Artificial intelligence in water quality monitoring: A review of water quality assessment applications

Water quality monitoring is an important responsibility of each state, to make sure that the population has access to safe water and the needs are met without creating pressures on the water resources. UN Sustainable Development Goal 6 aims to ensure the availability and sustainable management of water and sanitation for all. It focuses on providing safe and affordable drinking water, access to proper sanitation facilities, and promoting good hygiene practices. The goal also emphasizes the importance of protecting and restoring water ecosystems to maintain water quality and encourages efficient and sustainable management of water resources.

Water management is very complex and includes monitoring water quality and quantity parameters as well as biodiversity and aquatic life, identifying pollution sources and pollutants removal, sanitation, flood protection, resource allocation, etc.

The main sources of water pollution are discharges from urban agglomerations, leakage, and run-off from agriculture and industrial activities. Water monitoring includes a large number of hydrological, physical, chemical, and biological parameters, some of which are measured on site and others by sample analysis in the laboratory. Most countries have monitoring programs that specify sampling locations, parameters to be determined and frequency of sampling that are associated with efforts and costs for labour, reagents, equipment, etc.

Each country or region has its own quality standards that define limit values for parameters and classification systems to evaluate the state of a water body and its adequacy for different uses. For example, in the European Union, this field is regulated by the Water Framework Directive (WFD) and monitoring data are reported to the European Environment Agency by all Member States. The classification of the water bodies follows the One Out – All Out principle, which means that the status is given by the ranking of the worst parameter. For example, a water body cannot have ‘good’ status if one parameter ranks as ‘poor’. Other countries have developed classification systems based on water quality indexes (WQIs), which are dimensionless numbers that aggregate the values of several selected indicators.

There are different WQIs, but the process of calculating the WQI usually includes the following steps:

• - Selection of relevant water quality parameters;

• - Assigning a weight to each parameter;

• - Calculation of sub-indexes (e.g. based on limit values for a certain quality class);

• - Aggregation of sub-indexes into the WQI.

According to the value of the WQI, the water is then categorized into quality classes, depending on the calculation method and water uses.

The WQI method is laborious and has several limitations that may lead to inaccurate results, but for many years it has been a good instrument to assess the overall water quality and its long-term trends.

Artificial intelligence (AI) refers to the simulation of human intelligence processes by machines, particularly computer systems, which include learning, reasoning, self-correction, perception, and interaction. AI can analyse in a short period of time huge amounts of data, identify patterns or anomalies, calculate indicators, provide visual representations of data, etc., which can support the assessment of water quality, as well as identification of pollution sources and remediation measures.

The recent development of AI tools for assessing water quality has the potential to bring significant improvements in this sector. It may reduce monitoring efforts and costs and increase the accuracy of WQI prediction and water quality classification (WQC) in several ways, some of which are mentioned below:

• - Advanced analysis of monitoring data, followed by calculation of WQI may indicate which parameters have the strongest influence in determining the WQI value and allow reducing the numbers of monitored parameters or their frequency;

• - Complex modelling of WQI values may allow the prediction of WQI;

• - AI algorithms can assign datasets into quality classes based on raw monitoring data without the need of calculating WQIs.

In addition, the combination of AI and remote sensing may be able in the future to replace traditional monitoring methods with satellite data and real-time, on-site sensors data (IoT – Internet of Things), reducing the cost and efforts of water quality monitoring.

The objective of this paper is to conduct a comprehensive review of recently published research utilizing AI techniques in practical WQI and WQC applications. By examining results from diverse geographical locations and datasets, this study aims to provide a detailed analysis of available tools, their applicability, limitations, and areas for future research.

The databases such as scholar.google.com, scopus.com, and sciencedirect.com have been searched for publications on ‘artificial intelligence water quality index’ and ‘machine learning water quality index’ for the period 2015–2024 using terminologies as indicated in Table 1. Ninety publications were retrieved and were further analysed in terms of addressing the topic of the review, and the information regarding the WQI methodology and AI tools that were used in the research. Fifty-six original research articles were included in the review, as they were found to include sufficient and relevant information about the water quality datasets and the methods used for processing them. The papers were reviewed for:

• - Location and type of study site;

• - Reference period of monitoring data;

• - WQI methodology and included parameters;

• - AI tools used in data processing and results regarding their performance.

WQI prediction tools are based on regression, which means that data from the past are analysed to identify patterns and relationships between independent variables (water quality parameters) and dependent variables (WQI value) and to predict how values will evolve. For a model to be reliable, it needs to be built on sufficient data that covers a relevant period of time. As an example, if water monitoring data are available for a period of 10 years, the first seven years can be used for model training, and the last three years to verify if the WQI values predicted by the model are in line with those calculated from actual data.

WQC classification tools are ML models that assign a label to a dataset, for instance ‘good water quality’. These models also need to be trained on real data and checked for accuracy.

A selection of AI tools and their application is presented in Table 2.

The research papers examined in this study encompassed a wide variety of geographical locations, time-spans of data collection, WQI calculation methods, and AI tools employed for data processing.

Three main categories can be distinguished based on the purpose and methodology of the studies:

• - Complex datasets, including a large number of parameters that are used to calculate WQIs, followed by analysis of the influence of individual parameters on the final results and comparison between results obtained with all data vs. smaller datasets that still allow accurate prediction of water quality, in order to reduce the number of variables.

• - Different methodologies for WQI calculation applied to the same dataset in order to assess their performance and improve the calculation method.

• - Different ML methods applied to the same dataset and same WQI methodology, followed by assessment of performance by comparing the results of the AI models with those given by real monitoring data.

For instance, in a study carried out on monitoring data from Thailand for the period 2016–2021, it has been possible to reduce the number of parameters for WQI calculation from 13 to four, using ANN and the Bootstrap method (Chawishborwornworng et al. 2024). A long-term study on lake water in Finland (1980–2023) has shown that the long short-term memory model was the least sensitive model when COD and TP were removed from inputs, compared with SVR, RF, and ANN, using the British Columbia WQI (Kim et al. 2024). Another study in China has allowed the reduction of the number of parameters from 22 to nine, using redundancy analysis (RDA), on a dataset from 2017 (Li et al. 2021). A study on irrigation water in Vietnam found that coliform, dissolved oxygen, turbidity, and total suspended solids are the most important parameters for water quality assessment (Lap et al. 2023). In another study, electrical conductivity had the highest influence on WQI for groundwater, while pH had the lowest influence (Raheja et al. 2022). However, there is also a report where 14 of 17 parameters included in WQI calculations were found to be significant. Moreover, the model was found to perform well also with only 12 parameters instead of 17 (Fernández Del Castillo et al. 2022). A noteworthy result is that, despite the superior performance of the 10-parameter model, a significantly simpler model incorporating only BOD, turbidity, and phosphate demonstrated a remarkably high level of accuracy in predicting river water quality (Asadollah et al. 2021).

Some of the reviewed papers aim to improve the water quality index methodology with the help of AI tools. The fuzzy-GIS-based groundwater quality index (FGQI), combining geographical information system (GIS) with the groundwater quality index, is proposed as a reliable tool for groundwater quality assessment (Jha et al. 2020). Another study made the first attempt to include heavy metal concentrations in groundwater assessment, demonstrating the robustness of the model (Sajib et al. 2023). The automated water extraction index with no shadows (AWEI-nsh) proposed by another group, using remote sensing data, may be applied when shadows are excluded from the image (Li et al. 2021). The WQI model could also be improved by adjusting parameter weight values and using new aggregation functions, namely sinusoidal weighted mean and log-weighted quadratic mean (Ding et al. 2023). In addition, using ML methods, satellite data can be used to improve water quality monitoring in coastal waters that are optically complex (Hafeez et al. 2019).

A comparison between models using a minimum number of parameters (coliform, pH, temperature, turbidity, and total dissolved solids) indicates that gradient boosting and polynomial regression performed better in predicting WQI, whereas MLP performed better in predicting WQC (Ahmed et al. 2019). However, it should be mentioned that the regression model only included a relatively short period of time (2009–2012). A study carried out by Abba et al. (2020) on a limited number of parameters indicated that the Adaptive Neuro-Fuzzy Inference System (ANFIS) model performed best at one location and BPNN at two other locations, which seems to indicate that the results may be specific to certain regions.

Two studies indicate that MLs are efficient tools to predict irrigation water quality and AdaBoost can predict all parameters (El Bilali et al. 2021; Trabelsi & Bel Hadj Ali 2022).

Several studies report successful application of regression models to predict WQI, but it is not clear how regression was possible for monitoring data collected during one year or even one set of samples collected from different locations (Kadam et al. 2019, 2022; Hussein et al. 2023; Ibrahim et al. 2023; Uddin et al. 2023b; Abbas et al. 2024). Aslam et al. (2022) used data for two years (2020–2021) and found that the hybrid RT-ANN algorithm can produce good results for short-term data, but the stability of the model would be improved using long-term datasets.

Both for regression models, as well as for classifiers, the research data show that there is no model that performs best for a majority of the datasets. In fact, almost each study reports a different method that gave the best results in their case, which seems to indicate that the outcomes depend on the AI tool, but also on the characteristics of the dataset and perhaps on the WQI or classification methodology. In each situation, there was either a different type of water or a different set of variables or WQI calculation method, so it is not possible to make comparisons between studies.

The AI tools reported by several studies to perform best in the case of WQI prediction are gradient boosting (Ahmed et al. 2019; Nguyen et al. 2023; Abbas et al. 2024), ANFIS (Tiwari et al. 2018; Hmoud Al-Adhaileh & Waselallah Alsaade 2021; Ibrahim et al. 2023), and SVM (Li et al. 2021; Shamsuddin et al. 2022; Ibrahim et al. 2023).

In the case of classifiers, several studies indicate good performance of XGBoost or other gradient boost method (Shams et al. 2023; Uddin et al. 2023b; Singh et al. 2024; Solangi et al. 2024), as well as SVM (Derdour et al. 2022; Hussein et al. 2023).

The review has revealed a diversity of datasets, water quality assessment methodologies and AI tools used for modelling, so this field appears to be very far from a world-wide standardization.

ML methods, including gradient boosting and polynomial regression, have shown effectiveness in predicting WQI, although results can vary based on the dataset and region. The research indicates that no single AI model consistently outperforms others across different datasets, suggesting that outcomes depend on the specific AI tool, dataset characteristics, and WQI methodology used. Overall, AI tools like gradient boosting, ANFIS, and SVM have been reported to perform well in WQI prediction, while classifiers like XGBoost and SVM have shown good performance in classification tasks.

The reviewed studies and the main results are presented in Table 3.

There is a wide variety of AI tools available to support WQI and WQC prediction, and some of them have been demonstrated to have high accuracy and reliability. Choosing the appropriate AI tool for WQI modelling depends on the specific requirements of the problem, the nature of the dataset, and the desired performance metrics. Ensemble methods like random forest and XGBoost, as well as neural network-based approaches like ANN and LSTM, are particularly effective for complex water quality assessment tasks. Combining these models or using hybrid approaches can further enhance prediction accuracy and robustness (Shams et al. 2023).

It should be mentioned though that the review has revealed a wide variety of datasets and locations, and that not all results may be easily replicated. Although some short-term studies may be justified by the lack of long-term monitoring data, the methods should be tested before application on different datasets.

AI and ML tools offer advantages such as the ability to process vast amounts of information and handle complex, non-linear relationships in water quality data, deal with partial or missing data, and provide real-time or near-real-time predictions, which can be crucial for water quality monitoring and management.

To make use of these recently developed tools, water managers may develop gradual approaches to reduce costs and improve the efficiency of monitoring, including:

• - Analysis of historical datasets to identify critical parameters, pollution hotspots, and water quality trends;

• - Integration of the spatial dimension (GIS) and real-time monitoring systems (sensors);

• - Set alerts for key parameters that indicate when threshold values are exceeded for quick intervention;

• - Develop new models for data integration and/or WQC based on individual parameters' inputs.

Some AI tools that can be applied include linear and logistic regression, decision trees, random forests, support vector machines, neural networks, etc.

Open access to water quality monitoring data has allowed data scientists to make fast and significant progress in developing new tools for extracting relevant information from the data, which can further support authorities to identify pollution sources, design remediation measures, and assess their efficiency in improving water quality.

In general, a model trained on data specific to a certain region is only able to produce accurate results for that region. If more monitoring data would become available, further research could explore how models developed in one part of the world perform in other regions and even contribute to a global model. Combinations of traditional monitoring and real-time data from in-situ probes and remote sensing could lead to quick identification of pollution sources and allow early interventions for remediation. In addition, climate change considerations could improve the accuracy of the models.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.