Introduction

Achieving carbon neutrality by 2050 depends on our ability to accurately quantify, manage, and protect carbon stocks across the world’s terrestrial ecosystems [1]. Terrestrial vegetation regulates global climate by sequestering and storing atmospheric carbon. However, increasing urbanization and agricultural expansion encroach upon natural ecosystems, while forests face climate impacts like droughts and wildfires.

Models are crucial for mapping carbon storage and guide conservation and restoration efforts by simulating carbon cycle regulation, helping assess the potential of nature based solutions. As we experience an escalating climate crisis, we must understand three key factors: i) how much carbon is stored, ii) where it is stored, and iii) how storage capacity may change over time (as well as other services that ecosystems offer to society). The complexity of this task has led researchers to develop models with varying degrees of sophistication, each serving different purposes in the broader effort to manage carbon stocks. Estimating carbon stocks in a transparent, accurate manner that is consistent over time is critical to estimating the effects of past, present and future changes in emissions and removals of CO₂ from the atmosphere.

However, incoherent modelling approaches can create confusion for decision makers. For example, current global land use emissions estimates differ by approximately 7 billion tonnes of CO₂ per year (20% of global fossil CO₂ emissions) between global models used in the Intergovernmental Panel on Climate Change (IPCC) Assessment Reports and national inventories under the United Nations Framework Convention on Climate Change (UNFCCC) [2].

More integrated science can better help drive informed policies and action. Two recent studies developed at the Basque Centre of Climate Change (BC3) [3,4], which sit on different ends of the model complexity spectrum, offer insights on the value of integrating multi-tiered carbon storage modelling approaches.

Aligning carbon models with modelling purpose

Carbon storage models address three broad issues: (1) historical time-series reconstruction, (2) nowcasting and (3) forecasting of potential estimations. Like in most modelling problems, there is no universally accepted model solution. This results in a variety of approaches suited to specific contexts of application.

Recent global-scale studies focusing on mapping carbon storage in the first two decades of the 21st century exemplify divergent approaches to carbon storage estimation. For example, [5] used machine learning to derive Above Ground Biomass (AGB) estimates from satellite data and environmental features. In contrast, [3] employed the IPCC-endorsed tier one lookup tables based on environmental and historical factors. The machine learning approach offers superior resolution, but the IPCC methodology offers biological interpretability, greater transparency for practitioners, and direct compatibility with national carbon accounting frameworks.

The emergence of real-time monitoring capabilities (i.e. nowcasting) marks a significant advancement in carbon storage modelling. Real-time monitoring through LiDAR enables global scale AGB measurement [6], but requires extrapolation based on additional earth observation data due to incomplete coverage. Traditional approaches inferring AGB directly from multispectral data [7] have faced saturation limitations, particularly in tropical regions [8]. Newer deep learning models, such as [9] use a two-stage structure, first inferring canopy height and then applying allometric calculations to calculate AGB. These models achieve high accuracy but remain constrained by their regional scope and temporal specificity.

A third category of models focuses on estimating carbon storage potential to guide conservation and restoration efforts. These models [10–12] rely on environmental and social factors, training on data from minimally disturbed regions to avoid anthropogenic bias. Recent work by [3] showed that a model using four climatic variables could match models using hundreds of predictors, such as [12], suggesting that simpler, more interpretable models may be equally effective for potential storage estimation.

While modelling has seen a significant leap forward, challenges remain in harmonizing data across scales and integrating these models into actionable frameworks. There is a trade-off between model complicatedness (as opposed to complexity - see [13]) and interpretability. This is particularly important for practitioners, such as those defining and monitoring Nationally Determined Contributions (NDCs), who often evaluate models based on criteria like understandability, credibility, replicability, and ease of independent use.

From models to actionable insights

Broadly usable and integrated carbon storage models are vital for both scientific understanding and effective climate policy. Global carbon markets, reforestation projects, and national commitments under the Paris Agreement (i.e. NDCs and the new 2030 targets) all depend on reliable carbon estimates. Current fragmented approaches, primarily driven by top-down models accessible only to experts, can lead to inconsistencies which can in turn result in misguided policy [2,14].

We argue it is time to shift focus from the mere development of ever-accurate carbon storage models at any cost and focus on making them understandable, usable, accessible by all, integrated, and impactful for policy-making and real-world application. Achieving this requires significant progress in data and model integration, as well as enhanced interoperability [15].

The recent proliferation of carbon storage models is in stark contrast with their integration. Model results published in the literature and data repositories lack seamless actionability: they often require considerable efforts to integrate into new studies, leading to duplication of efforts by independent research teams. This challenge is even more pronounced when it comes to the reusability of models, reflecting a broader crisis in scientific reproducibility. For example, limited access to pre-trained models and incomplete documentation or lack of transparency often force researchers to reimplement entire workflows, increasing the risk of inconsistencies, errors, and further reducing replicability. The semantically-driven approaches developed by BC3 [3] move in the opposite direction, making pre-trained models of aboveground biomass, based on fully open-access and transparently integrated climate data, usable by researchers and practitioners without a need for intermediaries.

From fragmentation to integration by bridging scales

Carbon storage modelling is essential for developing climate policies, but its effectiveness is complicated by the varying scales at which these policies are implemented. Global assessments often use coarse-resolution data unsuitable for local decision-making, while detailed regional models lack broader applicability.

Artificial intelligence and machine reasoning offer promising pathways for the integration of different-tier models and automated selection of optimal approaches for specific regions. However, data harmonization remains a significant challenge. Inconsistent definitions of forest types, extent, and condition across datasets mean that inputs validated for one model are often incompatible with others. This fragmentation results in redundant data collection and creates barriers to model interoperability. The development of standardized definitions for forests, shrubland, grassland and other vegetation types and the corresponding land uses, as in [3], represents significant progress toward seamless multi-scale integration through semantic technologies.

Most carbon storage models face a crucial limitation related to their temporally static nature. They require periodic updates or separate implementations for time-series analysis. This constraint becomes particularly evident in models based on earth observation data, where coherence between years is often problematic [9]. We need time-agnostic models that can both process new data in real-time and adapt to different scenarios. Such models would bridge the gap between static historical analyses and future projections, providing a more comprehensive framework for carbon storage assessment.

Conclusion

Carbon storage modelling must become more actionable, transparent, integrated, and accessible to diverse stakeholders. The current state of fragmentation limits the ability to produce coherent, adaptable to context and multi-scale insights essential for effective policy and climate action. Addressing these challenges requires progress in interoperability, harmonization, and the development of models that balance sophistication with usability and interpretability.

Integration-oriented approaches that bridge global, regional, and local scales align with nested policy frameworks and support multi-level governance. For example, the development of standardized forest ontologies as in [3] and fully open-access models and data inputs, such as those used by [4], exemplify how transparent, user-friendly models can bridge gaps between scientific research and practical applications.

Technical advancements in carbon modelling must align with real-world applications. The value of transparent and robust carbon storage models lies in their ability to support evidence-based decision-making, facilitate cross-scale collaboration, and empower stakeholders to design realistic climate mitigation pathways based on biological carbon sequestration.

Carbon storage models must transcend academic boundaries to become accessible tools that guide policy decisions, and monitor their effectiveness. A coordinated push toward integrated, open, and scalable solutions can unlock their potential to guide effective conservation, restoration, and climate mitigation and adaptation efforts, ensuring these critical tools contribute to a sustainable and resilient future.