Gated-LNN: Gated Liquid Neural Networks for Accurate Water Quality Index Prediction and Classification

Abstract

Surface water quality is of utmost significance to ensure public health and facilitate sustainable economic development. Traditional water quality assessment methods are typically time-consuming and labor-intensive and require numerous field measurements and laboratory analyses, which are costly and impractical to implement in large-scale water quality monitoring. Recent advances in machine learning (ML) have brought new approaches to predicting water quality index (WQI) and classifying water quality in real time to enhance decision-making in environmental management. In this study, we propose a novel gated liquid neural network (gated-LNN) that can predict WQI and classify water quality with high accuracy. As opposed to typical ML models, the proposed gated-LNN includes a gating mechanism that enhances temporal learning and noise robustness, making it well-suited for dynamic environmental data. For ascertaining the effectiveness of the proposed approach, we conducted rigorous experiments on a publicly available water quality dataset with 1897 examples collected from varied water bodies of India between the years 2005 and 2014. The dataset comprises seven most significant parameters of water quality, i.e., dissolved oxygen, pH, conductivity, biological oxygen demand, nitrate, fecal coliform, and total coliform. The proposed gated-LNN model achieved a high R2 of 0.9995 for WQI prediction and 99.74% accuracy for three-class water quality classification into Good, Poor, and Unsuitable classes, outperforming state-of-the-art models in both regression and classification tasks. While these results highlight the model's potential as a highly accurate and efficient tool for real-time water quality assessment, its generalizability to different regions remains an important consideration. Future work will focus on enhancing computational efficiency and conducting generalization tests on datasets from diverse geographic regions and time periods to evaluate adaptability.