Climate–ecosystem modelling made easy: The Land Sites Platform

Dynamic Global Vegetation Models (DGVMs) provide a state‐of‐the‐art process‐based approach to study the complex interplay between vegetation and its physical environment. For example, they help to predict how terrestrial plants interact with climate, soils, disturbance and competition for resources. We argue that there is untapped potential for the use of DGVMs in ecological and ecophysiological research. One fundamental barrier to realize this potential is that many researchers with relevant expertize (ecology, plant physiology, soil science, etc.) lack access to the technical resources or awareness of the research potential of DGVMs. Here we present the Land Sites Platform (LSP): new software that facilitates single‐site simulations with the Functionally Assembled Terrestrial Ecosystem Simulator, an advanced DGVM coupled with the Community Land Model. The LSP includes a Graphical User Interface and an Application Programming Interface, which improve the user experience and lower the technical thresholds for installing these model architectures and setting up model experiments. The software is distributed via version‐controlled containers; researchers and students can run simulations directly on their personal computers or servers, with relatively low hardware requirements, and on different operating systems. Version 1.0 of the LSP supports site‐level simulations. We provide input data for 20 established geo‐ecological observation sites in Norway and workflows to add generic sites from public global datasets. The LSP makes standard model experiments with default data easily achievable (e.g., for educational or introductory purposes) while retaining flexibility for more advanced scientific uses. We further provide tools to visualize the model input and output, including simple examples to relate predictions to local observations. The LSP improves access to land surface and DGVM modelling as a building block of community cyberinfrastructure that may inspire new avenues for mechanistic ecosystem research across disciplines.


| INTRODUC TI ON
Vegetation strongly influences energy, water and biogeochemical (e.g., carbon) exchanges between the land surface and the atmosphere (Bonan & Doney, 2018). The terrestrial biosphere and its interplay with the Earth system remain one of the largest sources of uncertainties for ecosystem and climate predictions (IPCC, 2022;Urban et al., 2016), and complex models are necessary to understand the numerous physical, biogeochemical and ecological processes that govern these interactions (Blyth et al., 2021).
Recent developments open up new avenues for integrating biological and Earth system sciences Kyker-Snowman et al., 2022): increasing computational capacity facilitates the use of more complex models on higher spatial resolutions, with more heterogeneity and more sophisticated ecological process representations Pettorelli et al.,  At the same time, increased complexity and realism in these models invariably present new uncertainties, fuelling a continuous need to strengthen the models' empirical foundations and expose process representations to rigorous testing (Collier et al., 2018;. A plethora of model parameters must be derived individually from empirical relationships or experiments. This poses limitations on the processes and locations we can currently model with high confidence (Hartig et al., 2012). Managing the comprehensive ambition and complexity of DGVMs thus requires the mobilization of knowledge from experts across diverse fields . We argue that their use and development by a broader community will benefit both the science of DGVMs and enhance the role of trait-based ecology and process understanding in ecological science.
Widespread utilization and development of DGVMs by specialists in, for example, ecology, hydrology, snow and soil science faces numerous barriers. These models are often linked to and housed within the software architecture of Earth system models, which have some of the most complex scientific code in existence and significant technical requirements. Proprietary code or limited access to computational resources often impose additional hurdles.
New modellers experience steep technical and theoretical learning curves to set up meaningful experiments. At the same time, some trained modellers lack experience with field-and laboratory-based approaches and whole-organism ecology. Arguably, the exchange of technical know-how, knowledge and empirical data could be accelerated if these barriers to communication and hands-on collaboration were alleviated. Enhancing integration between mechanistic ecosystem modelling and statistical-, laboratory-or field-oriented ecology requires community cyberinfrastructure tools to improve accessibility to complex model frameworks (Fer et al., 2020;Lombardozzi et al., n.d.).
To lower the technical barriers to modelling, we present the Land Sites Platform (LSP; Karimi-Asli, Keetz, Lieungh, Yilmaz, et al., 2022). It simplifies simulations with recent model versions of a demographic DGVM embedded in a land surface modelling framework (Section 2.1). The software is open-source and userfriendly: it includes a Graphical User Interface (GUI), an Application Programming Interface (API; Section 2.2) and self-guided analysis tools for input and output (Section 2.3). Owing to their complexity visualize the model input and output, including simple examples to relate predictions to local observations. The LSP improves access to land surface and DGVM modelling as a building block of community cyberinfrastructure that may inspire new avenues for mechanistic ecosystem research across disciplines.

K E Y W O R D S
Application Programming Interface (API), Community Land Model (CLM), Docker container, Dynamic Global Vegetation Model (DGVM), ecological modelling, Functionally Assembled Terrestrial Ecosystem Simulator (FATES), Graphical User Interface (GUI), Land Surface Model (LSM) and driver requirements, many land model frameworks require users to choose and install the correct versions of source code, numerous external libraries and compilers (Fer et al., 2020). Setting this up on any computer is dependent on the operating system and requires substantial knowledge of software engineering. The LSP works on many operating systems and automates the installation of all software dependencies by virtualizing the software environment with Docker containers (Section 2.4; Table 1

| THE L AND S ITE S PL ATFORM
The LSP software architecture relies on GitHub repositories for code management and on Docker containers to run the models and present easy user interfaces ( Figure 1). Our software simplifies the interface between models and users, and between different software components, thereby reducing the number of technical challenges for running a simulation with the model framework.

| Model framework: NorESM, CLM and FATES models
The LSP's demographic DGVM relies on the software infrastructure of an Earth system model, such as a driver and case control system.

Container image Docker
Containers are isolated, virtualized computer environments based on a read-only image file with source code, libraries, dependencies and tools. Docker Inc. is a company that provides container solutions.

Application Programming Interface (API)
An interface for computer programs to communicate with each other efficiently.
Graphical User Interface (GUI) A visual interface between humans and software or hardware, with clickable buttons.
Command line interface Also called a terminal emulator; a computer program that receives commands from a user in the form of lines of text.

Jupyter notebook JupyterLab Jupyter Server
Jupyter notebooks combine computer code and rich text elements. They enable data analysis and viewing and plotting model output. JupyterLab allows users to develop and run notebooks interactively in a browser. Jupyter Server is responsible for storing and organizing data for Jupyter applications.

Git GitHub Repository
Git software enables version control by tracking changes in files. GitHub is an online host of repositories: data structures of files and their version history.

| Accessible interfaces: API and GUI
We created an API as a user-friendly and scalable solution enabling communication between different components of the software (Karimi-Asli, Keetz, & Lieungh, 2022a). An API allows users to interact with a program by sending action commands with optional inputs (requests) to trigger processes. After performing a task, the API re-

| Data processing and analysis
The

| Containerization with Docker
Containerization with Docker (Docker, 2023) permits the LSP to work on current versions of macOS, Linux and Windows. Through this setup, we provide images and containers ( Table 1)

| Integrated field sites
The TA B L E 2 An example API request and success response for creating a 1-year default case for an integrated site (BOR1). Long response strings were truncated and some entries removed for readability. See the NorESM-LSP Development Team (2023) for an overview of the API endpoints.
Existing data from these sites include microclimate, weather, vegetation composition, plant traits and demography (Vandvik et al., 2022).

| Educational example application
We present a simplistic, educational application of the LSP to showcase its capabilities and the model outputs . We The coarse-scale reanalysis climate data used to force the model, while capturing the seasonal variability, exhibit differences to the temperature measurements at the site (Figure 4a). This is not surprising, as the averaging of environmental variables across heterogeneous grid cells rarely represents the conditions at specific locations accurately.
Considering input data uncertainties for model output interpretation is especially important in areas with high topographic variation, such as Norway. To adjust model experiments to local conditions, the forcing could be replaced with data products with higher spatial resolution or with local observations. While observations can also be valuable to adjust (e.g. bias-correct) coarse datasets (e.g. Yang et al., 2011), the default data provide continuous environmental information that can complement incomplete or short-term field observations.
Tracking the modelled growth of plant cohorts through time shows community assembly through trait filtering. The succession of simulated potential natural vegetation with all default PFTs ( Figure 4b) predicts needleleaf evergreen trees as the only remaining PFT after reaching a steady state from initial bare-ground conditions. This is in line with the potential natural vegetation at this site (Bryn et al., 2013), and highlights how locally realized and modelled potential natural vegetation can diverge (Somodi et al., 2021).
Previous DGVM studies over Scandinavia show limited ability to capture realized vegetation patterns Horvath et al., 2021). At the BOR1 site, an important explanation is likely the traditional farming practices, which are not currently included as explicit processes in the model. To give a simplistic example of how changing the model configuration with the LSP allows emulating specific ecological conditions, we narrowed down the PFT list to two grass PFTs most closely F I G U R E 3 The LSP Graphical User Interface (edited screenshot). Users can choose among integrated sites (green markers), add custom sites (purple markers) and are guided through creating and running a case (i.e. simulation or model experiment) by customizing a selection of settings and model parameters. Existing cases are listed along with options to view model settings, run the case, copy and edit the case, or to delete it. The GUI sends requests to the API to operate the model. Background map by Carto. Map lines delineate study areas and do not necessarily depict accepted national boundaries.
resembling the realized vegetation (Figure 4c,d). CLM-FATES could also be configured to, for example, prescribe PFT distributions from satellite observations . Once a suitable model experiment is set up, CLM-FATES yields, for example, productivity measures that are difficult to collect continuously in situ but are important for the ecosystem's impact on the global carbon budget under given climatic conditions (Figure 4c).
The biomass comparison to observed data in this simple example underlines that the global default PFTs are not expected to represent specific regional plant traits, nor to represent this semi-natural ecosystem (Figure 4d). The CLM-FATES currently models potential natural vegetation without domestic grazing, which would, for example, remove biomass from the vegetation. Adding such processes and calibrating model parameters with local observations could improve the model's fit to the site (Lambert et al., 2022). Both process understanding and data are required to disentangle biased predictions like overestimated biomass (Figure 4d). Biomass is an emergent property of productivity and turnover. Establishing which of these is biased high or low is the first task. For instance, new model sensitivity experiments where individual parameters are varied can be evaluated against data such as flux measurements of carbon assimilation and respiration.
While the basic capabilities of the LSP are targeted at new modellers, the software may serve as a shortcut to advanced model experiments, data integration and efficient cross-disciplinary communication that could inspire future model development ( Figure 5).

The advanced DGVM model framework (CLM-FATES) presents
opportunities to test ecological hypotheses and scale up concepts through time and space in future research (Kyker-Snowman et al., 2022). We already know that ecological data and theory can improve DGVM parameters and structure (Nevalainen et al., 2022;Norby et al., 2016;Pastorello et al., 2020;Wullschleger et al., 2011Wullschleger et al., , 2014. Observations, for example in plant trait databases, can inform the direct or inverse estimation of model parameters (Dietze et al., 2014;Hartig et al., 2012;LeBauer et al., 2013). Other modelling approaches using ecological assumptions can refine or validate DGVM parameterization, such as machine learning (Beigaitė et al., 2022) and distribution modelling (Horvath et al., 2021).  et al., 1987). Concepts from community assembly theory may inform DGVMs' process representation (Scheiter et al., 2013). Vice versa, DGVMs may complement ecological hypothesis testing by tracing simulated plant growth from parameters through model processes to predicted states in a controlled environment. The LSP is well suited for interdisciplinary work on these issues because it lowers the technical thresholds and provides supporting materials for analysis and interpretation, thereby helping researchers focus on the scientific questions.
Modellers and field ecologists increasingly adopt open development and FAIR data sharing to facilitate community engagement, accessibility and data flow (e.g., Danabasoglu et al., 2020;Melton et al., 2020;Wilkinson et al., 2016). Tools and protocols conveniently relating observations to model outputs exist in different contexts and configurations (e.g. Fer et al., 2018;Hoffman et al., 2017).
Several software-oriented efforts already aim to ease land surface model and DGVM usability and interpretation Xu et al., 2017) and provide beginner-friendly instruments for education (e.g., LPJ-GUESS Education; Smith et al., 2001). Each solution comes with individual advantages but also limitations, such as missing cross-platform support, proprietary code or restricted possibilities to adapt the incorporated models and versions. The latter is crucial for collaborative efforts aiming at model advancement. The LSP is designed to be a flexible research tool, while simultaneously serving as a modern interface for education. We chose software solutions that are transferable to similar model frameworks, for example, with CESM instead of NorESM, or that could be merged with other initiatives to build on synergies. Further developments with the LSP may expand modelling capabilities with additional input datasets, climatic scenarios, model calibration workflows and alternative modes of the FATES model (see . The LSP development was driven by needs to overcome technical challenges with land surface modelling, including knowledge transfer across institutional and geographical barriers, training resources for new modellers and setting up a shared software environment available to everyone. We combined ideas and experiences from biologists and geoscientists and involved research software engineers to pull the developments together and create the GUI, API and containers. Because the Earth system's complexity inspires increasingly complex scientific code structures, engaging software engineers is crucial to enhance the efficiency and scientific standards of model development (Purgar et al., 2022;Wieters & Fritzsch, 2018). The LSP was tested on many different computers by varied users, including students and professors at an interdisciplinary Master-and PhD-level course in Ecological Climatology ; see also the NorESM-LSP Development Team, 2023, for performance and disc usage tests).
Their feedback confirmed that they felt enabled to run the model and understand the results. They also stressed the need for training in Earth system science and DGVMs to set up robust simulations and interpret the outputs properly. This underlines the need for cross-disciplinary collaboration and training, and for creating community cyberinfrastructure beyond the mainly technical facilitation of the LSP.

F I G U R E 5
The LSP (centre blue box), adds layers of user-friendly software around the vegetation demographic model CLM-FATES (grey trapezoid). Technical facilitation reduces software and hardware barriers to modelling, while user-friendly interfaces simplify the workflow. Many disciplines contribute process understanding to the models (top left). Data (top right) can force the model and provide parameter estimates for model processes and plant functional types. Dotted arrows indicate possible outcomes of the LSP's facilitation: interconnections between cross-disciplinary communication, process understanding, new model experiments, improved understanding of model data needs, existing ecological data and how observations can be compared to model input and output.

| CON CLUS ION
The scientific community needs interdisciplinary, open-access modelling tools and data, and stronger connections between geo-and biosciences to integrate global climate change and biodiversity science-policy agendas Pettorelli et al., 2021).
Our experiences with building the LSP showed that this task requires lowering technical thresholds. Cross-disciplinary education with the LSP may stimulate collaboration and mutual understanding that lead to model development and improved process understanding. The full importance of the many aspects of natural diversity and ecological processes for global systems is still unknown (IPBES, 2019), and will remain unknown until collaborative research is strengthened.
Dynamic, process-based models are increasingly important and sophisticated and can help understand the role of ecology within the Earth system, but require improved accessibility to be more widely adopted (Fer et al., 2020;Kyker-Snowman et al., 2022). (ESM2025) and #821003 (4C). Thanks are due to EOSC-Nordic for support with Galaxy.eu and metadata creation; to the Vestland Climate Grid team for discussions and data insights; numerous students and collaborators for data collection; to the NCAR/NEON team, especially Will Wieder, for helpful discussions and code inspiration; to Rukaya Johaadien and Michal Torma for discussion and code contributions; and to three anonymous referees for valuable feedback.

CO N FLI C T O F I NTER E S T S TATEM ENT
The authors declare no conflicts of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
The LSP code and data are available on GitHub and Zenodo, divided into repositories with the main code (Karimi-Asli, Keetz, Lieungh,