Machine learning reveals distinct aquatic organic matter patterns driven by soil erosion types

Abstract

Chromophoric dissolved organic matter (CDOM), characterized by unique optical properties, is an essential indicator for understanding aquatic organic matter dynamics within global carbon cycles. Soil erosion, a major source of CDOM received by lakes, transports terrestrial organic matter to water bodies, altering sources, bioavailability and molecular complexity of CDOM significantly. Yet, the spatial patterns of CDOM in lakes from different soil erosion regions are still unknown. Here, we developed a robust machine learning framework (RMSE_calibration = 0.87 m^-1) to estimate CDOM concentrations in lakes by integrating over 1300 in situ water samples with Landsat 8 OLI surface reflectance data. We then applied this model to map the spatial distribution of CDOM across lakes larger than 0.1 km² in 2020. Our analysis revealed distinct spatial patterns, with mean CDOM absorption coefficients at 355 nm of 3.73 m^-1 in freeze-thaw erosion regions, 6.31 m^-1 in wind erosion regions, and 3.72 m^-1 in hydraulic erosion regions, reflecting significant variations driven by erosion intensity. Two axes of PCA analysis explained over 48 % variations of CDOM for different soil erosion types. Chemical characterization indicated that polycyclic aromatic predominated in wind and hydraulic erosion regions, whereas freeze-thaw erosion regions exhibited higher proportions of peptides and unsaturated aliphatic compounds. This study highlights the crucial connection between terrestrial soil erosion processes and aquatic DOM composition, providing vital insights for evaluating global carbon cycling and carbon storage within inland ecosystems.