Landsat data reveal lake deoxygenation worldwide.

ABSTRACT

Dissolved oxygen (DO) is a fundamental requirement for the survival of aquatic organisms, which plays a crucial role in shaping the structure and functioning of aquatic ecosystems. However, the long-term DO change in global lakes remains unknown due to limited available data. To address this gap, we integrate Landsat data and geographic features to develop DO estimation models for global lakes using machine learning approaches. The results demonstrated that the trained random forest (RF) model has better performance (R2 = 0.72, and RMSE = 1.24 mg/L) than artificial neural network (ANN) (R2 = 0.66, and RMSE = 1.39 mg/L), support vector machine regression (SVR) (R2 = 0.62, and RMSE = 1.45 mg/L) and extreme gradient boosting (XGBoost) (R2 = 0.72, and RMSE = 1.29 mg/L). Then, we used the trained RF model to reveal the DO long-term (1984-2021) change in surface water (epilimnetic) of 351,236 global lakes with water area ≥ 0.1 km2. The results show that the average epilimnetic DO concentration of global lake was 9.72 ± 1.07 mg/L, with higher DO in the polar regions (latitude > 66.56 °) (10.87 ± 0.54 mg/L) and lower in the equatorial regions (-5 ° < latitude < 5 °) (6.29 ± 0.63 mg/L). We also find widespread deoxygenation in surface water of global lakes, with a rate of - 0.036 mg/L per decade. Meanwhile, the number of lakes and surface area that experiencing DO stress are continuously increasing, with rate of 39 and 212.85 km2, respectively. Our results offer a comprehensive dataset of DO variation spanning nearly 40 years, furnishing valuable insights for formulating effective management strategies, and enhancing the maintenance of the health of aquatic ecosystems.