Drinking water potability prediction using machine learning approaches: a case study of Indian rivers

One of the most critical aspects of a healthy environment is water quality (WQ). Clean water is essential for the survival of a wide range of plants and animals. Though it may appear unrelated at first, our land-based activities impact WQ. Pollutants, excess nutrients from fertilizers, and silt are commonly transferred into local lakes and rivers by runoff from cities and agricultural fields (Water Quality 2022a). Since rivers are more accessible than other water sources, they have become a better-suited option for the growth of the human society (Motagh et al. 2017). Due to these reasons and the fact that the ocean and groundwater come with their own set of hindrances: ocean water is salty, is unsuitable for drinking, and difficult to transport, and groundwater in many places has a slow recharge rate. Hence, the utilization of rivers has received a lot more attention. There has been an eminent amount of research in this area, and a new field of engineering called river engineering has been proposed, which contains studies on morphological changes, WQ, sediment movement, and transmission of pollutants (Julien 1992). In earth sciences, the investigation of river WQ is a popular subject (Kashefipour 2002). Detailed quality criteria are used to assess the quality of most ambient water bodies, such as rivers, lakes, and streams. Water specifications for various applications/uses also have their standards. Irrigation water, for example, must not be overly saline or contain hazardous elements that can be passed to plants or soil, harming ecosystems. WQ for industrial applications necessitates a variety of qualities depending on unique industrial activities. Natural water resources, such as ground and surface water, are some of the cheapest freshwater sources. Human/industrial activity and other natural processes can pollute such resources (Aldhyani et al. 2020).

According to the 2022 World Health Organization (WHO) report, over 2 billion people inhabit regions facing severe water stress, a condition expected to worsen in certain areas due to the effects of climate change and population growth. While the public expresses concerns about emerging pollutants like pharmaceuticals, pesticides, per- and polyfluoroalkyl substances, and microplastics, it is noteworthy that the most significant chemical risks in drinking water still arise from substances such as arsenic, fluoride, and nitrate. The contamination of water sources by microorganisms remains a pressing issue, leading to the spread of diseases such as cholera, dysentery, typhoid, and polio, causing an estimated number of 485,000 diarrheal-related deaths annually (Drinking-water 2022).

This has a higher fatality rate than those brought on by crimes, accidents, and terrorist acts (Clean Water for a Healthy World). As a result, it is critical to provide innovative ways for analyzing and, if possible, forecasting WQ. The WQ environment has been negatively impacted by the rapid increase in population, the industrial revolution, and the widespread use of pesticides and fertilizers (Cabral Pinto et al. 2019). As a result, having models for predicting the WQ is extremely useful for monitoring water contamination.

This paper is dedicated to the implementation of regression algorithms and artificial neural networks (ANNs) for the purpose of detecting the water quality index (WQI). This, in turn, enables us to determine whether water is suitable for consumption. Initially, a regression method is employed to assess water potability, followed by a classification process based on this metric. This study utilizes data from the Indian government website, considering the parameters typically used to ascertain WQ. The significance of this endeavor lies in its potential to identify the potability of water based on specific parameters. This, in turn, aids in addressing and eliminating the factors responsible for rendering water unfit for drinking.

In 2019, water from different locations of Al-Hilla River, southern Baghdadesh was studied. The study discovered that the water required some treatment to be suitable for marine life but was appropriate for agriculture (Al-ani 2019). Adeyemi et al. evaluated the chemical effects of leachate on the consistency of the surrounding water in Nigeria. According to the study, leachate-affected water samples had greater biochemical oxygen demand (BOD) and chemical oxygen demand (COD) levels in the dry season compared to the rainy season. Bacteria were more prevalent during the wet season than during the dry season. As a result, the findings of this study revealed that using leachate-contaminated water is hazardous and should be avoided (Sagar et al. 2015).

According to findings from the WQI (Hassan 2019), the problem of water contamination is on the rise. In response to this concern, a study conducted by (Batur & Maktav 2019; Liu et al. 2020) applied the principal component analysis (PCA) method to predict the WQ in Lake Gala, located in Turkey. What sets this research apart is the incorporation of satellite imagery, which offers a unique and innovative approach to assessing and monitoring WQ in the region. By utilizing PCA and satellite data, the study aimed to provide a more comprehensive understanding of the dynamics and trends in WQ, contributing to better-informed environmental management strategies.

Shailesh Jaloree & Rajput (2014) attempted to evaluate the WQ of the Narmada River using five WQ markers and used a decision tree approach. Another study advised developing a precise WQ forecasting system for smart agriculture utilizing the deep Bidirectional Stacked Simple Recurrent Unit (Bi-S-SRU) (Liu et al. 2020). Yan-jun & Qian (2012) used a support vector machine-based clustering model. It was discovered that this model is susceptible to vacancies.

Solanki et al. (2015) used a deep learning network model to assess and forecast the chemical components of water, particularly dissolved oxygen (DO) and pH value, which was shown to produce more accurate findings than supervised learning-based techniques. Khan and See used an ANN to construct a WQ model that incorporated the values of chlorophyll, DO, turbidity, and conductivity (Khan & See 2016).

Ali employed three evaluation strategies or assessment processes to measure the model's effects. The initial evaluation method was based on the segmentation of neural network connection weights, which determined the relevance of each network input parameter. On the other hand, the second and third assessment processes determined the most effective input with the greatest potential (Ahmed et al. 2019).

The method used by Sagan et al. (2020) combined WQ data with hyperspectral imaging, satellite data derived from laboratory analysis of grab samples, in situ real-time monitoring sensors installed in several water bodies across the Midwest, and grab sample analysis. The results showed that satellite-based and proximal sensors could give accurate estimates of optically active parameters, but the indirect estimation of non-optically active components is still difficult.

Dogo et al. (2019) concluded that despite the advantages of the extreme learning machine (ELM), its use is hardly ever investigated in this domain. According to the study, a hybrid deep learning (DL)-ELM framework was also put out as a workable choice that may be investigated further and used to find anomalies in WQ data. In a subsequent approach, Bucak & Karlik (2011) describe the design, implementation, and performance evaluations of an application developed for the real-time detection of drinking WQ using cerebellar model articulation controller (CMAC) ANNs. The CMAC ANN algorithm has a substantially faster learning rate than the multilevel perceptron (MLP) backpropagation (BP) approach.

Patel et al. (2022) used the Synthetic Minority Oversampling Technique (SMOTE) to balance the dataset as it was unbalanced. They proposed Random Forest (RF) and Gradient Boost methods achieving an accuracy of only 81%. In Aldhyani et al. (2020), long short-term memory (LSTM) and Support Vector Machine (SVM) techniques are used to predict the WQ, and the accuracy achieved was 96%. A 10k-fold cross-validation technique was performed, which resulted in an R² score of 0.99 with a possibility of ±0.08 deviation. In the study conducted by Al Duhayyim et al. (2022), a fuzzy deep neural network has been proposed for WQ prediction along with the atom search optimization technique. They used the Fuzzy Deep Convolutional Network (F-DCN) model followed by hyperparameter tuning, and the achieved accuracy was 98.1%. Rustam et al. (2022) proposed two models, namely, nonlinear autoregressive neural network (NARNET) and LSTM. They were able to achieve an accuracy of 96%.

In contrast to the previous study, our research focuses on the utilization of the XGBoost classifier for assessing water potability in Indian rivers. We conducted a comprehensive comparison of various classifiers, including ANNs. Notably, our findings demonstrate that the machine learning approach with the XGBoost classifier outperforms the ANN method. This disparity in performance underscores the superior accuracy and effectiveness of XGBoost in the context of WQ assessment for Indian rivers, as opposed to the approach employed in the earlier study.

The dataset chosen contains the values of various parameters of WQ in different regions of India collected over the years 2003–2014. The sources of these data are various Indian Government websites. Twelve different parameters and their descriptions are shown in Table 1.

Table 2 shows a few rows of the dataset, including parameters used to calculate the WQI. The dataset was taken from the Indian Government website (Water Quality 2022b). Important parameters like water temperature (Temp), pH value, electrical conductivity (EC), DO, fecal coliform and total coliform counts, and BOD are used to calculate the WQI. WQ will be defined as satisfactory or unsatisfactory based on the measured ambient concentrations and corresponding criteria.

Next, we initialize six new columns npH, ndo, nco, nbdo, nec, and nna in our data frame, which are calculated on the basis of DO, pH, EC, BOD, nitrate, fecal coliform, and total coliform. The null values in the dataset are replaced by their standard means and modes depending on the type of value that needs to be replaced.

Once the WQI is calculated, we classify the water sample into potable or non-potable water based on its value, as shown in Table 3 (Wong & Rylko 2014).

We have proposed different machine learning models like KNN, XGBoost, and RF to get the best performance. We have also used an ANN, an advanced machine learning approach to get the model performance.

We consider five scenarios, as shown in Table 4. Each scenario takes one more input variable than the previous one to observe how the presence of each input parameter affects the accuracy, precision, recall, and F1 score of the model.

Machine learning is a field of study in computer science where the machine learns from data and gives out predictions and insights. There are many machine learning algorithms formulated by mathematical and statistical concepts that help the machine predict or derive insights from the fed information. Then, these insights influence critical growth and key performance measures through decision-making within applications and companies. The need for data scientists will increase as Big Data develops, requiring their assistance in determining the most pertinent business issues and, as a result, the data required to answer them.

Our problem statement is a classification problem that finds the portability of water. Various models can be used to solve a classification problem. Some of the approaches being used are as follows.

KNN: K neighbors-based categorization is a form of lazy learning because it only saves instances of the training data rather than attempting to create a general internal model. Each point is classified by a simple majority vote of its KNN. According to to its name, the KNN algorithm works in such a way that a new case is compared to its closest neighbors, which already exist in the model weights, thus matching the name ‘nearest neighbors’. This algorithm is one of the simplest and most intuitive algorithms under supervised learning.

The KNN algorithm uses the previously stored data points to classify and categorize the new data points. Therefore, the algorithm is helpful in classifying and categorizing new data into previously well-defined classes. The algorithm can be used for both regression as well as classification problems, but it is more common to use it for classification purposes. Following a nonparametric approach, the KNN algorithm does not assume anything about the data.

RF: The RF classifier, a meta-estimator, uses an average to increase the model's predictive accuracy and reduce overfitting by fitting several decision trees on various subsamples of datasets. The samples are generated via replacement; however, the subsize of the sample is always the same as the original input sample's size. Because of their simplicity, on the one hand, and typically excellent performance, on the other, RFs, like naive Bayes and KNN algorithms, are well-liked. In contrast to the first two methods, RFs are somewhat unpredictable in terms of how the trained model will be structured. Due to the stochastic nature of tree construction, this is an inevitable result. One of the main reasons why this feature of RFs can be problematic in regulatory settings is that clinical adoption often necessitates a high level of repeatability, not only in terms of the algorithm's ultimate performance but also in terms of the mechanics by which a particular decision is made. The decision tree is the basic building block of RF classifiers. A decision tree is a hierarchical structure created from a dataset's characteristics (or independent variables). The decision tree is divided into nodes based on a measure connected with a subset of the features. A RF is built upon multiple decision trees. A decision tree is a machine learning algorithm where a tree is created based on attribute impact hierarchy created from the dataset's characteristics. The nodes of the tree are divided. The RF is made up of a set of decision trees that are linked to a set of bootstrap samples created from the original dataset. The nodes are divided using the entropy of a subset of the characteristics. Subsets from the original dataset that are the same size as the original dataset are created via bootstrapping. Through the idea of overlap thinning, the bootstrapping technique makes it easier to create a RF with the appropriate number of decision trees to improve classification accuracy. After that, the best trees are chosen through a voting process and a process called bagging (bootstrap aggregate).

XGBoost: The XGBoost classifier got the greatest results after attempting several different methods to get the best performance out of our model. XGBoost is an ensemble learning technique. Relying just on a single machine learning model's results might not always be sufficient. A technique for systematically combining the prediction skills of several learners is ensemble learning. The final model integrates the outputs of various other models into one. In boosting, the trees are constructed one after the other, with each succeeding tree attempting to minimize the errors of the one before it. Since it is a boosting method, each tree improves its accuracy by learning from the errors of the previous tree. In boosting strategy, the weak learners are combined into strong learners as each weak learner provides information which makes the next model stronger. The last strong learner lowers the bias as well as the variance. Boosting employs trees with fewer splits compared to bagging methods such as RF, which uses trees that have been fully formed. Such small trees, which are not very deep, are quite simple to interpret. We can enhance the capabilities of this model by performing some hyperparameter tuning such as controlling the number of trees, adjusting the learning rate that affects the speed and efficiency with which the model reaches global minima, and controlling the depth of the tree as well so that bias in the model is not increased. There could be overfitting if there are a lot of trees present. Therefore, hyperparameter tuning helps in reducing the chances of overfitting and gives a smooth model which gives accurate results.

Using the train test split function from the sklearn package, we divided our dataset into test and train datasets before implementing the XGBoost classifier. This assists in dividing our dataset in the desired ratio so that a portion is utilized to develop our machine learning model and the remaining portion is used to test our model. By determining how closely our actual values match our predicted values, we can then determine the model's accuracy.

ANNs are a set of interconnected neurons that imitate the neurons of the biological brain. Just like the animal brain where the neurons are interconnected, in ANNs, the nodes are connected to each other in a similar manner that exchange information with one another despite being on different layers. Edges are the connections that form between two neurons or nodes. The weight of the neurons and edges often changes as learning progresses. These weights control how strong a signal is at a connection. In order to develop probability-weighted correlations between the two that are stored within the network's data structure, neural networks learn (or are trained) by examining samples with known ‘input’ and ‘output.’ When training a neural network from a given sample, it is usually assessed how the network's processed output (generally a prediction) differs from a target output. The flaw is this difference. Using this error value and a learning technique, the network then adjusts its weighted associations. The output produced by the neural network will get even more similar to the target output with each adjustment. When enough of these improvements have been made, the training is adjudicated based on predetermined criteria. The architecture of a neural network primarily consists of three layers:

• The input layer: This layer accepts inputs in several different formats as provided by the user.

• Hidden layers: This layer, which is concealed between the input and output layers, may include one or more layers. All computations are made in order to uncover patterns and hidden features.

• Output layer: This layer is finally produced using this layer after the input layer has undergone a number of changes using the hidden layer.

There are 123 true-positive values and 244 true-negative values, as shown in the matrix. We do hyperparameter adjustments to further increase the accuracy of our model. Hyperparameter tuning is the process of identifying the ideal model architecture from a set of parameters known as hyperparameters, which determine the model architecture. The model parameters indicate how to convert the input data into the desired output, whereas the hyperparameters control how our model is organized. Upon using GridSearchCV to find the best suitable parameter values from our given parameter grid, we apply these values as the hyperparameter values in our model to increase its accuracy as shown in Table 6.

In our study, we conducted a comprehensive assessment of the predicted WQI values in comparison to the actual, observed values shown in Table 7. This evaluation is vital for gauging the accuracy and reliability of the models or methods employed for WQI estimation. By meticulously examining the disparities between our predictions and the ground truth, we gain insights into the effectiveness and precision of our chosen approach. This comparison not only helps us to understand the trustworthiness of our model but also holds significance for WQ management, with far-reaching implications for environmental and public health.

Table 8 contains the accuracy, precision, recall value, F1 score, and AUC-ROC, which have been calculated for both our models, and it is evident that the XGBoost model is a better approach when analyzing the potability of water.

To illustrate the importance of the proposed method suggested in this study, a comparison of its performance was conducted. The results presented in Table 9 clearly demonstrate that, in both scenarios, the proposed approach outperformed previous studies.

This paper focuses on the implementation of machine learning algorithms and ANNs for the detection of the WQI, which further helps us to conclude if the water is viable or not. Machine learning implies using already available data to make predictions further. Here, since we know the parameters that are used to determine the WQI, we take data from the Indian government site to predict the potability of water. It is useful as if we can determine the potability of water for a given amount of the parameters the waste or causes that make the water unfit for drinking can be eliminated. There were five different machine learning approaches, such as KNN, XGBoost with and without hyperparameter tuning, RF, and sequential ANN were tested, to see which one would be the most optimal for our cause, that is predicting the potability of water. Out of all five models, the most accuracy was obtained for hyperparameter tuning to the XGBClassifier model. The model has given an accuracy of 98.93%, an F1 score of 98.42%, a precision of 100%, and a recall of 95.60%. This model can be implemented to create a website or mobile application to help people check the WQI. It can also be integrated with environmental monitoring systems and Internet of Things (IoT) devices for automated data collection. This can be used to send alerts if the WQI of a region being monitored drops below a threshold so that preventive measures can be taken well ahead.

With the steady increase in global population and industrialization, the imperative to monitor and preserve WQ is more pressing than ever. Water, a fundamental resource for life and industry, necessitates vigilant oversight. By deploying the designed model over cloud services with an IoT device embedded with an array of essential sensors including pH, turbidity, DO, and temperature sensors, we can enhance the capabilities of both, making the whole ecosystem more efficient and helpful. These sensors are strategically positioned to monitor the WQ at its source, allowing us to observe critical parameters directly impacting purity. The real-time connectivity of these devices to predictive models will empower us to swiftly detect deviations in the WQ and anticipate potential issues. Employing data visualization techniques will help provide comprehensive insights into the state of water. This technological advancement is pivotal not only in safeguarding public health but also in promoting the sustainable management of industries dependent on water. As we confront environmental challenges and the looming impacts of climate change, this technology takes a proactive stance toward addressing WQ concerns. By bridging the worlds of IoT, data analysis, and visualization, this solution equips us to better protect and preserve one of our most indispensable resources, water.

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

The authors declare there is no conflict.