Abstract: Dissolved oxygen (DO) is essential for the growth, development, and community structure stability of marine organisms, and plays a crucial role in regulating the biogeochemical cycles of trace elements, toxic metals, and nutrients. Hypoxia, defined as DO<2 mg∙L^-1, threatens the health of coastal marine ecosystems and the sustainable utilization of marine resources. Due to intensifying global climate change and human activities, hypoxia in the Pearl River Estuary has become increasingly serious, resulting in substantial economic losses by damaging marine fishery resources and leading to the deterioration of key ecosystem characteristics that are vital for marine ecosystem sustainability. Accurate DO prediction is therefore essential for enhancing hypoxia prevention and management in the Pearl River Estuary. Based on time-series hydrometeorological and water quality parameters obtained from in situ measurements and numerical simulations, rapid prediction models for 24-hour-ahead, 48-hour-ahead, and 72-hour-ahead DO concentration were constructed using Randon Forest (RF), and CatBoost, respectively. These RF models demonstrated excellent performance, showing strong agreement with in situ measurements, with R²=0.965、0.921、0.889，RMSE=0.162、0.243、0.289 mg∙L^-1，MAPE=1.8%、2.9%、3.5. And these CatBoost models also performed well, aligning closely with numerical simulations, with R²=0.965、0.921、0.889，RMSE=0.162、0.243、0.289 mg∙L^-1，MAPE=1.8%、2.9%、3.5. To directly assess the hypoxia conditions, hypoxia levels were classified based on the model predictions, which could provide scientific support for water resource management and ecological protection in the Pearl River Estuary.

Key words: dissolved oxygen, machine learning, Pearl River Estuary, hypoxia