Long-term Continuous Monitoring of Dissolved Oxygen Concentration Based on Multi-Spectral Sensors and Machine Learning

Abstract

Dissolved oxygen (DO) is a crucial indicator of water quality and requires continuous monitoring across various applications. Although optical sensors are widely used for DO measurement, their large-scale deployment for long-term monitoring remains challenging due to high costs and the need for periodic replacement of sensing probes. This study presents a novel non-contact DO monitoring system that integrates a low-cost multispectral sensor with an optimized machine learning framework, offering a practical solution for long-term, continuous monitoring. A compact spectroscopic sensing unit was developed to continuously acquire absorbance data from water samples across 18 wavebands ranging from 410 to 940 nm. Multiple machine learning models were trained under different configurations, and several waveband selection algorithms were applied to identify the optimal predictive model. The neural network regression model utilizing four wavebands (460, 585, 680, and 760 nm) achieved the best result with a coefficient of determination and a root mean square error of 0.99 and 0.22 mg/L, respectively. These findings demonstrate the high accuracy and practical potential of the proposed system for long-term DO monitoring in aquaculture and environmental applications.