Dissolved Oxygen Inversion Based on Himawari-8 Imagery and Machine Learning: A Case Study of Lake Chaohu

Abstract

Dissolved oxygen (DO) concentration is a widely used and effective indicator for assessing water quality and pollution in aquatic environments. Continuous and large-scale inversion of water environments using remote sensing imagery has become a hot topic in water environmental research. Remote sensing technology has been extensively applied in water quality monitoring, but its limited sampling frequency necessitates the development of a high-frequency dynamic water quality monitoring model. In this study, we utilized Lake Chaohu as a case study. Firstly, we constructed a dynamic water quality inversion model for monitoring DO concentrations using machine learning methods, with Himawari-8 (H8) satellite imagery as input data and DO concentrations in Lake Chaohu as output data. Secondly, the developed DO concentration inversion model was employed to estimate the overall grid-based DO concentration in the Lake Chaohu region for the years 2019 to 2021. Lastly, Pearson correlation analysis and significance tests were performed to examine the correlation and significance between the estimated grid-based DO concentration and the ERA5 reanalysis dataset. The results demonstrate that the Random Forest (RF) model performs best in DO concentration inversion, with a high R2 score of 0.84, and low RMSE and MAE values of 0.69 and 0.54, respectively. Compared to other models, the RF model improves average performance with a 38% increase in R2, 13% decrease in RMSE, and 33% decrease in MAE. The model accurately predicts DO concentrations. Furthermore, the inversion results reveal seasonal differences in DO concentrations in Lake Chaohu from 2019 to 2021, with higher concentrations in spring and winter, and lower concentrations in summer and autumn. The average DO concentrations in the northwest, central-south, and northeast regions of Lake Chaohu are 10.12 mg/L, 9.98 mg/L, and 9.96 mg/L, respectively, with higher concentrations in the northwest region. Pearson correlation analysis indicates a significant correlation (p < 0.01) between DO concentrations and temperature, surface pressure, latent heat flux from the atmosphere to the surface, and latent heat flux from the surface to the atmosphere, with correlation coefficients of −0.615, 0.583, −0.480, and 0.444, respectively. The results verify the feasibility of using synchronous satellites for real-time inversion of DO concentrations, providing a more efficient, economical, and accurate means for real-time monitoring of DO concentrations. This study has practical value in improving the efficiency and accuracy of water environmental monitoring.

Keywords

• Himawari-8
• machine learning
• water quality monitoring
• dissolved oxygen