Identifying Parameters in Water Potability Analysis Through Machine Learning

Predicting whether the water is potable or not can be helpful for people who are reliant on bodies of water and redirect them to safer options. It will also be beneficial to apply the algorithm to other places where it is expensive and inefficient to send people out and collect water samples. Over the past couple of decades, researchers have often commented on the lack of funding as a source of error when it comes to data analysis and the accuracy of the research.

Machine learning has become a rapid and prevalent tool in the growth of environmental science. With the onset of emerging data, data analysts have relied on ML models to reveal hidden trends. In past papers, machine learning has been brought up countless times as an innovative method to rapidly sort through data in big groups. However, developing an artificial neural network model is a sparsely used tool in research due to its lack of human interface, risk factors, and not enough experimentation in real-life applications. In this research, we implemented multiple ANN models to identify factors in a water’s chemical levels to identify which has the largest weight or bias in determining whether the water is potable or not. Subsequently, we created a model that uses these parameters to analyze and predict whether a water source will be safe to drink. Water is an essential resource for all beings on Earth, and as the climate changes, the water levels rise, and along comes pollutants to once safe drinking water. By comparing the accuracy between both classifier and regressor ML models, each model was hyper-parameterized to develop a better accuracy, as well as identify the weight each parameter carries towards the potability of the source. The results demonstrate a promising conclusion that the number of solids affects the potability versus other chemical levels. That said, there was not a strong, distinct correlation and every factor plays an important part. The accuracy of these models returned a sub-par 64.77% accuracy with a 0.53 precision, a 0.46 recall score, and a 0.49 f1 score. In conclusion, while the accuracy of the models is ineffective to be used just yet on a bigger scale, they offer a stepping stone in the development of machine learning models in the environmental science fields, resourcing digital infrastructures to analyze and dissect the Earth and its trajectory through climate changes.

Explore More!

Research Paper