Prediction of nitrate in different catchments using domain adaptation for regression method

Abstract

Surface water quality is increasingly at risk due to anthropogenic activities and climate change, leading to issues such as eutrophication that threaten aquatic ecosystems and human well-being. This study harnesses the power of Artificial Intelligence (AI), specifically deep learning and domain adaptation techniques, to predict nitrate concentrations using readily measurable parameters such as electrical conductivity (EC), pH, and temperature. We propose the Multi-Domain Adaptation for Regression under Conditional Shift (DARC) framework, designed to tackle data scarcity and marginal shifts between catchments. By incorporating a Modified Pairwise Similarity Preserver (MPSP) loss function, our model achieved an NSE value of 0.44 using only seven data points from the target dataset, outperforming traditional linear regression, which failed to reach comparable performance even with more than 20 data points. This study highlights the potential of AI-based domain adaptation methods as cost-effective, scalable solutions for water quality monitoring, addressing global environmental challenges through improved prediction and management of surface water resources.