Reconstruction and downscaling of historical land surface boundary conditions with Machine Learning

Understanding the role of land surface physics and biogeochemistry is crucial for improving climate models and weather prediction, particularly in the context of long-term variability, local feedbacks, and extreme events. Accurate boundary conditions—such as land cover (LC) and land use (LU)—are key to enhancing the realism of climate simulations by better representing land-atmosphere interactions that influence surface energy balance, and ecosystem processes. Moreover, they provide the substratum for a realistic representation of the terrestrial carbon cycle components, such as vegetation and soil biogeochemistry.

The CERISE project aims to produce high-resolution (1 km) LC and Leaf Area Index (LAI) datasets covering the period 1925–2020, contributing to novel reanalysis datasets (e.g., ERA6-Land), and seasonal forecasts (e.g., SEAS6). In the first phase, we reconstructed historical LU and LAI by leveraging machine learning (ML) models to downscale coarse-resolution LU datasets (LUH2f, HILDA+). Our workflow integrates multiple ML techniques, such as Random Forest and XGBoost, to train models over high-resolution LC and LAI satellite observations, while actively exploring methods to enhance both performance and interpretability. To capture monthly LAI variations from annual LU inputs, we developed an auxiliary network to model intra-annual variability. Initial results show promising performance in reconstructing LC and LAI across various test years and regions, demonstrating the feasibility and robustness of this ML-based approach for historical reconstructions.

Future phases, including the CONCERTO and TerraDT projects, will extend this work to generate consistent high-resolution LU datasets for the historical (1850-present) and future scenarios (present–2100), supporting CMIP7 climate simulations and scenario-based studies. These efforts will incorporate additional auxiliary data (e.g., elevation, soil types, climate indices) to improve feature representation and develop autoregressive models that account for long-term temporal dependencies and dynamic changes. Ultimately, our goal is to build a robust ML-based emulator for generating scalable, high-resolution land surface boundary conditions to support digital twin applications, thereby enhancing climate simulation and prediction capabilities.