Feilong Shen, Jifeng Li, Tianwei Wang, Nian Li, Jiawei Yang, Mengyang Xu, Yange Yan, Linhuan He, Yulong Guo, Zhongbing Chen
• Landsat and XGBoost reconstructed 40-year DOCc and DOCs in Lake Taihu. • Persistent spatial heterogeneity and nearshore–offshore DOC gradients were resolved. • Lake-wide DOCs peaked in 2010 (16,906 t C) and declined at −138.51 t C/decade. • Meteorological factors dominated DOC variability over other driver groups. Dissolved organic carbon (DOC) plays a crucial role in lake carbon cycling, yet its long-term variations and key drivers remain poorly understood, particularly in large eutrophic lakes. Using 40 years of Landsat data and four machine learning (ML) models, we reconstructed DOC concentration (DOCc) and storage (DOCs) in Lake Taihu for the whole lake (T0) and six open-water zones: Gonghu Bay (T1), Meiliang Bay (T2), Zhushan Bay (T3), Central Lake (T4), Northwest Lake (T5), and Southwest Lake (T6). The XGBoost model achieved high accuracy (validation: R 2 = 0.73, RMSE = 0.51 mg/L, MAPE = 8.66%, RPD = 1.70), revealing spatiotemporal DOC variability (2.49–6.36 mg/L). DOCc peaked at 4.90 mg/L in 1996 and declined after 2000, with lower levels in several zones after 2010. Lake-wide DOCs showed interannual variability and a slight decrease, peaking at 16,906 t C in 2010 and showing a linear trend of −138.51 t C/decade. Regional differences highlight strong hydrodynamic flushing in T6 (−106.63 t C/decade), while localized increases in T1 and T5 may reflect nutrient-driven productivity and sedimentary carbon accumulation. Meteorological factors were the strongest direct controls on DOCc (−0.72) and DOCs (−0.98), exceeding human-activity proxies and other driver groups. These results highlight the potential of combining remote sensing with ML to provide a spatially explicit, long-term baseline for DOC monitoring and attribution in large optically complex lakes, thereby improving understanding of lake carbon dynamics under anthropogenic pressures and climate change