Scholarlink

<p>Dissolved oxygen (DO) in reservoirs regulate biodiversity, nutrient biogeochemistry, water quality, and greenhouse gas emissions. Maintaining healthy DO levels is essential for achieving United Nations Sustainable Development Goals (SDGs), particularly SDG 6 (Clean Water and Sanitation) and SDG 14 (Life Below Water). However, our full understanding of long-term DO dynamics of global reservoirs remains unknown, due to limited observations. Here, we develop a satellite-based machine learning model to reveal DO dynamics of major reservoirs (area &gt; 100 km<sup>2</sup>) at the global scale. We first based on continuous DO in-situ records (containing ~ 32,065 samples) to comprehensively evaluate the performance of estimating DO using three widely-used machine learning methods (e.g., Random Forest, RF; eXtreme Gradient Boosting, XGBoost; and Support Vector Regression, SVR). The RF outperforms other methods and can reliably estimate DO with R<sup>2</sup> = 0.73 and RMSE = 1.23&#xa0;mg/L in testing set. Our results demonstrate that global reservoirs show widespread deoxygenation (74%, 264 out of 357) from 1984 to 2023, with an average DO decreasing rate of 0.13&#xa0;mg/L per decade, which is faster than that observed in the lakes, oceans, and rivers. Reservoir DO exhibits pronounced spatial heterogeneity, with DO in cold northern systems is approximately 1.5 times that of in tropical and regions, reflecting latitudinal, climatic, and continental contrasts. These rapidly declining DO are mainly controlled by climate changes (contributing ~ 46%), human perturbations (contributing ~ 31%, through land use change and nutrient inputs), and the biogeochemical processes (contributing ~ 23%, through primary production and turbidity), as quantified by a state-of-the-art machine learning-based attribution analysis (SHapley Additive exPlanations, SHAP). Our study presents a practical method for spatiotemporal reconstruction of global reservoir DO dynamics using remote sensing and contributes to better understanding of driving factors behind DO changes in major reservoirs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Satellites reveal widespread deoxygenation of large global reservoirs from 1984 to 2023

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月29日

查看全文
  • Liangwei Liao,
  • Xinge Cai

摘要

Dissolved oxygen (DO) in reservoirs regulate biodiversity, nutrient biogeochemistry, water quality, and greenhouse gas emissions. Maintaining healthy DO levels is essential for achieving United Nations Sustainable Development Goals (SDGs), particularly SDG 6 (Clean Water and Sanitation) and SDG 14 (Life Below Water). However, our full understanding of long-term DO dynamics of global reservoirs remains unknown, due to limited observations. Here, we develop a satellite-based machine learning model to reveal DO dynamics of major reservoirs (area > 100 km2) at the global scale. We first based on continuous DO in-situ records (containing ~ 32,065 samples) to comprehensively evaluate the performance of estimating DO using three widely-used machine learning methods (e.g., Random Forest, RF; eXtreme Gradient Boosting, XGBoost; and Support Vector Regression, SVR). The RF outperforms other methods and can reliably estimate DO with R2 = 0.73 and RMSE = 1.23 mg/L in testing set. Our results demonstrate that global reservoirs show widespread deoxygenation (74%, 264 out of 357) from 1984 to 2023, with an average DO decreasing rate of 0.13 mg/L per decade, which is faster than that observed in the lakes, oceans, and rivers. Reservoir DO exhibits pronounced spatial heterogeneity, with DO in cold northern systems is approximately 1.5 times that of in tropical and regions, reflecting latitudinal, climatic, and continental contrasts. These rapidly declining DO are mainly controlled by climate changes (contributing ~ 46%), human perturbations (contributing ~ 31%, through land use change and nutrient inputs), and the biogeochemical processes (contributing ~ 23%, through primary production and turbidity), as quantified by a state-of-the-art machine learning-based attribution analysis (SHapley Additive exPlanations, SHAP). Our study presents a practical method for spatiotemporal reconstruction of global reservoir DO dynamics using remote sensing and contributes to better understanding of driving factors behind DO changes in major reservoirs.