Machine learning-derived dose-response relationships considering interactions in mixtures: Applications to the oxidative potential of particulate matter.

Conventional environmental health research is primarily focused on isolated chemical exposures, neglecting the complex interactions between multiple pollutants that may synergistically or antagonistically influence toxicity, thereby posing unexpected health risks. In this study, we address this knowledge gap by introducing an explainable machine learning (ML) approach with Feature Localized Intercept Transformed-Shapley Additive Explanations (FLIT-SHAP) designed to extract the dose-response relationships of specific pollutants in mixtures. In contrast to traditional SHAP, FLIT-SHAP can localize the global intercept to elucidate mixture effects, which is crucial for understanding the oxidative potential (OP) of ambient particulate matter (PM). Assessing multi-pollutant OP using FLIT-SHAP revealed both synergistic (55-63 %) and antagonistic (25-42 %) effects in laboratory-controlled OP data, but an antagonistic (33-66 %; lower OP) effect in ambient PM. Notably, the FLIT-SHAP approach demonstrated higher prediction accuracy (R2 = 0.99) compared to the additive model (R2 = 0.89) when evaluated against real-world PM samples. Quinones, such as phenanthrenequinone, play a more significant role in PM2.5 than previously recognized. Through this study, we highlighted the potential of FLIT-SHAP to enhance toxicity predictions and aid decision-making in the field of environmental health.

Spatiotemporal estimation of groundwater and surface water conditions by integrating deep learning and physics-based watershed models.

The impacts of climate change on hydrology underscore the urgency of understanding watershed hydrological patterns for sustainable water resource management. The conventional physics-based fully distributed hydrological models are limited due to computational demands, particularly in the case of large-scale watersheds. Deep learning (DL) offers a promising solution for handling large datasets and extracting intricate data relationships. Here, we propose a DL modeling framework, incorporating convolutional neural networks (CNNs) to efficiently replicate physics-based model outputs at high spatial resolution. The goal was to estimate groundwater head and surface water depth in the Sabgyo Stream Watershed, South Korea. The model datasets consisted of input variables, including elevation, land cover, soil type, evapotranspiration, rainfall, and initial hydrological conditions. The initial conditions and target data were obtained from the fully distributed hydrological model HydroGeoSphere (HGS), whereas the other inputs were actual measurements in the field. By optimizing the training sample size, input design, CNN structure, and hyperparameters, we found that CNNs with residual architectures (ResNets) yielded superior performance. The optimal DL model reduces computation time by 45 times compared to the HGS model for monthly hydrological estimations over five years (RMSE 2.35 and 0.29 m for groundwater and surface water, respectively). In addition, our DL framework explored the predictive capabilities of hydrological responses to future climate scenarios. Although the proposed model is cost-effective for hydrological simulations, further enhancements are needed to improve the accuracy of long-term predictions. Ultimately, the proposed DL framework has the potential to facilitate decision-making, particularly in large-scale and complex watersheds.

Deep learning-based efficient drone-borne sensing of cyanobacterial blooms using a clique-based feature extraction approach.

Recent advances in remote sensing techniques provide a new horizon for monitoring the spatiotemporal variations of harmful algal blooms (HABs) using hyperspectral data in inland water. In this study, a hierarchical concatenated variational autoencoder (HCVAE) is proposed as an efficient and accurate deep learning (DL) based bio-optical model. To demonstrate its usefulness in retrieving algal pigments, the HCVAE is applied to bloom-prone regions in Daecheong Lake, South Korea. By abstracting the similarity between highly related features using layer-wise clique-based latent-feature extraction, HCVAE reduces the computational loads in deriving outputs while preventing performance degradation. Graph-based clique-detection uses information theory-based criteria to group the related reflectance spectra. Consequently, six latent features were extracted from 79 spectral bands to consist of a multilevel hierarchy of HCVAE that can simultaneously estimate concentrations of chlorophyll-a (Chl-a) and phycocyanin (PC). Despite the parsimonious model architecture, the Chl-a and PC concentrations estimated by HCVAE closely agree with the measured concentrations, with test R2 values of 0.76 and 0.82, respectively. In addition, spatial distribution maps of algal pigments obtained from HCVAE using drone-borne reflectance successfully capture the blooming spots. Based on its multilevel hierarchical architecture, HCVAE can provide the importance of latent features along with their individual wavelengths using Shapley additive explanations. The most important latent features covered the spectral regions associated with both Chl-a and PC. The lightweight neural network DNNsel, which uses only the spectral bands of highest importance in latent-feature extraction, performed comparably to HCVAE. The study results demonstrate the utility of the multilevel hierarchical architecture as a comprehensive assessment model for near-real-time drone-borne sensing of HABs. Moreover, HCVAE is applicable to a wide range of environmental big data, as it can handle numerous sets of features.

Long short-term memory models of water quality in inland water environments.

Water quality is substantially influenced by a multitude of dynamic and interrelated variables, including climate conditions, landuse and seasonal changes. Deep learning models have demonstrated predictive power of water quality due to the superior ability to automatically learn complex patterns and relationships from variables. Long short-term memory (LSTM), one of deep learning models for water quality prediction, is a type of recurrent neural network that can account for longer-term traits of time-dependent data. It is the most widely applied network used to predict the time series of water quality variables. First, we reviewed applications of a standalone LSTM and discussed its calculation time, prediction accuracy, and good robustness with process-driven numerical models and the other machine learning. This review was expanded into the LSTM model with data pre-processing techniques, including the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise method and Synchrosqueezed Wavelet Transform. The review then focused on the coupling of LSTM with a convolutional neural network, attention network, and transfer learning. The coupled networks demonstrated their performance over the standalone LSTM model. We also emphasized the influence of the static variables in the model and used the transformation method on the dataset. Outlook and further challenges were addressed. The outlook for research and application of LSTM in hydrology concludes the review.

Hierarchical deep learning model to simulate phytoplankton at phylum/class and genus levels and zooplankton at the genus level.

Harmful algal blooms (HABs) have become a global issue, affecting public health and water industries in numerous countries. Because funds for monitoring HABs are limited, model development may be an alternative approach for understanding and managing HABs. Continuous monitoring based on grab sampling is time-consuming, costly, and labor-intensive. However, improving simulation performance remains a major challenge in modeling, and current methods are limited to simulating phytoplankton (e.g., Microcystis sp., Anabaena sp., Aulacoseira sp., Cyclotella sp., Pediastrum sp., and Eudorina sp.) and zooplankton (e.g., Cyclotella sp., Pediastrum sp., and Eudorina sp.) at the genus level. The traditional modeling approach is limited for evaluating the interactions between phytoplankton and zooplankton. Recently, deep learning (DL) models have been proposed for solving modeling problems because of their large data handling capabilities and model structure flexibilities. In this study, we evaluated the applicability of DL for simulating phytoplankton at the phylum/class and genus levels and zooplankton at the genus level. Our work was an explicit representation of the taxonomic and ecological hierarchy of the DL model structure. The prerequisite for this model design was the data collection at two taxonomic and hierarchical levels. Our model consisted of hierarchical DL with classification transformer (TF) and regression TF models. These DL models were hierarchically connected; the output of the phylum/class level model was transferred to the genus level simulation model, and the output of the genus level model was fed into the zooplankton simulation model. The classification TF model determined the phytoplankton occurrence initiation date, whereas the regression TF model quantified the cell concentration of plankton. The hierarchical DL showed potential to simulate phytoplankton at the phylum/class and genus levels by producing average R2, and root mean standard error values of 0.42 and 0.83 [log(cells mL-1)], respectively. All simulated plankton results closely matched the measured concentrations. Particularly, the simulated cyanobacteria showed good agreement with the measured cell concentration, with an R2 value of 0.72. In addition, our simulated result showed good agreement in peak concentration compared to observations. However, a limitation remained in following the temporal variation of Tintinnopsis sp. and Bosmia sp. Using an importance map from the TF model, water temperature, total phosphorus, and total nitrogen were identified as significant variables influencing phytoplankton and zooplankton blooms. Overall, our study demonstrated that DL can be used for modeling HABs at the phylum/class and genus levels.

A novel method for micropollutant quantification using deep learning and multi-objective optimization.

Micropollutants (MPs) released into aquatic ecosystems have adverse effects on public health. Hence, monitoring and managing MPs in aquatic systems are imperative. MPs can be quantified by high-resolution mass spectrometry (HRMS) with stable isotope-labeled (SIL) standards. However, high cost of SIL solutions is a significant issue. This study aims to develop a rapid and cost-effective analytical approach to estimate MP concentrations in aquatic systems based on deep learning (DL) and multi-objective optimization. We hypothesized that internal standards could quantify the MP concentrations other than the target substance. Our approach considered the precision of intra-/inter-day repeatability and natural organic matter information to reduce instrumental error and matrix effect. We selected standard solutions to estimate the concentrations of 18 MPs. Among the optimal DL models, DarkNet-53 using nine standard solutions yielded the highest performance, while ResNet-50 yielded the lowest. Overall, this study demonstrated the capability of DL models for estimating MP concentrations.

Analysis of micropollutants in a marine outfall using network analysis and decision tree.

The presence of micropollutants (MPs), including pharmaceutical, industrial, and pesticidal compounds, threatens both human health and the aquatic ecosystem. The development and extensive use of new chemicals have also inevitably led to the accumulation of MPs in aquatic environments. Recreational beaches are especially vulnerable to contamination, affecting humans and aquatic animals via the absorption of MPs in water during marine activities (e.g., swimming, sailing, and windsurfing). Additionally, marine outfalls in an urbanized coastal city can cause serious chemical and microbial pollution on recreational beaches, leading to an increase in adverse effects on public health and the ecological system. Therefore, the aim of this study was to, with the use of network and decision tree analyses, identify the features and factors that influence the change in MP concentrations in a marine outfall. These analyses were conducted to inspect the relationship between each MP and its hierarchical structure as well as hydrometeorological variables. Additionally, a risk analysis was conducted in this study in which the MPs were prioritized based on their optimized risk quotient values. During our monitoring of MP concentrations over time at the marine outfall, high concentrations of pharmaceutical and industrial compounds were detected when the tide level was low after rainfall. Furthermore, results of the risk analysis and the prioritization revealed that a total of 18 substances identified in our study posed a risk to the ecosystem; these include major ecotoxicologically hazardous substances such as telmisartan, mevinphos, and methiocarb. Results of the network analysis demonstrated distinct trends for pharmaceutical and industrial substances, whilst those for pesticide compounds were irregular. Additionally, the hierarchical structures for most MPs consisted of rainfall, tide level, and antecedent dry hours; this implies that these factors influence MP dynamics. These findings will be helpful for establishing chemical contamination management plans for recreational beaches in the future.

Cyanobacteria cell prediction using interpretable deep learning model with observed, numerical, and sensing data assemblage.

Massive cyanobacterial blooms in river water causes adverse impacts on aquatic ecosystems and water quality. Complex and diverse data sources are available to investigate the cyanobacteria phenomena, including in situ data, synthetic measurements, and remote sensing images. Deep learning attention models can process these intricate sources to forecast cyanobacteria by identifying important variables in the data sources. However, deep learning attention models for predicting cyanobacteria have rarely been studied using an assemblage of various datasets. Thus, in this study, a convolutional neural network (CNN) model with a convolutional block attention module (CNNan) was developed to predict cyanobacterial cell concentrations by using the observed cell data from field monitoring, chlorophyll-a distribution map from hyperspectral image sensing, and simulated water quality outputs from a hydrodynamic model. Then, the prediction performance of the CNNan model was compared to an environmental fluid dynamics code (EFDC) simulation and a CNN model without an attention network. The seasonal variations of the predicted cyanobacteria that was obtained from CNNan showed the best agreement with the observed variations with Nash-Sutcliffe efficiency values higher than 0.76 when compared to the EFDC and CNN predictions. The daily hydrodynamic outputs allowed the prediction of cyanobacteria cells, while the rich information of the chlorophyll-a map contributed to the improvement of the prediction performance at certain periods. Moreover, the attention network visualized the importance of the additional chlorophyll-a map and improved the CNNan model prediction performance by refining the input features. Therefore, this study demonstrated that a deep learning model with data assemblage is practically feasible for predicting the presence of harmful algae in inland water.

Monitoring the vertical distribution of HABs using hyperspectral imagery and deep learning models.

Remote sensing techniques have been applied to monitor the spatiotemporal variation of harmful algal blooms (HABs) in many inland waters. However, these studies have been limited to monitor the vertical distribution of HABs due to the optical complexity of inland water. Therefore, this study applied a deep neural network model to monitor the vertical distribution of Chlorophyll-a (Chl-a), phycocyanin (PC), and turbidity (Turb) using drone-borne hyperspectral imagery, in-situ measurement, and meteoroidal data. The pigment concentrations were measured between depths of 0 m and 5.0 m with 0.05 m intervals. Here, four state-of-the-art data-driven model structures (ResNet-18, ResNet-101, GoogLeNet, and Inception v3) were adopted for estimating the vertical distributions of the harmful algal pigments. Among the four models, the ResNet-18 model showed the best performance, with an R2 value of 0.70. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) substantially provided informative reflectance band ranges near 490 nm and 620 nm in the hyperspectral image for the vertical estimation of pigments. Therefore, this study demonstrated that the explainable deep learning model with drone-borne hyperspectral images has the potential to estimate Chl-a, PC, and Turb vertical distributions and to show influential features that contribute to describing the vertical profile phenomena.

Application of Machine Learning for eutrophication analysis and algal bloom prediction in an urban river: A 10-year study of the Han River, South Korea.

The increasing release of nutrients to aquatic environments has led to great concern regarding eutrophication and the risk of unwanted algal blooms. Based on observational data of 20 water quality parameters measured on a monthly basis at 40 stations from 2011 to 2020, this study applied different Machine Learning (ML) algorithms to suggest the best option for algal bloom prediction in the Han River, a large river in South Korea. Eight different ML algorithms were categorized into several groups of statistical learning, regression family, and deep learning, and were then compared for their suitability to predict the chlorophyll-derived trophic index (TSI-Chla). ML algorithms helped identify the most important water quality parameters contributing to algal bloom prediction. The ML results confirmed that eutrophication and algal proliferation were governed by the complex interplay between nutrients (nitrogen and phosphorus), organic contaminants, and environmental factors. Of the models tested, the adaptive neuro-fuzzy inference system (ANFIS) exhibited the best performance owing to its consistent and outperforming prediction both quantitatively (i.e., via regression) and qualitatively (i.e., via classification), which was evidenced by the lowest value of mean absolute error (MAE) of 0.09, and the highest F1-score, Recall and Precision of 0.97, 0.98 and 0.96, respectively. In a further step, a representative web application was constructed to assist common users to predict the trophic status of the Han River. This study demonstrated that ML techniques are not only promising for highly accurate water quality modeling of urban rivers, but also reduce time and labor intensity for experiments, which decreases the number of monitored water quality parameters, providing further insights into the driving factors of water quality deterioration. They ultimately help devise proactive strategies for sustainable water management.

Effect of hyperspectral image-based initial conditions on improving short-term algal simulation of hydrodynamic and water quality models.

Hydrodynamic and water quality modeling have provided valuable simulation results that have enhanced the understanding of the spatial and temporal distribution of algal blooms. Typical model simulations are performed with point-based observational data that are used to configure initial and boundary conditions, and for parameter calibration. However, the application of such conventional modeling approaches is limited due to cost, labor, and time constraints that preclude the retrieval of high-resolution spatial data. Thus, the present study applied fine-resolution algal data to configure the initial conditions of a hydrodynamic and water quality model and compared the accuracy of short-term algal simulations with the results simulated using conventional point-based initial conditions. The environmental fluid dynamics code (EFDC) model was calibrated to simulate Chlorophyll-a (Chl-a) concentrations. Hyperspectral images were used to generate Chl-a maps based on a two-band ratio algorithm for configuring the initial condition of the EFDC model. The model simulation with hyperspectral-based initial conditions returned relatively accurate results for Chl-a, compared to the simulation based on point-based initial conditions. The simulations exhibited percent bias values of 9.93 and 14.23, respectively. Therefore, the results of this study demonstrate how hyperspectral-based initial conditions could improve the reliability of short-term algal bloom simulations in a hydrodynamic model.

Identification of influencing factors of A. catenella bloom using machine learning and numerical simulation.

Alexandrium catenella (A. catenella) is a notorious algal species known to cause paralytic shellfish poisoning (PSP) in Korean coastal waters. There have been numerous studies on its temporal and spatial blooms in Korea. However, its bloom dynamics have not been fully understood because of the complexity in physical, chemical, and biological environments. This study aims to identify the factors that influence A. catenella blooms by applying a numerical model and machine learning. Intensive monitoring of A. catenella was conducted to investigate temporal variations in its population and its spatial distribution in the area with frequent occurrences of PSP bloom initiation. Moreover, a numerical model was built to analyze the ocean physical factors related to the bloom of A. catenella. Based on the information obtained from the monitored and simulated results, the decision tree (DT) method was applied to identify factors that caused the bloom. The outbreak of A. catenella was observed in the eastern coastal water of Geoje Island in 2017, recording a peak density of 4 × 104 (cell L-1). Retention time and particle scattering demonstrated that the physical force in 2017 was weaker than that in 2018, as shown by the smaller effects of advection and dispersion in 2017. The decision tree model showed that (1) water temperature below 17.21 °C was ideal for the growth of A. catenella, (2) phosphate influenced the growth of the species, and (3) cell density was accelerated with increasing retention time. The results from DT can contribute to the prediction of A. catenella blooms by determining the conditions that cause bloom initiation. Further, they can be used as a practical approach for mitigating HABs. Thus, machine learning and numerical simulation in this study can be a potential approach for effectively managing the bloom of A. catenella.

Designing a marine outfall to reduce microbial risk on a recreational beach: Field experiment and modeling.

A marine outfall can be a wastewater management system that discharges sewage and stormwater into the sea; hence, it is a source of microbial pollution on recreational beaches, including antibiotic resistant genes (ARGs), which lead to an increase in untreatable diseases. In this regard, a marine outfall must be efficiently located to mitigate these risks. This study aimed to 1) investigate the spatiotemporal variability of Escherichia coli (E. coli) and ARGs on a recreational beach and 2) design marine outfalls to reduce microbial risks. For this purpose, E. coli and ARGs with influential environmental variables were intensively monitored on Gwangalli beach, South Korea in this study. Environmental fluid dynamic code (EFDC) was used and calibrated using the monitoring data, and 12 outfall extension scenarios were explored (6 locations at 2 depths). The results revealed that repositioning the marine outfall can significantly reduce the concentrations of E. coli and ARGs on the beach by 46-99%. Offshore extended outfalls at the bottom of the sea reduced concentrations of E. coli and ARGs on the beach more effectively than onshore outfalls at the sea surface. These findings could be helpful in establishing microbial pollution management plans at recreational beaches in the future.

Replacing the internal standard to estimate micropollutants using deep and machine learning.

Similar to the worldwide proliferation of urbanization, micropollutants have been involved in aquatic and ecological environmental systems. These pollutants have the propensity to wreak havoc on human health and the ecological system; hence, it is important to persistently monitor micropollutants in the environment. Micropollutants are commonly quantified via target analysis using high resolution mass spectrometry and the stable isotope labeled (SIL) standard. However, the cost-intensiveness of this standard presents a major obstacle in measuring micropollutants. This study resolved this problem by developing data-driven models, including deep learning (DL) and machine learning (ML), to estimate the concentration of micropollutants without resorting to the SIL standard. Our study hypothesized that natural organic matter (NOM) could replace internal standards if there was a specific mass spectrum (MS) subset, including NOM information, which correlated with an SIL standard peak. Therefore, we analyzed the MS to find the specific MS subsets for replacing the SIL standard peak. Thirty-five alternative MS subsets were determined for applying DL and ML as input data. Thereafter, we trained four different DL models, namely, ResNet101, GoogLeNet, VGG16, and Inception v3, as well as three different ML models, i.e., random forest (RF), support vector machine (SVM), and artificial neural network (ANN). A total of 680 MS data were used for the model training to estimate five different micropollutants, namely Sulpiride, Metformin, and Benzotriazole. Among the DL models, ResNet 101 exhibited the highest model performance, showing that the average validation R2 and MSE were 0.84 and 0.26 ng/L, respectively, while RF was the best in the ML models, manifesting R2 and MSE values of 0.69 and 0.58 ng/L. The trained models showed accurate training and validation results for the estimation of the five micropollutant concentrations. Therefore, this study demonstrates that the suggested analysis has a potential for alternative micropollutant measurement that has rapid and economic vantages.

Using convolutional neural network for predicting cyanobacteria concentrations in river water.

Machine learning modeling techniques have emerged as a potential means for predicting algal blooms. In this study, synthetic spatio-temporal water quality data for a river section were generated with a 3D water quality model and used to investigate the capability of a convolutional neural network (CNN) for predicting harmful cyanobacterial blooms. The CNN model displayed a reasonable capacity for short-term predictions of cyanobacteria (Microcystis) biomass. In the nowcasting of Microcystis, the CNN performance had a Nash-Sutcliffe Efficiency (NSE) of 0.87. An increase in the forecast lead time resulted in a decrease in the prediction accuracy, reducing the NSE from 0.87 to 0.58. As the spatial observation density increased from 20% to 100% of the input image grids, the CNN prediction NSE had improved from 0.70 to 0.84. Adding noise to the data resulted in accuracy deterioration, but even at the noise amplitude of 10%, the accuracy was acceptable for some applications, with NSE = 0.76. Visualization of the CNN results characterized its performance variations across the studied river reach. Overall, this study successfully demonstrated the capability of the CNN model for cyanobacterial bloom prediction using high temporal frequency images.

Estimation of heavy metals using deep neural network with visible and infrared spectroscopy of soil.

Heavy metal contamination in soil disturbs the chemical, biological, and physical soil conditions and adversely affects the health of living organisms. Visible and near-infrared spectroscopy (VNIRS) shows a potential feasibility for estimating heavy metal elements in soil. Moreover, deep learning models have been shown to successfully deal with complex multi-dimensional and multivariate nonlinear data. Thus, this study implemented a deep learning method on reflectance spectra of soil samples to estimate heavy metal concentrations. A convolutional neural network (CNN) was adopted to estimate arsenic (As), copper (Cu), and lead (Pb) concentrations using measured soil reflectance. In addition, a convolutional autoencoder was utilized as a joint method with the CNN for dimensionality reduction of the reflectance spectra. Furthermore, artificial neural network (ANN) and random forest regression (RFR) models were built for heavy metal estimation. Principal component analysis was utilized for dimensionality reduction of the ANN and RFR models. Among these models, the CNN model with convolutional autoencoder showed the highest accuracies for As, Cu, and Pb estimates, having R2 values of 0.86, 0.74, and 0.82, respectively. The convolutional autoencoder disentangled the relevant features of multiple heavy metal elements and delivered robust features to the CNN model, thereby generating relatively accurate estimates.

Evaluating physico-chemical influences on cyanobacterial blooms using hyperspectral images in inland water, Korea.

Understanding harmful algal blooms is imperative to protect aquatic ecosystems and human health. This study describes the spatial and temporal distributions of cyanobacterial blooms to identify the relations between blooms and environmental factors in the Baekje Reservoir. Two-year cyanobacterial cell data at one fixed station and four remotely sensed distributions of phycocyanin (PC) concentrations based on hyperspectral images (HSIs) were used to describe the relation between the spatial and temporal variations in the blooms and the affecting factors. An artificial neural network model and a three-dimensional hydrodynamic model were implemented to estimate the PC concentrations using remotely sensed HSIs and simulate the hydrodynamics, respectively. The statistical test results showed that the variations in the cyanobacterial biomass depended significantly on variations in the water temperature (slope = 0.13, p-value < 0.01), total nitrogen (slope = -0.487, p-value < 0.01), and total phosphorus (slope = 20.7, p-value < 0.05), whereas the variation in the biomass was moderately dependent on the variation in the outflow (slope = -0.0097, p-value = 0.065). Water temperature was the main factor affecting variations in the PC concentrations for the three months from August to October and was significantly different for the three months (p-value < 0.01). Hydrodynamic parameters also had a partial effect on the variations in the PC concentrations in those three months. Overall, this study helps to describe spatial and temporal variations in cyanobacterial blooms and identify the factors affecting the variation in the blooms. This study may play an important role as a basis for developing strategies to reduce bloom frequency and severity.