Machine learning-based water quality prediction using octennial in-situ Daphnia magna biological early warning system data

Application of an improved LSTM model based on FECA and CEEMDAN VMD decomposition in water quality prediction

To address the limitations of existing water quality prediction models in handling non-stationary data and capturing multi-scale features, this study proposes a hybrid model integrating Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Variational Mode Decomposition (VMD), Long Short-Term Memory Network (LSTM), and Frequency-Enhanced Channel Attention (FECA). The model aims to improve prediction accuracy and robustness for complex water quality dynamics, which is critical for environmental protection and sustainable water resource management. First, CEEMDAN and Sample Entropy (SE) were used to decompose raw water quality data into interpretable components and filter noise. Then, a VMD-enhanced LSTM architecture embedded with FECA was developed to adaptively prioritize frequency-specific features, thereby improving the model's ability to handle nonlinear patterns. Results show that the model is successful in predicting all six water quality indicators: NH₃-N (ammonia nitrogen), DO (dissolved oxygen), pH, TN (total nitrogen), TP (total phosphorus), and CODMn (chemical oxygen demand, permanganate method). The model achieved Nash-Sutcliffe Efficiency (NSE) values ranging from 0.88 to 0.99. Using dissolved oxygen (DO) as an example, the model reduced the Mean Absolute Percentage Error (MAPE) by 0.12% and increased the coefficient of determination (R2) by 0.20% compared to baseline methods. This work provides a robust framework for real-time water quality monitoring and supports decision making in pollution control and ecosystem management.

Water quality parameters-based prediction of dissolved oxygen in estuaries using advanced explainable ensemble machine learning

The dissolved oxygen (DO) is crucial for the ecological health of estuaries and bays. However, human activities, land-sea interactions, and the unclear impact mechanisms of water quality parameters (WQPs) pose challenges to DO prediction. Water quality models and statistical methods were used to achieve predictions previously, with gaps of low accuracy and unclear impact mechanism between WQPs and DO. Here, we present an interpretable ensemble machine learning (EML) framework for DO prediction and reveal that the impact mechanism of WQPs on DO variation of six estuaries in China. The results show: 1) DO have significant short-term fluctuations and a decreasing trend in most rivers from November 2020 to December 2023. Bagging-boosting model (BBM) have best performance in most rivers, while stacking model (SM) achieves better prediction in Jilong River (R2 is 0.71 and RMSE is 0.55), due to its capability of utilizing more WQP information (lag features of Electrical conductivity (EC), permanganate concentration (CODMn), pH and total nitrogen (TN)) to predict DO under larger variation condition. 2) 1-3 days lag features of DO and water temperature (WT) play a crucial role in DO prediction and the 1-day lagged DO makes the largest contribution (mean absolute SHapley Additive explanation (SHAP) value higher than 0.9). The lag features of pH have a positive impact on DO in most rivers. 3) The factors that makes largest influence on DO in the same day differ across different rivers. The impact of 1-day lagged DO on model prediction has a threshold around 10 mg/L and the interaction between WT and DO on the same day shows spatial heterogeneity across different rivers. EC, pH, and TN have a positive impact on the DO of the same day, while WT, ammonia nitrogen (NH3-N) and total phosphorus (TP) have a negative impact. 4) Errors in model prediction stem from two aspects, one is insufficient driving force of features when they correctly guide predictions toward true value. Another is features with insignificant contribution pushing predictions in the opposite direction of true value. The proposed framework and the findings will allow more accurate understanding of the impact mechanism of WQPs on DO and provide important insights for hypoxia management in coastal rivers.

Spatial prediction of groundwater salinity in multiple aquifers of the Mekong Delta region using explainable machine learning models

Groundwater salinization is a prevalent issue in coastal regions, yet accurately predicting and understanding its causal factors remains challenging due to the complexity of the groundwater system. Therefore, this study predicted groundwater salinity in multi-layered aquifers spanning the entire Mekong Delta (MD) region using machine learning (ML) models based on an in situ dataset and using three indicators (Cl-, pH, and HCO3-). We applied nine different decision tree-based models and evaluated their prediction performances. The models were trained using 13 input variables: weather (2), hydrogeological conditions (4), water levels (3), groundwater usage (2), and relative distance from water sources (2). Subsequently, by employing model interpretation techniques, we quantified the significance of factors within the model prediction. Performance evaluations of the ML models demonstrated that the Extra Trees model exhibited superior performance and demonstrated generalization capabilities in predicting Cl- concentration, whereas the Bagging and Random Forest models outperformed the other models in predicting pH and HCO3- concentration. The coefficients of determination were determined to be 0.94, 0.67, and 0.78 for Cl-, pH, and HCO3-, respectively Additionally, the model interpretation effectively identified significant factors that depended on the target variables and aquifers. In particular, salinity indicators and aquifers that were strongly influenced by the artificial usage of groundwater were identified. Therefore, our research, which provides accurate spatial predictions and interpretations of groundwater salinity in the MD, has the potential to establish a foundation for formulating effective groundwater management policies to control groundwater salinization.

Distribution coefficient prediction using multimodal machine learning based on soil adsorption factors, XRF, and XRD spectrum data

The distribution coefficient (Kd) plays a crucial role in predicting the migration behavior of radionuclides in the soil environment. However, Kd depends on the complexities of geological and environmental factors, and existing models often do not reflect the unique soil properties. We propose a multimodal technique to predict Kd values for radionuclide adsorption in soils surrounding nuclear facilities in Republic of Korea. We integrated and trained three sub-networks reflecting different data domains: soil adsorption factors for physicochemical conditions, X-ray fluorescence (XRF) data, and X-ray diffraction (XRD) spectra for inherent soil properties. Our multimodal model achieved high performance, with a coefficient of determination (R2) of 0.84 and root mean squared error (RMSE) of 0.89 for natural log-transformed Kd. This is the first study to develop a multimodal model that simultaneously incorporates inherent soil properties and adsorption factors to predict Kd. We investigated influential peaks in XRD spectra and also revealed that pH and calcium oxide (CaO) were significant variables in soil adsorption factors and XRF data, respectively. These results promote the use of a multimodal model to predict Kd values by integrating data from different domains, providing a cost-effective and novel approach to elucidate the mechanisms of radionuclide adsorption in soil.

Using machine learning models to estimate Escherichia coli concentration in an irrigation pond from water quality and drone-based RGB imagery data

The rapid and efficient quantification of Escherichia coli concentrations is crucial for monitoring water quality. Remote sensing techniques and machine learning algorithms have been used to detect E. coli in water and estimate its concentrations. The application of these approaches, however, is challenged by limited sample availability and unbalanced water quality datasets. In this study, we estimated the E. coli concentration in an irrigation pond in Maryland, USA, during the summer season using demosaiced natural color (red, green, and blue: RGB) imagery in the visible and infrared spectral ranges, and a set of 14 water quality parameters. We did this by deploying four machine learning models - Random Forest (RF), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGB), and K-nearest Neighbor (KNN) - under three data utilization scenarios: water quality parameters only, combined water quality and small unmanned aircraft system (sUAS)-based RGB data, and RGB data only. To select the training and test datasets, we applied two data-splitting methods: ordinary and quantile data splitting. These methods provided a constant splitting ratio in each decile of the E. coli concentration distribution. Quantile data splitting resulted in better model performance metrics and smaller differences between the metrics for both the training and testing datasets. When trained with quantile data splitting after hyperparameter optimization, models RF, GBM, and XGB had R2 values above 0.847 for the training dataset and above 0.689 for the test dataset. The combination of water quality and RGB imagery data resulted in a higher R2 value (>0.896) for the test dataset. Shapley additive explanations (SHAP) of the relative importance of variables revealed that the visible blue spectrum intensity and water temperature were the most influential parameters in the RF model. Demosaiced RGB imagery served as a useful predictor of E. coli concentration in the studied irrigation pond.

A hybrid approach to improvement of watershed water quality modeling by coupling process-based and deep learning models

Watershed water quality modeling to predict changing water quality is an essential tool for devising effective management strategies within watersheds. Process-based models (PBMs) are typically used to simulate water quality modeling. In watershed modeling utilizing PBMs, it is crucial to effectively reflect the actual watershed conditions by appropriately setting the model parameters. However, parameter calibration and validation are time-consuming processes with inherent uncertainties. Addressing these challenges, this research aims to address various challenges encountered in the calibration and validation processes of PBMs. To achieve this, the development of a hybrid model, combining uncalibrated PBMs with data-driven models (DDMs) such as deep learning algorithms is proposed. This hybrid model is intended to enhance watershed modeling by integrating the strengths of both PBMs and DDMs. The hybrid model is constructed by coupling an uncalibrated Soil and Water Assessment Tool (SWAT) with a Long Short-Term Memory (LSTM). SWAT, a representative PBM, is constructed using geographical information and 5-year observed data from the Yeongsan River Watershed. The output variables of the uncalibrated SWAT, such as streamflow, suspended solids (SS), total nitrogen (TN), and total phosphorus (TP), as well as observed precipitation for the day and previous day, are used as training data for the deep learning model to predict the TP load. For the comparison, the conventional SWAT model is calibrated and validated to predict the TP load. The results revealed that TP load simulated by the hybrid model predicted the observed TP better than that predicted by the calibrated SWAT model. Also, the hybrid model reflects seasonal variations in the TP load, including peak events. Remarkably, when applied to other sub-basins without specific training, the hybrid model consistently outperformed the calibrated SWAT model. In conclusion, application of the SWAT-LSTM hybrid model could be a useful tool for decreasing uncertainties in model calibration and improving the overall predictive performance in watershed modeling. PRACTITIONER POINTS: We aimed to enhance process-based models for watershed water-quality modeling. The Soil and Water Assessment Tool-Long Short-Term Memory hybrid model's predicted and total phosphorus (TP) matched the observed TP. It exhibited superior predictive performance when applied to other sub-basins. The hybrid model will overcome the constraints of conventional modeling. It will also enable more effective and efficient modeling.

The drug loading capacity prediction and cytotoxicity analysis of metal-organic frameworks using stacking algorithms of machine learning

Metal-organic frameworks (MOFs) have shown excellent performance in the field of drug delivery. Despite the synthesis of a vast array of MOFs exceeding 100,000 varieties, certain formulations have exhibited suboptimal performance characteristics. Therefore, there is a pressing need to enhance their efficacy by identifying MOFs with superior drug loading capacities and minimal cytotoxicity, which can be achieved through machine learning (ML). In this study, a stacking regression model was developed to predict drug loading capacity and cytotoxicity of MOFs using datasets compiled from various literature sources. The model exhibited exceptional predictive capabilities, achieving R2 values of 0.907 for drug loading capacity and 0.856 for cytotoxicity. Furthermore, various model interpretation methods including partial dependence plots, individual conditional expectation, Shapley additive explanation, decision tree, random forest, CatBoost Regressor, and light gradient-boosting machine were employed for feature importance analysis. The results revealed that specific metal atoms such as Zn, Cr, Fe, Zr, and Cu significantly influenced the drug loading capacity and cytotoxicity of MOFs. Through model validation encompassing experimental validation and computational verification, the reliability of the model was thoroughly established. In general, it is a good practice to use ML methods for predicting drug loading capacity and cytotoxicity analysis of MOFs, guiding the development of future property prediction methods for MOFs.

Adsorption of uranyl ion on hexagonal boron nitride for remediation of real U-contaminated soil and its interpretation using random forest

Acid leaching has been widely applied to treat contaminated soil, however, it contains several inorganic pollutants. The decommissioning of nuclear power plants introduces radioactive and soluble U(VI), a substance posing chemical toxicity to humans. Our investigation sought to ascertain the efficacy of hexagonal boron nitride (h-BN), an highly efficient adsorbent, in treating U(VI) in wastewater. The adsorption equilibrium of U(VI) by h-BN reached saturation within a mere 2 h. The adsorption of U(VI) by h-BN appears to be facilitated through electrostatic attraction, as evidenced by the observed impact of pH variations, acidic agents (i.e., HCl or H2SO4), and the presence of background ions on the adsorption performance. A reusability test demonstrated the successful completion of five cycles of adsorption/desorption, relying on the surface characteristics of h-BN as influenced by solution pH. Based on the experimental variables of initial U(VI) concentration, exposure time, temperature, pH, and the presence of background ions/organic matter, a feature importance analysis using random forest (RF) was carried out to evaluate the correlation between performances and conditions. To the best of our knowledge, this study is the first attempt to conduct the adsorption of U(VI) generated from real contaminated soil by h-BN, followed by interpretation of the correlation between performance and conditions using RF. Lastly, a. plausible adsorption mechanism between U(VI) and h-BN was explained based on the experimental results, characterizations, and a. comparison with previous adsorption studies on the removal of heavy metals by h-BN.

The predictive model of hydrobiological diversity in the Asana-Tumilaca basin, Peru based on water physicochemical parameters and sediment metal content

The hydrobiological diversity in the basin depends on biotic and abiotic factors. A predictive model of hydrobiological diversity for periphyton and macrobenthos was developed through multiple linear regression analysis (MLRA) based on the physicochemical parameters of water (PPW) and metal content in sediments (MCS) from eight monitoring stations in the Asana-Tumilaca Basin during the dry and wet seasons. The electrical conductivity presented values between 47.9 and 3617 μS/cm, showing the highest value in the Capillune River due to the influence of geothermal waters. According to Piper's diagram, the water in the basin had a composition of calcium sulfate and calcium bicarbonate-sulfate. According to the Wilcox diagram, the water was found to be between good and very good quality, except for in the Capillune River. The Shannon-Wiener diversity indices (H') were 2.62 and 2.88 for periphyton, and 2.10 and 2.44 for macrobenthos, indicating moderate diversity; for the Pielou's evenness index (J'), they were 0.68 and 0.70 for periphyton, and 0.68 and 0.59 for macrobenthos, indicating similar equity, in the dry and wet seasons, respectively, for both indices. In the model there were three cases, where the first two cases only worked with PPW or MCS, and case 3 worked with PPW and MCS. For case 3, the predicted values for H' and J' of periphyton and macrobenthos concerning those observed presented correlation coefficients of 0.7437 and 0.6523 for periphyton and 0.9321 and 0.8570 for macrobenthos, respectively, which were better than those of cases 1 and 2. In addition, principal component analysis revealed that the As, Pb, and Zn contents in the sediments negatively influenced the diversity, uniformity, and richness of the macrobenthos. In contrast, Cu and Cr had positive impacts because of the adaptation processes.