Wavelet transform couple long short-term memory neural network for profiling of Ganga water pollution and its prediction approaches

Abstract

The River Ganges is the most pious and holds significant regional, cultural, spiritual, and economic strength. It is the lifeline for millions of people and directly influences their lifestyle and livelihood. However, illegitimate anthropogenic practices severely diminish its antiquity and rank this riverine system as the fifth most polluted river in the world. To alleviate it, many scientists have investigated and attempted to predict water pollution levels using various approaches. Most of these studies are focused on univariate prediction and perform poorly when it comes to predicting multiple river water pollutants. The present investigation focused on the Ganga water pollution profiling based on physiological parameters (DO, BOD, TDS, Conductivity & Metal analysis) on a monthly dataset for three years. Furthermore, it proposes a Wavelet-LSTM model that may achieve a balance between local univariate prediction accuracy and overall accuracy to observe the trend of river water pollution more widely. The model integrates signal processing and deep learning techniques by utilizing wavelet decomposition at a specific scale to obtain the low and high-frequency features of the target data, constructing a feature matrix, and feeding it into an LSTM network for prediction. The model was used to make predictions on the DO and BOD dataset of Ganga pollution in Uttar Pradesh, India. The observed results indicate that the proposed model obtained robust performance for multiple water pollutant factors. The R² value ranges between 94 % and 99 % for the complete prediction of all pollutants. This investigation shows the effectiveness of the model, which can give theoretical along with practical observations for the forecast towards prevention, control, and mitigate the overall river water pollution and stabilize the sustainable environment for the future.