Time-series modelling of harmful cyanobacteria blooms by convolutional neural networks and wavelet generated time-frequency images of environmental driving variables.

Early warning systems for harmful cyanobacterial blooms (HCBs) that enable precautional control measures within water bodies and in water works are largely based on inferential time-series modelling. Among deep learning techniques, convolutional neural networks (CNNs) are widely applied for recognition of pictorial, acoustic and thermal images. Time-frequency images of environmental drivers generated by wavelets may provide crucial signals for modelling of HCBs to be recognized by CNNs. This study applies CNNs for time-series modelling of HCBs of Microcystis sp. in four South Korean rivers between 2016 and 2022 by means of time-frequency images of environmental drivers within the lead time of HCBs. After estimating the cardinal dates of beginning, peak, and ending of HCBs, wavelet analysis identified key drivers by phase analysis and generated time-frequency images of the drivers within the cardinal dates for 3, 4 and 5 years. Performances of CNNs were compared in terms of four determinants of input images: methods of estimating critical timings, the number of segments, time-series continuity, and image size. The resulting CNNs predicted high or low intensities of HCBs with a mean accuracy of 97.79 ± 0.06% and F1-score 97.49 ± 0.06% for training dataset, and a mean accuracy of 95.01 ± 0.06% and F1-score 93.30 ± 0.07% for testing dataset. Predictions of Microcystis abundances by CNNs achieved a mean MSE of 2.58 ± 2.46 and a mean R2 of 0.78 ± 0.20 for training, and a mean MSE of 2.76 ± 2.42 and a mean R2 of 0.55 ± 0.20 for testing dataset. Precipitation and discharge appeared to be the best performing drivers for qualitative and quantitative predictions of HCBs pointing at the nonstationary nature of river habitats. This study highlights the opportunities of time-series modelling by CNNs driven by wavelet generated time-frequency images of key environmental variables for forecasting of HCBs.

Dissolved organic nitrogen, dinoflagellates and cyanobacteria in two eutrophic lakes: Analysis by inferential modelling.

The Lake Suwa (Japan) has a history of non-N-fixing Microcystis blooms. Lake Kinneret (Israel) experienced multiannual periods of sole domination by the dinoflagellate Peridinium gatunense and periods dominated seasonally by P. gatunense or cyanobacteria. Extensive studies have been carried out in both lakes regarding the role of dissolved inorganic nitrogen and phosphorus as drivers of primary productivity. There is growing evidence that dissolved organic nitrogen (DON) compounds also influence not only biomass and structure of phytoplankton communities but also microcystin production. This study focuses on relationships of DON with: (1) population dynamics of Microcystis spp. and concentrations of microcystins in Lake Suwa, and (2) population dynamics of P. gatunense as well as N- and non-N-fixing cyanobacteria in Lake Kinneret. Modelling results for historical data of Lake Suwa by means of the hybrid evolutionary algorithm HEA revealed that the prediction of abundances of four Microcystis species and concentrations of cyanotoxins achieved higher coefficients of correlation when DON/DIN-ratios were included as drivers. Population dynamics of P. gatunense in Lake Kinneret appeared to have a strong inverse relationships with DON/DIN-ratios reflected by inferential models of HEA with higher coefficients of correlation when driven by DON/DIN-ratios. When DON/DIN-ratios were included as drivers, models of Microcystis spp. in Lake Kinneret performed higher coefficients of determination compared to models of N-fixing cyanobacteria. The study highlights the need to consider DON for improved understanding and management of population dynamics of cyanobacteria and dinoflagellates in freshwater lakes.

Implications of flow regulation for habitat conditions and phytoplankton populations of the Nakdong River, South Korea.

Anthropogenic regulation of hydrographs is a widespread approach to river management; however, the effects of river regulation on habitat conditions and aquatic communities have rarely been studied. In this study, we analyzed the physical, chemical, and biological data from the lower Nakdong River in South Korea from 2005 to 2009 before weir construction and from 2012 to 2016 after weir construction. A partial least square path model (PLS-PM) was applied to delineate the complex interrelationships of diatoms and cyanobacteria with physicochemical parameters, nutrients, zooplankton grazing, and hydrological parameters. Inferential modeling using the hybrid evolutionary algorithm (HEA) allowed the identification of differences in the importance and threshold conditions of population dynamics drivers of diatoms and cyanobacteria before and after flow regulation. The annually averaged trajectories of limnological variables displayed significant shifts in seasonality and magnitudes of phytoplankton, zooplankton, and nutrient concentrations between the two periods. The results of PLS-PM indicated that, after flow regulation, diatoms and cyanobacteria were directly affected by nutrients and zooplankton densities and the path coefficients of hydrological parameters decreased or even were insignificant. The inferential models suggested that diatom dynamics were essentially shaped by threshold conditions of water temperature (WT) and pH before regulation, but mainly by those of rotifers (below 51.1 ind. L-1) after regulation. As for cyanobacteria dynamics, WT was identified as a critical threshold condition before and after regulation, and the threshold of PO4- concentration above 145.4 L-1 was identified as the reason for occasional blooms during the post-regulation period. Overall, the results suggest that flow regulation gradually alters habitat conditions typically of rivers to those of stagnant waters. These findings must be taken into account for sustainable management strategies of regulated rivers.

Comparison of two commonly used methods for identifying water quality thresholds in freshwater ecosystems using field and synthetic data.

Defining ecological thresholds has become increasingly relevant for water resource management. Despite the fact that there has been a rapid expansion in methods to evaluate ecological threshold responses to environmental stressors, evaluation of the relative benefits of various methods has received less attention. This study compares the performance of Gradient Forest (GF) and Threshold Indicator Taxa Analysis (TITAN) for identifying water quality thresholds in both field and synthetic data. Analysis of 14 years of macroinvertebrates data from the Mediterranean catchments of the Torrens and Onkaparinga Rivers, South-Australia, identified electrical conductivity (EC) and total phosphorus (TP) as the most important water quality variables affecting macroinvertebrates. Water quality thresholds for macroinvertebrates identified by both methods largely corresponded at low EC (GF: 400-900 µS cm-1 vs. TITAN: 407-951 µScm-1), total phosphorus (TP) (GF: 0.02-0.18 mg L-1 vs. TITAN: 0.02-0.04 mg L-1) and total nitrogen (TN) (GF: 0.2 mg L-1 vs. TITAN: 0.28-0.67 mg L-1) concentrations. However, multiple GF-derived thresholds, particularly at high stressor concentrations, were representative of low data distribution, and thus need to be considered with caution. In another case study of South Australian diatom data, there were marked differences in GF and TITAN identified thresholds for EC (GF: 5000 µScm-1 vs. TITAN 1004-2440 µS cm-1) and TP (GF: 250-500 µg L-1 vs. TITAN: 11-329 µg L-1). These differences were due to the fact that while TITAN parsed species responses into negative and positive taxa, GF overestimated thresholds by aggregating the response of taxa that increase and decrease along environmental gradients. Given these findings, we also evaluated the methods' performance using different distributions of synthetic data i.e. with both skewed and uniform distribution of samples and species responses. Both methods identified similar change-points in the case of a uniform environmental gradient, except when species optima were simulated at centre of the gradient. Here GF detected the change-points but TITAN failed to do so. GF also outperformed TITAN when four simulated species change-points were present. Thus, the distribution of species responses and optima and the evenness of the environment gradient can affect the models' performance. This study has shown that both methods are robust in identifying change in species response but threshold identification differs depending both on the analysis used and the nature of ecological data. We recommend the careful application of GF and TITAN, noting these differences in performance, will improve their application for water resource management.

Causal relationships of Raphidiopsis (formerly Cylindrospermopsis) dynamics with water temperature and N:P-ratios: A meta-analysis across lakes with different climates based on inferential modelling.

Raphidiopsis raciborskii is a tropical toxic cyanobacterium that is rapidly expanding to diverse lake habitats in different climate zones by sophisticated adaptation mechanisms. This meta-analysis investigated correlations of R. raciborskii with water temperature and N:P (nitrogen to phosphorus)-ratios across four lakes with different climates and trophic states by means of long-term time series and the hybrid evolutionary algorithm HEA. The results have shown that in the lakes with temperate and Mediterranean climate, R. raciborskii is strongly correlated with water temperature since germination and growth rely on rising water temperatures in spring. In contrast, there was a weaker correlation with water temperature in subtropical and tropical lakes where pelagic populations of R. raciborskii are overwintering, and are present all year round. However, the highest abundances of R. raciborskii coincided with highest water temperature for the Mediterranean, subtropical and tropical lakes, whilst in the temperate Langer See the highest abundances of R. raciborskii occurred at 24.1 °C, even though temperatures of up to 27 °C were recorded in 2013 and 2014. The correlation of R. raciborskii with N:P-ratios proved to be strongest for the meso- to eutrophic Lake Kinneret (r2 = 0.8) and lowest for the eutrophic Lake Paranoa (r2 = 0.16). However, the assumption has been confirmed that R. raciborskii is growing fastest when waters are N-limited regardless of trophic states. In terms of phenology, the temperate and Mediterranean lakes displayed "fastest growth" in spring and early summer. In contrast, the growing season in subtropical and tropical lakes lasted from spring to autumn most likely because of overwintering populations, and growing importance of direct and indirect biotic regulating factors such as competition, grazing, remineralisation of nutrients along warming climate. In order to carry out a meta-analysis of time series across four different lakes, HEA served as powerful tool resulting in inferential models with predictive capacity for population dynamics of R. raciborskii just driven by water temperature or N:P-ratios, whilst coefficients of determination r2 served as criteria for hypotheses testing.

Comparison of the alternative models SOURCE and SWAT for predicting catchment streamflow, sediment and nutrient loads under the effect of land use changes.

Quantifying the water quantity and quality variations resulting from human induced activities is important for policy makers in view of increasing water scarcity and water pollution. Simple models can be robust tools in estimating the runoff from catchments, but do they also sufficiently reflect complex physio-chemical processes required for spatially-explicit simulation of soil-water interactions, and the resulting pollutant responses in catchments? Do these models respond sensitive to the impacts of different land use change representations? These questions are considered by applying the semi-distributed process-based catchment models SWAT and SOURCE to the Sixth Creek catchment in South Australia. Both models used similar data whereas inputs for SOURCE were generated from land-use based Functional Units (FUs), while FUs for SWAT were based on land use, soil and slope combinations. After satisfying calibration of both models for the outlet station of the catchment, the simulated flow by SOURCE produced high goodness of fit metrics, while nutrient loads simulated by SWAT were more realistic. Both models benefitted from using locally available Potential Evapotranspiration data for calibrating the hydrology. Scenarios of intensified land uses by two models showed more credible results for sediment and nutrient loads with the static approach when simulating the linear rather than the non-linear land use changes. The study has shown that informing decisions on the hydrology at catchment scale is well suited to less-complex models, whereas decisions on impact of land use change on water quality in catchments are better suited by models with process descriptions for soil-water interactions.

Early warning of limit-exceeding concentrations of cyanobacteria and cyanotoxins in drinking water reservoirs by inferential modelling.

An early warning scheme is proposed that runs ensembles of inferential models for predicting the cyanobacterial population dynamics and cyanotoxin concentrations in drinking water reservoirs on a diel basis driven by in situ sonde water quality data. When the 10- to 30-day-ahead predicted concentrations of cyanobacteria cells or cyanotoxins exceed pre-defined limit values, an early warning automatically activates an action plan considering in-lake control, e.g. intermittent mixing and ad hoc water treatment in water works, respectively. Case studies of the sub-tropical Lake Wivenhoe (Australia) and the Mediterranean Vaal Reservoir (South Africa) demonstrate that ensembles of inferential models developed by the hybrid evolutionary algorithm HEA are capable of up to 30days forecasts of cyanobacteria and cyanotoxins using data collected in situ. The resulting models for Dolicospermum circinale displayed validity for up to 10days ahead, whilst concentrations of Cylindrospermopsis raciborskii and microcystins were successfully predicted up to 30days ahead. Implementing the proposed scheme for drinking water reservoirs enhances current water quality monitoring practices by solely utilising in situ monitoring data, in addition to cyanobacteria and cyanotoxin measurements. Access to routinely measured cyanotoxin data allows for development of models that predict explicitly cyanotoxin concentrations that avoid to inadvertently model and predict non-toxic cyanobacterial strains.

Modelling the impacts of altered management practices, land use and climate changes on the water quality of the Millbrook catchment-reservoir system in South Australia.

Sustainable management of drinking water reservoirs requires taking into account the potential effects of their catchments' development. This study is an attempt to estimate the daily patterns of nutrients transport in the catchment - reservoir systems through the application of the ensemble of complementary models SWAT-SALMO. SWAT quantifies flow, nitrate and phosphate loadings originating in catchments before entering downstream reservoirs meanwhile SALMO determines phosphate, nitrate, and chlorophyll-a concentrations within the reservoirs. The study applies to the semi-arid Millbrook catchment-reservoir system that supplies drinking water to north-eastern suburbs of Adelaide, South Australia. The catchment hosts viti- and horticultural land uses. The warm-monomictic, mesotrophic reservoir is artificially aerated in summer. After validating the simulation results for both Millbrook catchment and reservoir, a comprehensive scenario analysis has been conducted to reveal cascading effects of altered management practices, land uses and climate conditions on water quality in the reservoir. Results suggest that the effect on reservoir condition in summer would be severe, most likely resulting in chlorophyll-a concentrations of greater than 40 µg/l if the artificial destratification was not applied from early summer. A 50% curbing of water diversion from an external pipeline to the catchment will slightly limit chlorophyll-a concentrations by 1.22% as an effect of reduced inflow phosphate loads. The simulation of prospective land use scenarios converting 50% of present pasture in the Millbrook catchment into residential and orchards areas indicates an increase of summer chlorophyll-a concentrations by 9.5-107.9%, respectively in the reservoir. Global warming scenarios based on the high emission simulated by SWAT-SALMO did result in earlier growth of chlorophyll-a but overall the effects on water quality in the Millbrook reservoir was not significant. However scenarios combining global warming and land use changes resulted in significant eutrophication effects in the reservoir, especially in the unmanaged condition with stratification in summer. This study has demonstrated that complementary model ensembles like SWAT-SALMO allow to comprehend more realistically cascading effects of distinct catchment processes on internal reservoir's processes, and facilitate integrated management scenarios.

Future climate and land uses effects on flow and nutrient loads of a Mediterranean catchment in South Australia.

Mediterranean catchments experience already high seasonal variability alternating between dry and wet periods, and are more vulnerable to future climate and land use changes. Quantification of catchment response under future changes is particularly crucial for better water resources management. This study assessed the combined effects of future climate and land use changes on water yield, total nitrogen (TN) and total phosphorus (TP) loads of the Mediterranean Onkaparinga catchment in South Australia by means of the eco-hydrological model SWAT. Six different global climate models (GCMs) under two representative concentration pathways (RCPs) and a hypothetical land use change were used for future simulations. The climate models suggested a high degree of uncertainty, varying seasonally, in both flow and nutrient loads; however, a decreasing trend was observed. Average monthly TN and TP load decreased up to -55% and -56% respectively and were found to be dependent on flow magnitude. The annual and seasonal water yield and nutrient loads may only slightly be affected by envisaged land uses, but significantly altered by intermediate and high emission scenarios, predominantly during the spring season. The combined scenarios indicated the possibility of declining flow in future but nutrient enrichment in summer months, originating mainly from the land use scenario, that may elevate the risk of algal blooms in downstream drinking water reservoir. Hence, careful planning of future water resources in a Mediterranean catchment requires the assessment of combined effects of multiple climate models and land use scenarios on both water quantity and quality.

Inductive reasoning and forecasting of population dynamics of Cylindrospermopsis raciborskii in three sub-tropical reservoirs by evolutionary computation.

Seven-day-ahead forecasting models of Cylindrospermopsis raciborskii in three warm-monomictic and mesotrophic reservoirs in south-east Queensland have been developed by means of water quality data from 1999 to 2010 and the hybrid evolutionary algorithm HEA. Resulting models using all measured variables as inputs as well as models using electronically measurable variables only as inputs forecasted accurately timing of overgrowth of C. raciborskii and matched well high and low magnitudes of observed bloom events with 0.45≤r2>0.61 and 0.4≤r2>0.57, respectively. The models also revealed relationships and thresholds triggering bloom events that provide valuable information on synergism between water quality conditions and population dynamics of C. raciborskii. Best performing models based on using all measured variables as inputs indicated electrical conductivity (EC) within the range of 206-280mSm-1 as threshold above which fast growth and high abundances of C. raciborskii have been observed for the three lakes. Best models based on electronically measurable variables for the Lakes Wivenhoe and Somerset indicated a water temperature (WT) range of 25.5-32.7°C within which fast growth and high abundances of C. raciborskii can be expected. By contrast the model for Lake Samsonvale highlighted a turbidity (TURB) level of 4.8 NTU as indicator for mass developments of C. raciborskii. Experiments with online measured water quality data of the Lake Wivenhoe from 2007 to 2010 resulted in predictive models with 0.61≤r2>0.65 whereby again similar levels of EC and WT have been discovered as thresholds for outgrowth of C. raciborskii. The highest validity of r2=0.75 for an in situ data-based model has been achieved after considering time lags for EC by 7 days and dissolved oxygen by 1 day. These time lags have been discovered by a systematic screening of all possible combinations of time lags between 0 and 10 days for all electronically measurable variables. The so-developed model performs seven-day-ahead forecasts and is currently implemented and tested for early warning of C. raciborskii blooms in the Wivenhoe reservoir.

Elucidation and short-term forecasting of microcystin concentrations in Lake Suwa (Japan) by means of artificial neural networks and evolutionary algorithms.

Non-supervised artificial neural networks (ANN) and hybrid evolutionary algorithms (EA) were applied to analyse and model 12 years of limnological time-series data of the shallow hypertrophic Lake Suwa in Japan. The results have improved understanding of relationships between changing microcystin concentrations, Microcystis species abundances and annual rainfall intensity. The data analysis by non-supervised ANN revealed that total Microcystis abundance and extra-cellular microcystin concentrations in typical dry years are much higher than those in typical wet years. It also showed that high microcystin concentrations in dry years coincided with the dominance of the toxic Microcystis viridis whilst in typical wet years non-toxic Microcystis ichthyoblabe were dominant. Hybrid EA were used to discover rule sets to explain and forecast the occurrence of high microcystin concentrations in relation to water quality and climate conditions. The results facilitated early warning by 3-days-ahead forecasting of microcystin concentrations based on limnological and meteorological input data, achieving an r(2)=0.74 for testing.