Machine learning for long-term reconstruction and attribution of carbon and water-quality dynamics in inland waters

Feilong Shen

Doctoral thesis

Abstract:

Inland waters play disproportionate roles in global carbon and nutrient cycles, yet their long-term dynamics remain poorly observed. Field monitoring networks provide direct measurements but are spatially sparse, temporally discontinuous, and particularly limited for small rivers that integrate multiple catchment pressures. Satellite remote sensing and machine learning (ML) can extend observations across space and time. Still, their feasibility depends on three conditions that are rarely evaluated together: whether the target variable is optically expressed, whether the sensor adequately resolves the water body, and whether model validation accounts for spatial dependence. This dissertation therefore aimed to reconstruct and attribute long-term changes in carbon dynamics and inland water quality using satellite observations and ML, while defining the physical and statistical limits within which such inferences remain scientifically defensible. Harmonized surface-reflectance data from Landsat 5, 7, 8, and 9, Sentinel-2, Harmonized Landsat and Sentinel-2 (HLS), and MODIS-Aqua were processed in Google Earth Engine. The analytical workflow included cloud screening, cross-sensor harmonization, and temporal matching with extensive in situ records. Object-level extraction, effective-pixel gating, and spatially explicit validation were used to reduce contamination and evaluate model transferability. Two contrasting systems were examined: eutrophic Lake Taihu in China and 103 width-stratified agricultural river reaches in South Bohemia, Czechia. Study 1 reconstructed dissolved organic carbon (DOC) concentration and storage in Lake Taihu from 1984 to 2023 using 200 Landsat–in situ matchups. Extreme gradient boosting (XGBoost) showed the best generalization among four evaluated algorithms, achieving a coefficient of determination (R2) of 0.73 on an independent validation dataset. The estimates were harmonized against MODIS-Aqua and combined with reconstructed lake water storage. Lake-wide DOC concentration averaged approximately 4.72 mg L-1 and followed a three-phase temporal trajectory, whereas DOC storage declined slightly over the full period and exhibited substantial spatial heterogeneity. Among the evaluated environmental drivers, structural equation modelling identified meteorological forcing as the dominant correlate of DOC concentration and storage, with human activity, bio-optical conditions, and vegetation contributing secondary or interacting influences. Study 2 evaluated the retrieval of ammonium nitrogen (NH4+-N), nitrate nitrogen (NO3–-N), and total phosphorus (TP) using 178,286 quality-gated reach-object observations and 52,521 in situ measurements. Under leave-one-site-out spatial cross-validation, R2 values were −0.03, 0.05, and 0.11, respectively, whereas random cross-validation inflated apparent performance by as much as ΔR2 = 0.57. The principal constraint was geometric: the median narrow reach occupied only 16% of a 30 m pixel, leaving measured reflectance dominated by land mixing and adjacency effects. The long-term in situ records nevertheless showed declines of approximately 80%, 48%, and 66% in the three nutrient variables, followed by a plateau around 2007. This recovery was regionally coherent and broadly consistent across river-width classes. At the same time, none of nine multi-scale tests based on riparian land-cover change explained the observed pattern under spatial validation. Study 3 compared direct optical retrieval of chemical oxygen demand (COD) and five-day biochemical oxygen demand (BOD5) with a random-forest model driven by satellite-observable catchment attributes and point-source pressures. The modelled quantity was concentration rather than flux. Optical models failed to transfer to unvisited reaches, producing Nash–Sutcliffe efficiencies of 0.02 for COD and 0.03 for BOD5, with no improvement from native 10 m imagery. In contrast, the catchment model achieved efficiencies of 0.

Abstract:

29 and 0.40 and Kling–Gupta efficiencies of 0.33 and 0.47, respectively, on the log₁₀ scale. BOD5 prediction was strongest in the narrowest reaches. Reconstruction for 1985–2025 indicated an initial decline followed by a partial rebound, while the 2050 sensitivity analysis suggested that land-use management exerted a substantially stronger influence than warming within the tested scenarios. Collectively, the three studies move inland-water monitoring beyond algorithm-centred accuracy comparisons towards a scale-aware framework of retrievability. They establish a double inseparability boundary. The first is an optical boundary, at which weak constituent signals cannot be separated from sub-pixel land contributions and adjacency contamination. The second is an attributional boundary, at which spatial dependence and regionally coherent forcing prevent basin-wide change from being decomposed into local causes. A near-zero score under spatial validation is therefore not merely evidence of an inadequate algorithm; it can indicate that the requested information is unavailable at the selected observational scale. Importantly, the dissertation also identifies a pathway beyond this limit: when water-surface reflectance cannot support defensible direct retrieval, catchment context can recover part of the missing information. Its central contribution is threefold: providing long-term reconstruction where optical and spatial conditions permit it, quantitatively diagnosing where physical observability and statistical transferability fail, and developing a spatially validated catchment-based alternative for narrow rivers. This framework assigns satellite, catchment, and in situ observations to the scales at which each is informative, thereby strengthening the credibility of long-term reconstruction, attribution, and management of inland-water change.

Keywords

Environmental drivers, Satellite remote sensing, Machine learning, Long-term reconstruction, Spatial cross-validation, Inland water quality

Language used: English

Date on which the thesis was submitted / produced: 12. 9. 2026

Thesis defence

• Supervisor: doc. Zhongbing Chen, Ph.D.
• Reader: Josef Krása, Bc., externi, Ottavia Zoboli, externi

Czech University of Life Sciences Prague

Faculty of Environmental Sciences

Doctoral programme:
Applied and Landscape Ecology