Performance evaluation of deep learning based stream nitrate concentration prediction model to fill stream nitrate data gaps at low-frequency nitrate monitoring basins

Accurate and frequent nitrate estimates can provide valuable information on the nitrate transport dynamics. The study aimed to develop a data-driven modeling framework to estimate daily nitrate concentrations at low-frequency nitrate monitoring sites using the daily nitrate concentration and stream discharge information of a neighboring high-frequency nitrate monitoring site. A Long Short-Term Memory (LSTM) based deep learning (DL) modeling framework was developed to predict daily nitrate concentrations. The DL modeling framework performance was compared with two well-established statistical models, including LOADEST and WRTDS-Kalman, in three selected basins in Iowa, USA: Des Moines, Iowa, and Cedar River. The developed DL model performed well with NSE >0.70 and KGE >0.70 for 67% and 79% nitrate monitoring sites, respectively. DL and WRTDS-Kalman models performed better than the LOADEST in nitrate concentration and load estimation for all low-frequency sites. The average NSE performance of the DL model in daily nitrate estimation is 20% higher than that of the WRTDS-Kalman model at 18 out of 24 sites (75%). The WRTDS-Kalman model showed unrealistic fluctuations in the estimated daily nitrate time series when the model received limited observed nitrate data (less than 50) for simulation. The DL model indicated superior performance in winter months' nitrate prediction (60% of cases) compared to WRTDS-Kalman models (33% of cases). The DL model also better represented the exceedance days from the USEPA maximum contamination level (MCL). Both the DL and WRTDS-Kalman models demonstrated similar performance in annual stream nitrate load estimation, and estimated values are close to actual nitrate loads.

Investigating the applicability and assumptions of the regression relationship between flow discharge and nitrogen concentrations for load estimation

The regression relationship between water discharge rates and nutrient concentrations can provide a quick and straightforward way to estimate nutrient loads. However, recent studies indicated that the relationship might produce large biases in load estimates and, therefore, may not be applicable in certain types of cases. The goal of this study is to explore the theoretical reasons behind the selective applicability of the regression relationship between flow rates and nitrate + nitrite concentrations. For this study, we examined daily flow and nitrate + nitrite concentration observations made at the outlets of 22 watersheds monitored by the Heidelberg Tributary Loading Program (HTLP). The statistical relationship between the flow rates and concentrations was explored using regression equations offered by the LOAD ESTimator (LOADEST). Results demonstrated that the use of the regression equations provided nitrate + nitrite load estimates at acceptable accuracy levels (N S E ≥ 0.35 and | P B I A S | ≤ 30.0 %) in 14 watersheds (64 % of 22 study watersheds). The regression relationships provided highly biased results at eight watersheds (36 %), implying their limited applicability. The heteroscedasticity of the residuals led to the high bias and resulting inaccurate regression, which was commonly found in watersheds where low flow had high nitrate + nitrite concentration variations. Conversely, the regression relationships provided acceptable accuracy for watersheds that had a relatively constant variance of the nitrate + nitrite concentrations. The results indicate that the homoscedasticity of residuals is the key assumption to be satisfied to estimate nitrate + nitrite loads from a statistical regression between flow discharge and nitrate + nitrite concentrations. The transport capacity (capacity-limited) concept implicitly assumed in the regression relationship between flow discharge and nitrate + nitrite concentrations is not always applicable, especially to agricultural areas in which nitrate + nitrite loads are highly variable depending on management practices (supply-limited). The findings suggest that the regression relationship should be carefully applied to areas in which intensive agricultural activities, including crop management and conservation practices, are implemented. Thus, the transport capacity concept is reasonably regarded to contribute to the homoscedasticity of residuals.

Uncertainty quantification in reconstruction of sparse water quality time series: Implications for watershed health and risk-based TMDL assessment

Despite the plethora of methods available for uncertainty quantification, their use has been limited in the practice of water quality (WQ) modeling. In this paper, a decision support tool (DST) that yields a continuous time series of WQ loads from sparse data using streamflows as predictor variables is presented. The DST estimates uncertainty by analyzing residual errors using a relevance vector machine. To highlight the importance of uncertainty quantification, two applications enabled within the DST are discussed. The DST computes (i) probability distributions of four measures of WQ risk analysis- reliability, resilience, vulnerability, and watershed health- as opposed to single deterministic values and (ii) concentration/load reduction required in a WQ constituent to meet total maximum daily load (TMDL) targets along with the associated risk of failure. Accounting for uncertainty reveals that a deterministic analysis may mislead about the WQ risk and the level of compliance attained with established TMDLs.

Assessment of the impact of climate change and mining activities on streamflow and selected metal's loading in the Chindwin River, Myanmar

The rapid expansion in mining activities is deteriorating the water quality in the Chindwin River of Myanmar. In addition, climate change may also aggravate this situation in future. Therefore, the aim of this study was to establish a connection between hydrology, mining area, heavy metal loading, and climate change in the Chindwin River. The hydrology of the upper Chindwin basin was modelled using SHETRAN hydrological model. Geochemical model PHREEQC was utilised to conduct speciation and saturation indexes modelling along the river in order to quantify the precipitated minerals along the river. Thereafter a regression relationship along with LOADEST model was used to quantify the heavy metal loads. Future climate was projected using four RCM's namely ACCESS1-CSIRO-CCAM, CCSM4-CSIRO-CCAM, CNRM-CM5-CSIRO-CCAM and MPI-ESM-LR-CSIRO-CCAM. Future discharges at water quality monitoring stations were simulated using the averaged ensembles. Finally, the heavy metal loading under future climate scenarios were analysed. Results indicate that climate change is likely to reduce future discharges by 3.4%-36.5% in all stations except in the Mokekalae station which shows 1.3%-9.4% increase in the near future discharges. Also, the projected metal loading under future climate conditions shows a decreasing pattern which is similar to the projected discharge pattern. In both baseline and future climate conditions, the area between stations Naung Po Aung and Uru downstream show the highest load effluent for both arsenic and mercury while the area between stations Uru downstream and Mokekalae show the highest load of iron effluent. Although future heavy metal loadings are expected to decrease, mining activities should be carefully monitored, since they discharge a large amount of toxic heavy metal loadings into the Chindwin River which is also expected to suffer a decrease streamflow in future.

Rating curve based assessment of seasonal variability of sulfate in streamflow

Though constituent concentrations and loads in rivers exhibit apparent seasonal fluctuations, they are characterized by event-driven nature of the fluctuations in respond to natural processes and seasonal anthropogenic activities. This study aimed at establishing relationship between streamflow and sulfate load in Gin River, the major water source in southern Sri Lanka and assessing seasonal sulfate levels in the streamflow following monsoon pattern and cultivation seasons. Rating curve, a load-streamflow regression model, was developed using adjusted maximum likelihood estimation. Following the assumptions of model fit, the regression model showed low correlation among explanatory variables and good empirical agreement with the measured data exhibiting its applicability to deduce sulfate loads from streamflow data, during non-sampling periods. Sulfate loads, highly dependent on streamflow, peaked annually in April-June (south-west monsoon contributing to Yala cultivation season) and October-December (north-east monsoon contributing to Maha cultivation season), following the bimodal monsoon pattern in the catchment. Median sulfate load exhibited fourfold increase from the lowest value 8,888 kg/day in August (non-cultivation season) to the highest value 38,185 kg/day in November (Maha cultivation season), despite the twofold increase of median streamflow between the two months. Flow-weighted sulfate concentrations showed varying flow dependence attributed to the seasonality. At low streamflows (above 70th percentile), sulfate concentration and streamflow were inversely related and at high streamflows (below 30th percentile), and sulfate concentration and streamflow were directly related. Elevated sulfate concentrations attributed to less soluble sulfate irons were clearly evident during the two cultivation seasons which coincided with the monsoon periods.

Calculating salt loads to Great Salt Lake and the associated uncertainties for water year 2013; updating a 48 year old standard

Effective management of surface waters requires a robust understanding of spatiotemporal constituent loadings from upstream sources and the uncertainty associated with these estimates. We compared the total dissolved solids loading into the Great Salt Lake (GSL) for water year 2013 with estimates of previously sampled periods in the early 1960s. We also provide updated results on GSL loading, quantitatively bounded by sampling uncertainties, which are useful for current and future management efforts. Our statistical loading results were more accurate than those from simple regression models. Our results indicate that TDS loading to the GSL in water year 2013 was 14.6 million metric tons with uncertainty ranging from 2.8 to 46.3 million metric tons, which varies greatly from previous regression estimates for water year 1964 of 2.7 million metric tons. Results also indicate that locations with increased sampling frequency are correlated with decreasing confidence intervals. Because time is incorporated into the LOADEST models, discrepancies are largely expected to be a function of temporally lagged salt storage delivery to the GSL associated with terrestrial and in-stream processes. By incorporating temporally variable estimates and statistically derived uncertainty of these estimates, we have provided quantifiable variability in the annual estimates of dissolved solids loading into the GSL. Further, our results support the need for increased monitoring of dissolved solids loading into saline lakes like the GSL by demonstrating the uncertainty associated with different levels of sampling frequency.

Spatial and temporal trends in estimates of nutrient and suspended sediment loads in the Ishikari River, Japan, 1985 to 2010

Nutrients and suspended sediment in surface water play important roles in aquatic ecosystems and contribute strongly to water quality with implication for drinking water resources, human and environmental health. Estimating loads of nutrients (nitrogen and phosphorus) and suspended sediment (SS) is complicated because of infrequent monitoring data, retransformation bias, data censoring, and non-normality. To obtain reliable unbiased estimates, the Maintenance of Variance-Extension type 3 (MOVE. 3) and the regression model Load Estimator (LOADEST) were applied to develop regression equations and to estimate total nitrogen (TN), total phosphorus (TP) and SS loads at five sites on the Ishikari River, Japan, from 1985 to 2010. Coefficients of determination (R(2)) for the best-fit regression models for loads of TN, TP, and SS for the five sites ranged from 71.86% to 90.94%, suggesting the model for all three constituents successfully simulated the variability in constituent loads at all studied sites. Estimated monthly average loads at Yishikarikakou-bashi were larger than at the other sites, with TN, TP, and SS loads ranging from 8.52×10(3) to 2.00×10(5) kg/day (Apr. 1999), 3.96×10(2) to 5.23×10(4) kg/ day (Apr. 1999), and 9.21×10(4) to 9.25×10(7) kg/day (Sep. 2001), respectively. Because of variation in river discharge, the estimated seasonal loads fluctuated widely over the period 1985 to 2010, with the greatest loads occurring in spring and the smallest loads occurring in winter. Estimated loads of TN, TP, and especially SS showed decreasing trends during the study period. Accurate load estimation is a necessary goal of water quality monitoring efforts and the methods described here provide essential information for effectively managing water resources.

Estimating Escherichia coli loads in streams based on various physical, chemical, and biological factors

Microbes have been identified as a major contaminant of water resources. Escherichia coli (E. coli) is a commonly used indicator organism. It is well recognized that the fate of E. coli in surface water systems is governed by multiple physical, chemical, and biological factors. The aim of this work is to provide insight into the physical, chemical, and biological factors along with their interactions that are critical in the estimation of E. coli loads in surface streams. There are various models to predict E. coli loads in streams, but they tend to be system or site specific or overly complex without enhancing our understanding of these factors. Hence, based on available data, a Bayesian Neural Network (BNN) is presented for estimating E. coli loads based on physical, chemical, and biological factors in streams. The BNN has the dual advantage of overcoming the absence of quality data (with regards to consistency in data) and determination of mechanistic model parameters by employing a probabilistic framework. This study evaluates whether the BNN model can be an effective alternative tool to mechanistic models for E. coli loads estimation in streams. For this purpose, a comparison with a traditional model (LOADEST, USGS) is conducted. The models are compared for estimated E. coli loads based on available water quality data in Plum Creek, Texas. All the model efficiency measures suggest that overall E. coli loads estimations by the BNN model are better than the E. coli loads estimations by the LOADEST model on all the three occasions (three-fold cross validation). Thirteen factors were used for estimating E. coli loads with the exhaustive feature selection technique, which indicated that six of thirteen factors are important for estimating E. coli loads. Physical factors included temperature and dissolved oxygen; chemical factors include phosphate and ammonia; biological factors include suspended solids and chlorophyll. The results highlight that the LOADEST model estimates E. coli loads better in the smaller ranges, whereas the BNN model estimates E. coli loads better in the higher ranges. Hence, the BNN model can be used to design targeted monitoring programs and implement regulatory standards through TMDL programs.