Accurately predicting terrestrial ecosystem responses to climate change is crucial for addressing global challenges. This relies on mechanistic modelling of ecosystem processes through Land Surface Models (LSMs). Despite their importance, LSMs face significant uncertainties due to poorly constrained parameters, especially in carbon cycle predictions. This paper reviews the progress made in using data assimilation (DA) for LSM parameter optimisation, focusing on carbon-water-vegetation interactions, as well as discussing the technical challenges faced by the community. These challenges include identifying sensitive model parameters and their prior distributions, characterising errors due to observation biases and model-data inconsistencies, developing observation operators to interface between the model and the observations, tackling spatial and temporal heterogeneity as well as dealing with large and multiple datasets, and including the spin-up and historical period in the assimilation window. We then outline how machine learning (ML) can help address these issues, proposing different avenues for future work that integrate ML and DA to reduce uncertainties in LSMs. We conclude by highlighting future priorities, including the need for international collaborations, to fully leverage the wealth of available Earth observation datasets, harness machine learning advances, and enhance the predictive capabilities of LSMs.

The atmospheric CO2 growth rate (CGR) is characterised by large interannual variability, mainly due to variations in the land carbon uptake, the most uncertain component in the global carbon budget. We explore the relationships between CGR and global terrestrial water storage (TWS) from the GRACE satellites. A strong negative correlation (r = -0.68, p < 0.01) between these quantities over 2001-2023 indicates that drier years correspond to a higher CGR, suggesting reduced land uptake. We then show regional TWS-CGR correlations and use a new metric to assess their contributions to the global correlation. The tropics account for the entire global TWS-CGR correlation, with small cancelling contributions from the Northern and Southern Hemisphere extratropics. Tropical America explains the dominant contribution (69%) to the global TWS-CGR correlation, despite occupying < 12% of the land surface. Aggregating TWS by MODIS land cover type, tropical forests exhibit the strongest CGR correlations and contribute most to the global TWS-CGR correlation (39%), despite semi-arid and cropland/grassland regions both having more interannual TWS variability. An ensemble mean of four atmospheric CO2 flux inversion products also indicate a 74% tropical contribution to CGR variability, with tropical America and Africa each contributing 30% and 27%. Regarding land cover type, semi-arid/tropical forests contribute almost equally (37%/35%) to CGR variability, although tropical forests cover a smaller surface area (25%/10%). Timeseries of global and regional TWS and CO2 flux inversions through 2001-2023 also show changing regional contributions between global CGR events, which are discussed in relation to regional drought and ENSO events.

Vegetation Gross Primary Productivity (GPP) is the single largest carbon flux of the terrestrial biosphere which, in turn, is responsible for sequestering $25-30\%$ of anthropogenic carbon dioxide emissions. The ability to model GPP is therefore critical for calculating carbon budgets as well as understanding climate feedbacks. Earth System Models (ESMs) have the capability to simulate GPP but vary greatly in their individual estimates, resulting in large uncertainties. We describe a Machine Learning (ML) approach to investigate two key factors responsible for differences in simulated GPP quantities from ESMs: the relative importance of different atmospheric drivers and differences in the representation of land surface processes. We describe the different steps in the development of our interpretable Machine Learning (ML) framework including the choice of algorithms, parameter tuning, training and evaluation. Our results show that ESMs largely agree on the physical climate drivers responsible for GPP as seen in the literature, for instance drought variables in the Mediterranean region or radiation and temperature in the Arctic region. However differences do exist since models don’t necessarily agree on which individual variable is most relevant for GPP. We also explore a distance measure to attribute GPP differences to climate influences versus process differences and provide examples for where our methods work (South Asia, Mediterranean)and where they are inconclusive (Eastern North America).

Forecasts of forest responses to climate variability are governed by climate exposure and ecosystem sensitivity, but ecosystem model projections and process representations are under-constrained by data at multidecadal and longer timescales. Here, we assess ecosystem sensitivity to centennial-scale hydroclimate variability, by comparing dendroclimatic and pollen-inferred reconstructions of drought, forest composition and biomass for the last millennium with five ecosystem model simulations. In both observations and models, spatial patterns in ecosystem responses to hydroclimate variability are strongly governed by ecosystem sensitivity rather than climate exposure. Ecosystem sensitivity was highest in simpler models and higher than observations, suggesting that interactions among biodiversity, demography, and ecophysiology processes dampen the sensitivity of forest composition and biomass to climate variability and change. By integrating ecosystem models with observations from timescales extending beyond the instrumental record, we can better understand and forecast the mechanisms regulating forest sensitivity to climate variability in a complex and changing world.