Search

In recent years, deep learning (DL) approaches have been proven effective in addressing high nonlinear relationships within complex systems. Although various scientific studies have primarily focused on optimizing model architecture and enhancing its computational efficiency through hybridization, they have neglected the crucial aspect of selecting appropriate predictor variables (PV) and their influence on the model's predictions. In this paper, a new multi-phase stochastic DL model is designed with a feature selection framework for monthly streamflow prediction. The multi-phase hybrid MCMC–BC–BiLSTM–BiGRU model is based on bidirectional long short-term memory (BiLSTM) and bidirectional gated recurrent unit (BiGRU) integrated with Markov Chain Monte Carlo (MCMC) based bivariate copulas (BC). First, Ant Colony Optimization (ACO) algorithm was employed for screening optimal PV, and then significant antecedent streamflow was coupled with it to generate optimized input combinations. In addition, 25 different MCMC–BC models were adopted for each input combination to determine the dependency of previous month's streamflow on current and future streamflow. Finally, best-fitted MCMC–BC model was integrated with hybrid BILSTM–BiGRU model to predict streamflow at nine different catchments in the Victoria region of the Upper Murray Basin (UMB), Australia. The experimental results show that the multi-phase hybrid MCMC–BC–BiLSTM–BiGRU model showed remarkable performance in comparison to its benchmark counterparts with respect to different robust statistical error metrics. Therefore, this study demonstrates that stochastic DL techniques can be effectively employed as a promising alternative predictive tool for upstream discharge predictions with high consistency and accuracy, specifically when a statistically significant relationship with the antecedent streamflow exists. [ABSTRACT FROM AUTHOR]

A large literature on the sensitivity of streamflow response to climate variations has emerged over the past several decades, but the underlying mechanisms are not fully understood. The usual approaches to this problem have been simple and do not fully address its complexity, which involves the individual and joint effects of large‐scale climate modes and smaller‐scale weather features on streamflow volumes, the presence of persistence in these time series and their effects on antecedent conditions. Ongoing improvements in observational and reanalysis datasets and online access allow for better quantification of the connections between streamflow response, climate variability modes and regional weather features. The purpose of this paper is to determine whether Bayesian networks can be used to better identify key factors and their associated pathways leading to streamflow generation. A Markov blanket approach is described and illustrated using monthly streamflow series recorded at eight gauging stations within or immediately adjacent to the coastal region of eastern mainland Australia. The method is compared and contrasted with conventional multiple linear regression with discussion around this type of approach as part of a growing interest in machine learning applications. One example is used to illustrate the application of both approaches to a streamflow series exhibiting strong non‐stationarity. Results for the other streamflow series are included in a concise synthesis. Overall, the results suggest that Bayesian networks have several desirable features in terms of transparency, interpretability and explanatory insight. The findings from this study lend support to the use of Bayesian networks for modelling connections between streamflow volume and the variability of climate and regional weather. This improved understanding of the key controls of streamflow variability is intended to help address growing needs around informing social, cultural, economic and ecological aspects of water planning and management. [ABSTRACT FROM AUTHOR]

The Murray-Darling River system is perhaps Australia's most important, with significant social, cultural and environmental values including 16 Ramsar listed wetlands. The MDB is home to 2.6 million people and produces about $24 billion worth in agricultural production each year (about one-third of total value for Australia). Hydrologic issues, typified by water availability and quality, have existed for many years, peaking during the Millennium drought from 1997 to 2010. Competing interests (i.e. irrigation, tourism, environmental heath), and the declining flows and water quality during droughts, led governments and water management agencies to consider the risks to water resources in the system in the early-mid 2000s. This paper reviews changes to risks associated with forest dynamics, as identified by - afforestation and bushfire – and considers new issues that have emerged since that analysis. It was found that the potential impacts of bushfire on stream flows were over-estimated in past studies, and that a planned significant afforestation expansion into agricultural and grazing land that was projected to reduce stream flows did not occur. While these two risks now do not seem likely to have significant future impacts on flows, or consequent effects on downstream users, the interaction of elevated CO2 and increasing temperatures on vegetation functioning and subsequent hydrologic consequences at catchment scale require further research and analysis. Reduced rainfall and increased temperatures under future climate change are likely to have an impact on inputs and flows. Uncertainties in how these changes, and feedbacks between climate, drought, more frequent fire and vegetation responses, impact on system hydrology also require further investigation. [ABSTRACT FROM AUTHOR]

Dominant hydrological processes of a catchment could shift due to a changing climate. This climate-induced hydrological nonstationarity could affect the reliability of future runoff projection developed using a hydrological model calibrated for the historical period as the model or parameters may no longer be suitable under a different future hydroclimate. This paper explores whether competing parameterization approaches proposed to account for hydrological nonstationarity could improve the robustness of future runoff projection compared to the traditional approach where the model is calibrated targeting overall model performance over the entire historical period. The modeling experiments are carried out using climate and streamflow datasets from southeastern Australia, which has experienced a long drought and exhibited noticeable hydrological nonstationarity. The results show that robust multicriteria calibration based on the Pareto front can provide a more consistent model performance over contrasting hydroclimate conditions, but at a slight expense of increased bias over the entire historical period compared to the traditional approach. However, the robust calibration does not necessarily result in a more reliable projection of future runoff. This is because the systematic bias in any parameterization approach would propagate from the historical period to the future period and would largely be cancelled out when estimating the relative runoff change. Ensemble simulations combining results from different parameterization considerations could produce a more inclusive range of future runoff projection as it covers the uncertainties due to model parameterization. [ABSTRACT FROM AUTHOR]

Although traditional physical models have been used to analyse groundwater systems, the emergence of novel machine learning models can improve the accuracy of the predictions. Deep learning has been prominent in environmental and climate change problems. In this paper, we present a framework for utilising deep learning models to predict groundwater levels based on nearby streamflow and rainfall data. We address the missing data problem using a Bayesian linear regression model within the deep learning framework. Our deep learning framework utilises models such as long-short term memory (LSTM) networks and convolutional neural networks (CNN) for multi-step ahead time series prediction. We examine the fluctuations in groundwater levels at various boreholes located near Middle Creek in New South Wales, Australia. We use the National Collaborative Research Infrastructure Strategy (NCRIS) groundwater database and utilise Bayesian linear regression to impute missing data. We investigate the accuracy of the selected models for individual and regional basins and univariate and multivariate strategies. Our results show that the LSTM-based regional model with multivariate strategy using rainfall data provided the best accuracy. • We predict groundwater levels based on stream flow and rainfall data. • Our framework uses Bayesian data imputation and deep learning for multi-step prediction. • We examine the groundwater fluctuations at various boreholes located in Australia. • Results show a fine-tuned regional model with rainfall data yielded the best accuracy. • We find that LSTM models outperformed CNNs using the multivariate approach. [ABSTRACT FROM AUTHOR]

• Nationwide streamflow projections for Australia informed by CMIP5 and CMIP6. • Vast majority of CMIP5 and CMIP6 GCMs project less winter rainfall across Australia. • Lower winter rainfall translates to a significant reduction in annual runoff. • Modelled median projection of runoff shows a 50% reduction in far southwest Australia. • Modelled median projection of runoff shows a 20% reduction in far southeast Australia. Projections of streamflow under future climate are essential for developing adaptation strategies in the water and related sectors. This paper presents nationwide streamflow projections for Australia informed by climate change signals in CMIP6 GCMs and compares the projections with those from CMIP5 GCMs. The modelled future runoff projections driven by CMIP5 and CMIP6 GCMs are relatively similar for far southern Australia. The median projection is a 50% reduction in mean annual runoff in far southwest Australia and 20% reduction in far southeast Australia by 2046–2075 relative to 1976–2015 under the high SSP5-8.5/RCP8.5 global warming scenario. The vast majority of CMIP5 and CMIP6 GCMs project less winter rainfall across Australia. As most of the runoff in far southern Australia occurs in winter and spring, the lower winter rainfall translates to a significant reduction in annual runoff. It is therefore prudent to plan and manage for a reduction in future water resources, particularly in the densely populated and important agricultural regions in southeast Australia. There is less agreement between GCMs in the summer rainfall projection, with the CMIP6 GCMs generally projecting a wetter future (or smaller decrease in rainfall) than the CMIP5 GCMs. The modelled median projection for mean annual runoff is a 10% reduction in the south-east coast, 5% reduction in the north-east coast and little change in northern Australia. More GCMs project an increase rather than a decrease in the interannual variability of rainfall, and this would further amplify multi-year hydrological droughts. [ABSTRACT FROM AUTHOR]

This paper presents the Australian edition of the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) series of datasets. CAMELS-AUS (Australia) comprises data for 222 unregulated catchments, combining hydrometeorological time series (streamflow and 18 climatic variables) with 134 attributes related to geology, soil, topography, land cover, anthropogenic influence and hydroclimatology. The CAMELS-AUS catchments have been monitored for decades (more than 85 % have streamflow records longer than 40 years) and are relatively free of large-scale changes, such as significant changes in land use. Rating curve uncertainty estimates are provided for most (75 %) of the catchments, and multiple atmospheric datasets are included, offering insights into forcing uncertainty. This dataset allows users globally to freely access catchment data drawn from Australia's unique hydroclimatology, particularly notable for its large interannual variability. Combined with arid catchment data from the CAMELS datasets for the USA and Chile, CAMELS-AUS constitutes an unprecedented resource for the study of arid-zone hydrology. CAMELS-AUS is freely downloadable from 10.1594/PANGAEA.921850 (Fowler et al., 2020a). [ABSTRACT FROM AUTHOR]

This paper presents the Australian edition of the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) series of datasets. CAMELS-AUS comprises data for 222 unregulated catchments, combining hydrometeorological timeseries (streamflow and 18 climatic variables) with 134 attributes related to geology, soil, topography, land cover, anthropogenic influence, and hydroclimatology. The CAMELS-AUS catchments have been monitored for decades (more than 85 % have streamflow records longer than 40 years) and are relatively free of large scale changes, such as significant changes in landuse. Rating curve uncertainty estimates are provided for most (75 %) of the catchments and multiple atmospheric datasets are included, offering insights into forcing uncertainty. This dataset, the first of its kind in Australia, allows users globally to freely access catchment data drawn from Australia's unique hydroclimatology, particularly notable for its large interannual variability. Combined with arid catchment data from the CAMELS datasets for the USA and Chile, CAMELS-AUS constitutes an unprecedented resource for the study of arid-zone hydrology. CAMELS-AUS is freely downloadable from https://doi.pangaea.de/10.1594/PANGAEA.921850 (Fowler et al., 2020a). [ABSTRACT FROM AUTHOR]

Accurately predicting river flows over daily timescales is considered as an important task for sustainable management of freshwater ecosystems, agricultural applications, and water resources management. In this research paper, artificial intelligence (AI) techniques, namely the cascade correlation neural networks (CCNN) and the random forest (RF) models, were employed in daily river stage and river flow prediction for two river systems (i.e., Dulhunty River and Herbert River) in Australia. To develop the CCNN and RF models, a significant 3-day antecedent river stage and river flow time series were used. 80% of the whole data were used for model training and the remaining 20% for model testing. A total of ten different model structures with different input combinations were used to evaluate the optimal model in the training phase, and the results were analyzed using statistical metrics including the root mean square error (RMSE), Nash–Sutcliffe coefficient (NS), Willmott's index of agreement (WI), and Legate and McCabe's index (ELM) in the testing phase. The inter-comparison of CCNN and RF models for both river systems showed that the CCNN model was able to generate a more accurate prediction of the river stage and river flow compared to the RF model. Due to hydro-geographic differences leading to a different underlying historical data characteristics, the optimal CCNN's performance for the Dulhunty River was found to be most accurate, in terms of ELM = 0.779, WI = 0.964, and ENS = 0.862 versus 0.775, 0.968, and 0.885 for the Herbert River. Following the performance accuracies, the authors ascertained that the CCNN model can be taken as a preferred data intelligent tool for river stage and river flow prediction. [ABSTRACT FROM AUTHOR]

• Daily flow statistics (indicators) are key to understanding flow-ecology interaction. • This study examined whether monthly data can be used to describe daily flow regimes. • A suite of 173 flow indicators were converted to a monthly timestep where possible. • Around half of monthly indicators had strong correlations with their daily versions. • Using only monthly data, good predictions were achieved for 90% of daily indicators. Flow indicators based on daily flow data are commonly used to characterize the hydrologic regimes of rivers and to investigate links between flow and ecological condition. Discharge data measured at a daily timestep is considered essential for such investigations because there is an underlying premise that longer timesteps (e.g. monthly) are not suitable for capturing the variability in flow behaviour of most relevance to ecologically important processes. This paper investigates this assumption by examining the information content of daily and monthly streamflow data for a diversity of perennial, intermittent and ephemeral flow regimes across Australia, using a novel approach of comparing daily and monthly flow indicators rather than observed flow timeseries. Results show that many flow indicators derived using daily and monthly data are strongly correlated (Spearman ρ greater than 0.9). Partial least squares regression shows that the majority of daily flow indicators can be predicted using monthly flow indicators, and that as few as six monthly flow indicators can be used to estimate a wide range of daily flow characteristics, including those relating to flow magnitudes, and durations of high and low flow events. Daily indicators relating to the frequency and seasonal timing of high and low flows cannot be predicted as accurately (R2 between 0.4 and 0.9), suggesting that information on these facets of the flow regime is partially lost when aggregating to a monthly timestep. Results from this study point to the utility of monthly-based flow indicators in river ecology, opening new opportunities to overcome limitations in gauge network coverage and data resolution in many parts of the world. [ABSTRACT FROM AUTHOR]

Modelling workflow for BRF-LSTM and BRF-GRU predictive models. [Display omitted] • Deep learning hybrid predictive model (BRF-LSTM) is built for streamflow forecasting. • Climate indices, rainfall and periodicity are factored for accurate predictions. • Adopts feature algorithm (i.e. Boruta) with long short-term memory and gated recurrent unit. • Six gauging sites tested; BRF-LSTM yields > 98% errors in ± 0.015 m, ~1.30% relative error. • The proposed model is useful for hydrological and strategic water resources planning. Long-term forecasting of any hydrologic phenomena is essential for strategic environmental planning, hydrologic and other forms of structural design, agriculture, and water resources management. Climate mode indices, utilising machine learning methods, are frequently considered as predictor variables in order to forecast several different hydrological variables. In this study, a feature selection algorithm based on two different deep learning models, i.e., long short-term memory and a gated recurrent unit, is applied to improve the forecasting capability of streamflow water levels at six gauging stations in the Murray Darling Basin of Australia. This paper therefore aggregates the significant antecedent lag memory of climate mode indices, rainfall, and the monthly factor based on the periodicity as the predictor variables to attain significantly accurate stream water level forecasts. This novel method identifies an improved relationship between the stream water level and climate mode indices through the aggregation of the significant lagged datasets capturing the historical features to predict the future streamflow water level. The boruta feature selection algorithm (BRF) was then applied in a two phase process before and after attaining the significant lagged inputs to screen the optimum predictor variables. The merits of the forecast models were evaluated through different performance evaluation criteria. The results show that the accumulated significant lagged inputs based on climate mode indices, along with the rainfall and periodicity factors are seen to provide improved forecasting of the SWL over the non-BRF deep learning approaches where no prior feature selection was applied. The hybrid LSTM method (i.e., BRF-LSTM model) achieved a unique advantage in terms of SWL forecasting, particularly attaining over 98% of the predictive errors lying within a band of +/-0.015 m with relatively low relative errors (RRMSE ≈1.30% and RMAE ≈ 0.882%), outperforming all of the benchmark models. It is also found that the periodicity factor has a potential influence on the accuracy of the forecast models for the four monitored study stations. This study concludes that the newly developed hybrid deep learning approaches, coupled with the BRF feature selection, provide improved forecasting performance. The hybrid approach developed in this paper can therefore be used to provide a strong provide predictive response algorithm for the hydrological variables that were influenced by the low-frequency variability of the climate model indices in respect to streamflow water level. [ABSTRACT FROM AUTHOR]

[Display omitted] • Wildfires are becoming more severe and frequent and their effects increasingly expose water resources to stress. • Climate, wildfire, streamflow records (1975 to 2018), topography, and landcover data from 92 catchments were combined to quantify the contribution of wildfire to streamflow variability across Australian temperate forests. • On average wildfire regimes relatively results in 8.8% increases in streamflow across Australian temperate forests. • Wildfire effects on regional streamflow varied spatially depending on hydroclimate factors, particularly precipitation. As wildfires become more frequent and severe, there are concerns regarding their impacts on water yield from forested catchments. While there are many studies in Australia about the effects of individual wildfires on streamflow at fine scales (<1 km2) in specific geographic settings, the effects of wildfire regimes on streamflow at broad spatial scales across temperate forests in Australia are not well understood. In this paper, we combined climate, wildfire, streamflow records (1975 to 2018), topographic, and landcover data from 92 catchments (74 – 4740 km2) in the Australian temperate zone to quantify the contribution of wildfire regimes over time to streamflow variability in different hydroclimatic settings (humid, dry sub-humid, and semi-arid) and geographic regions (Southeast Australia (SEA), Southwest Western Australia (SWWA), and Tasmania (TAS)). Wildfire regimes were represented by two metrics: the burnt area to drainage area (BDA) ratio for wildfire events in each year and a spatially averaged metric of Time Since Fire (TSF), which is a spatial average of time since the last wildfire in the catchment. By comparing prefire and postfire runoff ratios our study found that on average there was a short-term increase in runoff ratio (∼3% in year 1 and ∼6% in year 2 post-fire) after wildfires with BDA > 25%. No influence of fire was found in the long-term (15–20 years after the wildfire) runoff ratio. We found that wildfire regime, measured by TSF, explained ∼8.8% of the variation in annual streamflow across the Australian temperate zone, and that with decreasing TSF (i.e., increased wildfire impact), the average streamflow increased. Streamflow variation explained by wildfire regimes varied with hydroclimate. The explained variance of streamflow by wildfire regimes in semi-arid catchments (23%) and dry sub-humid catchments (13%) were higher than humid catchments (5%). Our results provide a broad-scale understanding of how wildfire regimes influence streamflow variability at broad temporal and spatial scales, and provided important context and baseline information for determining the implications of changes in climate and fire regimes for regional water availability. [ABSTRACT FROM AUTHOR]

This paper presents the results from a study on the application of an artificial neural network (ANN) model for regional flood frequency analysis (RFFA). The study was conducted using stream flow data from 88 gauging stations across New South Wales (NSW) in Australia. Five different models consisting of three to eight predictor variables (i.e., annual rainfall, drainage area, fraction forested area, potential evapotranspiration, rainfall intensity, river slope, shape factor and stream density) were tested. The results show that an ANN model with a higher number of predictor variables does not always improve the performance of RFFA models. For example, the model with three predictor variables performs considerably better than the models using a higher number of predictor variables, except for the one which contains all the eight predictor variables. The model with three predictor variables exhibits smaller median relative error values for 2- and 20-year return periods compared to the model containing eight predictor variables. However, for 5-, 10-, 50- and 100-year return periods, the model with eight predictor variables shows smaller median relative error values. The proposed ANN modelling framework can be adapted to other regions in Australia and abroad. [ABSTRACT FROM AUTHOR]