From Space to Soil: modelling and mapping soil organic carbon dynamics using machine learning and earth observation

Digital soil mapping (DSM) of soil organic carbon (SOC) has advanced rapidly due to growing demand for climate-relevant soil information, expanding SOC measurements, open data, Earth Observation archives, and machine learning. Yet pan-European 3D+T SOC mapping remains constrained by fragmented data, limited harmonized databases, and computational demands. This thesis develops an operational 30 m, wall-to-wall SOC pipeline for Europe (2000–2022), integrating a cloud-optimized Landsat data cube with harmonized SOC observations. Using Random Forest, Quantile Regression Forest, and neural network approaches, it generates multi-depth SOC density time series with pixel-level uncertainty and extrapolation diagnostics. Results provide the first temporally explicit European SOC baseline at high resolution, though pixel-scale change detection remains limited due to low signal-to-noise ratios. Hybrid soil-informed neural networks improve SOC density reconstruction under sparse bulk density data. Overall, the work clarifies methodological constraints, uncertainty communication needs, and practical limits of large-scale SOC monitoring, supporting more transparent and scalable DSM frameworks.