheless, RCMs will continue to improve and will become increasingly useful for hydrological applications as the bias in RCM simulations reduces.

Mapping spatial and temporal variation in tree water use with an elevation model and gridded temperature data

Tree water use is a major component of the water balance in forested catchments of semi-arid areas, as more than 80% of the incoming rainfall may be used by overstory trees. Managers are unable to easily predict water use and thus water yield, for the majority of eucalypt-dominated catchments in south-east Australia, owing to the variety of dominant and co-dominant species, their distributions with respect to landform, and the lack of species- and landform-specific knowledge of the regulation of water use. Moreover, the costs incurred to quantify input variables for available complex, process-based models, generally encourage finding alternative approaches. This study tested the adequacy of using just two easily measured variables for estimating rates of tree water use, using a model derived from data-learning techniques. The inputs are (1) measured daily atmospheric demand for water and (2) potential incoming radiation derived from surface topography and solar declination. Artificial neural networks (ANNs) and genetic programming (GP) models were trained and validated using in situ observations of vapour pressure deficit (VPD) and estimates of potential solar radiation (Qpot), for a period of two years, at each of 10 forest stands across the high country of the states of New South Wales and Victoria. The models were tested using a random 50% of the collected data that was independent, i.e. not used in model development.Atmospheric demand was selected because it strongly affects tree water use irrespective of site and species. Potential solar radiation was selected as a proxy for radiation, because it is relatively easy to estimate for any location for which elevation data are available in digital format, and since radiation strongly controls photosynthesis (through stomatal behaviour) and thermal balance.Genetic programming resulted in models better able to predict rates of sap flux. A selected GP model was able to describe the relationship between tree sap flux, VPD, and potential radiation with good accuracy, and was used to map tree water use across the catchment.

A similarity index for storm runoff due to saturation excess overland flow

• We propose and demonstrate an index for similarity in the storm event response of catchments to rainfall. • The similarity index is based on the steady state assumption. • The index is applicable to catchments where runoff is generated by saturation excess overland flow. • The index was applied to 236 independent storm events on eight catchments in southeast Australia. • The index’s effectiveness was assessed by comparing its groupings with those obtained from a flood frequency analysis. An index for the determination of hydrologic similarity is proposed and demonstrated. The index is based on the steady state assumption and is applicable for small- to medium-sized catchments where storm runoff is generated principally by the saturation excess mechanism – a key runoff generation mechanism in humid regions. The index uses variables that can be derived from rainfall and streamflow measurements and topographic and soil hydraulic attributes. The index is applied to eight gauged catchments located in southeast Australia. Comparisons of similarity index values with groupings obtained from a peaks-over-threshold (POT) analysis of daily maxima of hourly runoff series show good agreement. A sensitivity analysis of the index and the POT series indicated that these results are reasonably robust provided catchments are allowed to have partial memberships in all of the groupings identified.

Estimating water quality using linear mixed models with stream discharge and turbidity

Most water quality monitoring schemes rely on estimation methods as it is often far too expensive to monitor water quality properties continuously. Estimations are used to evaluate management strategies and long term trends. It is critical that the estimation methods provide accurate estimations and an accurate estimate of the associated uncertainty. Currently the most common estimation methods assume observations are sampled using a probabilistic sampling scheme, however this assumption is often not met. This paper evaluated the ability of a linear mixed model to estimate water quality concentration values based on observations collected using non-probabilistic sampling. The linear mixed models were used to predict total phosphorus and total nitrogen observations from two catchments in south east Australia. A comparison between stream discharge and turbidity as predictors is made to investigate the effectiveness of turbidity to estimate water quality. In addition to stream discharge and turbidity, several covariates were derived from stream discharge in an attempt to account for hydrological processes. To compare models and their covariates leave one out event cross validation was performed. Event cross validation evaluated predictions during periods of high stream discharge. The inclusion of temporal auto-correlation component improved the accuracy of all models for total phosphorus and total nitrogen. For both catchments the use of turbidity instead of stream discharge increased the accuracy of predictions by at least 15% for total phosphorus and total nitrogen. However, event based cross validation indicated that a combination of both turbidity and stream discharge based variables provided more accurate predictions, decreasing the event RMSE by 18% for total phosphorus and 24% for total nitrogen. In catchments characterised by long periods of base-flow and short rainfall events the addition of turbidity measurements provide more accurate predictions during base flow and during events.

Ensemble dressing for hydrological applications

AbstractThis manuscript presents a simple yet effective method to account for uncertainty in hydrologic ensemble forecasting applications. Most operational hydrological ensemble forecasting systems only account for uncertainty in future climate (e.g. precipitation) forcings, ignoring other sources of uncertainty (e.g. model error). The result can be under‐dispersive and overconfident forecasts. Ensemble dressing is a form of statistical post‐processing to include information about the uncertainty of individual ensemble members. First, the historical simulations can be corrected to remove systematic biases. Next, the simulated and observed flow data are transformed to ensure that model residuals can be fitted to a distribution (e.g. normal); this study uses a 2‐parameter log‐sinh transformation. Model residuals in transformed space determine the width of the error distribution. Ensemble forecasts are then generated and transformed. Each ensemble is dressed with the error distribution, the results are untransformed, and finally, a probabilistic forecast is derived from the collective distribution of the dressed ensembles. The method is applied to 128 catchments in southeast Australia, demonstrating that the raw ensemble can be made reliable through the use of this dressing method. Copyright © 2012 John Wiley & Sons, Ltd.

Crash testing hydrological models in contrasted climate conditions: An experiment on 216 Australian catchments

This paper investigates the actual extrapolation capacity of three hydrological models in differing climate conditions. We propose a general testing framework, in which we perform series of split‐sample tests, testing all possible combinations of calibration‐validation periods using a 10 year sliding window. This methodology, which we have called the generalized split‐sample test (GSST), provides insights into the model's transposability over time under various climatic conditions. The three conceptual rainfall‐runoff models yielded similar results over a set of 216 catchments in southeast Australia. First, we assessed the model's efficiency in validation using a criterion combining the root‐mean‐square error and bias. A relation was found between this efficiency and the changes in mean rainfall (P) but not with changes in mean potential evapotranspiration (PE) or air temperature (T). Second, we focused on average runoff volumes and found that simulation biases are greatly affected by changes in P. Calibration over a wetter (drier) climate than the validation climate leads to an overestimation (underestimation) of the mean simulated runoff. We observed different magnitudes of these models deficiencies depending on the catchment considered. Results indicate that the transfer of model parameters in time may introduce a significant level of errors in simulations, meaning increased uncertainty in the various practical applications of these models (flow simulation, forecasting, design, reservoir management, climate change impact assessments, etc.). Testing model robustness with respect to this issue should help better quantify these uncertainties.

Climate non-stationarity – Validity of calibrated rainfall–runoff models for use in climate change studies

This paper presents results from a modelling study carried out to investigate whether the calibrated parameter values for rainfall–runoff models based on historical observed data can be used to reliably predict runoff responses to changes in future climate inputs. The modelling experiments are carried out for four rainfall–runoff models using long records of historical daily climate and streamflow data from 61 catchments in southeast Australia. The results indicate that the models, when calibrated using more than 20 years of data, can generally be used for climate impact studies where the future mean annual rainfall is not more than 15% drier or 20% wetter than the mean annual rainfall in the model calibration period. It is generally more difficult for a model calibrated over a wet period to predict runoff over a dry period compared to a model calibrated over a dry period to predict runoff over a wet period. For southeast Australia, there is a good reason to use the recent records to calibrate rainfall–runoff models to represent the current prolonged drought over the region and for climate change impact studies where the large majority of climate models predict a drier future across this region.