Summary of issues, challenges, and opportunities

​

1. Determine when heterogeneity matters
1. How many tiles per grid cell?
2. At which scales can we revert to sub-grid homogeneity?
2. Connections between tiles
• Intra-cell connections (subsurface/surface/lower ABL)
• Inter-cell connections (connectivity of sub-grid parameterizations across grid cells)
3. New/complementary approaches for sub-grid heterogeneity
• Is tiling the best approach to couple land and atmosphere sub-grid heterogeneity?
• Use of RANS to model heterogeneity over land (perhaps for intra-tile heterogeneity)
4. Software engineering challenges
• Load-balancing issues when considering optimal tiling configurations per cell
• Does increased tiling complexity overwhelm the tiling benefits?
• Existing preprocessing steps/data stumbling blocks to broad adoption of “tile maps”
5. Evaluation
• Need to evaluate simulated sub-grid heterogeneity of surface fluxes and states
• Mapping tiles out in space enables more objective and robust evaluation (and use)
• Satellite remote sensing (LST) and/or intensive field campaigns (e.g., LIAISE)

Breakout 1: New approaches for subgrid heterogeneity

Potential collaborative next steps

​

10M idea - “Enabling stakeholder-scale land model predictions”

�Motivation

• Broad interest to develop modeling center capabilities to map �tiling simulations out in space (“recast tiling as a clustering exercise”)
• Direct approach to more objectively evaluate simulated tiles
• Enables exploration of tiling connections (inter/intra cells)
• Straightforward and self-consistent approach (not downscaling!)�to bring tiles down to stakeholder-relevant spatial scales

�Approach

• Generalized workflow (data and software package) to assemble tile structures (and associated ~30 m maps); tiling assembly workflow would be implemented at each participating modeling center
• Sufficiently flexible for centers to implement/adapt their existing tiling structures/hierarchies
• CMIP land modeling results would enable the capability to produce field-scale estimates
• Build APIs for stakeholders and provide country/center-level support for data accessibility
• Framework enables each center to then implement down the lince concepts around inter-tile connections, optimal tiling structures, among others.

Breakout 1: New approaches for subgrid heterogeneity

​

Summary of issues, challenges, and opportunities

​

• Challenge: how to measure complexity, and how to judge appropriate levels of complexity
• Challenge: how much of a concern is order of operation in calibration
• Challenge: how to prevent biases in forcing datasets (e,.g. LAI) from aliasing into biases in parameters?�
• Opportunity: Identification of shared logic in models and where the more general rungs are in complexity hierarchies and testing of models in these configurations (e.g. prescribed LAI already part of PLUMBER, what others are general?)
• Sharable workflows and standards
• Specific use cases, e.g. soil biogeochemistry�
• Opportunity: reduced complexity configurations of land models for educational and training purposes.
• Requires documentation, which is important anyway

​

​

Breakout 2: Managing model complexity

Potential collaborative next steps (Do not need to include something for all funnding levels)

​

Please consider breaking down options according to cost

(e.g., no cost, inexpensive (~$5000), medium cost (~$100,000), Super League (~$5m)

​

• Option 1
• Option 2
• …

Breakout 2: Managing model complexity

​

​

• All agree on the large potential benefit of defining common modules that can be shared !
• Which level (individual ‘small’ processes – large W/E/C components ) still unclear ?
• Drawing boundary between tightly connected processes is LSM dependent
• Defining common “LSM structure” / “module interfaces” / “level of complexity” are challenging !
• Associate benchmark / Evaluation data in the “modularity vision”
• Two direction of work have been identified / discussed:
1. Long term / large ambition approach with large funding (10 M€)
• Define an “Ideal Target” for LSM complexity / processes (⇒ This morning meeting) �⇒ Then have a community concept note/paper ⇒ A common voice to help raise funding �⇒ Define a flexible LSM framework across groups (Using concept of “Object”
2. Short term effort with very low / no funding

Breakout 3: Towards sharing of modules across LSMs

Summary of issues / challenges / opportunities

Potential collaborative next steps (Do not need to include something for all funnding levels)

No or low cost potential actions

• Start defining a common module (and the associated interfaces) that could be useful for a few modeling groups !
• Build from that first experience to design / define the bigger framework (for the whole LSM)
• Potential Candidates :
• CROP module (for Phenology / C allocation / Sowing & Harves dates / ….
• Natural vegetation leaf phenology (smaller ambition)
• Soil Organic Matter Dynamic
• Hydrology: Routing model including water / heat / components (DIC, DOC, ….)
• Radiative transfer scheme
• ....
• Common Module could be based on existing models or merging different models ?�

High cost potential actions

• Defining Groups to define collectively the scientific needs : �Demographie (Rosie), Biogeochemistry (Soenke), Crops (Sam; link with ISIMIP); Hydrology (Simon); Energy & Atm. coupling (?) ; Urban ? ; “Overall structure / design of Modularity” (?);
• Regular teleconf with an objective for a 6/12 month target : A common vision of “optimal / desirable model features” �(link model feature with key science questions or observational evidences to simulate)
• Need to gather at least 1-2 persons per LSM groups ?

Breakout 3: Towards sharing of modules across LSMs

​

Summary of issues, challenges, and opportunities

​

• Ancillary data, parameters and processing tools

- We need a common shared land-sea mask. Ideally one that overlaps with hydrological community (ECMWF

may have a ‘best’ one)?

- we need a plethora of parameters for each tile, but input datasets do not always overlap spatially and in

resolution.

- Centres used different approaches to generate these spatial inputs (e.g. nearest neighbour). Let’s share

our practices, e.g. issues with downscaling/upscaling and how to generate effective parameters

- Reproducibility of datasets, non-standard grids, clusters v. spatially explicit grids

- share ancillary data generation procedures (but often thousands of lines of code..)

- Models use different PFTs; Behaviour clustering v PFTs

- different land cover and LUC datasets are used

- ....

​

Breakout 4: Input and forcing datasets

​

Summary of issues, challenges, and opportunities

​

​

• Forcing datasets

- Different centres & ‘Projects’ (Global Carbon Project) use different driving data products:

Full 20th century until present day for TRENDY/Global carbon project; Global present day data, higher quality

1980 till present (ERA5?); Regional bespoke data; Flux tower dataset generated by the PLUMBER team (solved

problem); long-term bias-corrected (ISIMIP) projection datasets)

- Are some datasets becoming obsolete (e.g. GSWP3)? Diversity is perceived as useful and we do not recommend

all using the same dataset, but current (LMIP) diversity is opportunistic rather than scientific (could we develop an

ensemble of forcing data that captures uncertainty (ERA?)

- Quality/fidelity deteriorates moving from Global v regional v. national products: werecommend sharing

the national/regional products among groups

- Frequency of forcing should ideally be increased (hourly rather than 3 or than 6 hourly) to solve predictions during transition hours, for example

- Quality of forcing – more info, e.g., direct/diffuse radiation, precipitation intensity/area?

​

- What is the best product for S2S flood and stream-flow forecasts. S2S at global scale very coarse resolution. Although this is ultimately a predictability problem.

​

​

​

• …

Breakout 4: Input and forcing datasets

​

Potential collaborative next steps

​

Inexpensive (~$100K per year?)

​

• Working group and repository for input data (under GLASS umbrella?)
• Work on proper ancillary input data description
• Tools to process the data.
• Start with the highest resolutions and/or best datasets
• Create and maintain a dedicated repository, especially including a land input data guide, original and processed versions, version control
• File-types? Clear attributes in files and link to detailed scientific documents explaining the datasets
• How to get the centres to buy in?
• How to best use GoogleEarth Engine? Issues of data sharing/privacy
• ~$100K+/yr to support maintenance of repository

Breakout 4: Input and forcing datasets

​

​

Summary of issues, challenges, and opportunities

​

Topics of strong interest

• Capturing extreme event impacts on yields
• Agricultural management for C sequestration

​

Challenges

• Need for calibration/evaluation data
• Just simulating grasses for crops won’t get you the right biogeochemical/physical dynamics.
• So many management techniques to add
• Not much apparent interest in pasture…

​

Opportunities

• Existing intercomparison efforts can be leveraged.
• Don’t necessarily need to care about yield to answer questions relevant to atmosphere-coupled runs.
• Crop modules typically pretty self-contained; potentially good for modularization.

Breakout 5: Crop modelling and Forestry practice� ^v and pasture!

Potential collaborative next steps (order of magnitude)

• Free: Shared document describing/linking existing datasets that can be used for calibration/evaluation
• $10-50k: Agriculture-specific meeting to coordinate next steps (USDA?)
• $10-100k: Scenario generation to evaluate different management strategy impacts
• $100k-$1M: Scaled-up, coordinated point experiments (with observations) of yield, management, and soil C (AgMIP?)
• $++: Add crop & pasture sites to benchmarking efforts to catalyze process/property evaluation
• $1-10M: Multiple LU projections to achieve any given climate target, to explore uncertainty
• $10M+: Modularize a few models to enable plug-and-play of crop physiology, phenology, and management

Potential collaborative next steps:

​

• Free: Shared document describing/linking existing datasets that can be used for calibration/evaluation
• $10-50k: Datasets: Need to create higher level datasets past country level tiling.
• How to utilize remote sensed data biomass data? Improve land cover mapping to PFT mapping
• $10-100k: LUH3: coordination and discussion on how to improve datasets, make better match of what we need.
• $100k-1M: LUH3: Develop new LUH dataset, transitions, “sub”scenarios, etc.
• $10M+: IAM and ESM coupling

Breakout 5: Crop modelling and Forestry practice

Summary of issues, challenges, and opportunities

​

Topics of strong interest

• Forest management for C sequestration
• Climate impacts on growth of forests; important for permanence of CDR, forest harvest potential, water demand.
• Demography for secondary forest sink, reforestation
• Linking IAMs to ESMs more

​

Challenges

• What land mgmt practices are important, what inputs would be needed?
• Realistic forest harvests (need to coordinate with IAMs, national level harvest translation to gridcell)
• Forestry species - need to include which species are being planted for afforestation, plantations, fast growth

​

Opportunities

• Agroforestry (but should this be implemented by prescribed or process-based way?)
• Plantations in demography models

Summary of issues, challenges, and opportunities

​

Opportunities: linking agriculture to N&P biogeochem, resolving groundwater dynamics, shifts due to snow and glacier melt, and effects of conservation efforts on the water cycle

​

“Our moonshot”: Developing a human-water coupled system → $10M+ ($100M+?) project

​

For this, we need for a “HydroEconomic” approach at the Intersection of climate science, sociology and econometrics to develop a human-aware prognostic model

​

Challenges to this objective: Widely unknown (state and evolution of ) reservoir operations, water use for agriculture, industry, and other human uses. (limitations in both available data and unknown processes).

​

Constrained by ongoing efforts: E.g., 1) Groundwater (GRACE), channel routing (SWOT)

​

Potential of local and crowd-sourced data, cell phone data, … But, how can we upscale them?

​

Breakout 6: Water & Land management

​

• Short term challenges:

​

Increasing model resolution makes it more challenging to model water and land management

​

• (solution to) Long term challenges:

​

Crowdsourcing for water management data collection (e.g., participation of municipalities)

Contribute policymaking

Business, farming, water use …

Groundwater sciences

Water temperature and quality

Potential collaborative next steps (Do not need to include something for all funding levels)

​

Please consider breaking down options according to cost

(e.g., no cost, inexpensive (~$5000), medium cost (~$100,000), Super League (~$5m)

​

• Low cost: Unify existing data sources on human water use and reservoir operation

​

• Medium cost: Advocate for exapansion existing groundwater observational networks

​

• High cost: Advocate for costellations (Cube sat, Nano sat, ..) for water-surface monitoring.

​

• VH cost: Advocate for longer satellite mission to constrain human-water dynamics (“SWOT2”) following the success of GRACE.

​

​

Breakout 6: Water & Land management

​

​

​

Summary of issues, challenges, and opportunities

​

• When do you couple models? This decision needs to consider benefits but ALSO costs (incl. opportunity costs)
• Criteria - clearly show model(s) are fit for purpose before coupling
• Benefits - clearly define (predictive skill and/or capability)
• Costs
• Get higher with increasing disparity in models’ timesteps, approaches, resolution
• Increased complexity and performance loss
• Data degradation, e.g. with PFT mismatches
• Have to maintain models and coupling code in the future, indefinitely
• Too often these kinds of things happen because “it’s a challenge”
• Opportunity costs!
• Extremes and LUC as examples of why one might need coupling
• Extremes: daily (at least) timescale needed; drive societal responses; can IAMs take advantage of this information? Unclear
• IAMs can’t (?) yet take advantage of distribution changes/extremes even when coupled (but we really needed an IA modeler in the room)

​

Breakout 7: Coupling external models to LSMs

​

Potential collaborative next steps (most of these are inexpensive to medium cost)

​

• Standardization of land model couplers where possible
• Be aware, and take advantage, of specialist community (e.g. 2023 Toulouse workshop)
• Much work has been done: YAC, LandSymm, etc. Learn from it
• Integrated reduced order models (and/or imaginative experiments, e.g. adaptive emissions scenarios) as tests before proceeding with an expensive coupling
• Understand order of magnitude of likely effect/benefit ; will coupling signal emerge from the noise?
• This also might allow isolation of behavior(s) you care about
• And/or regional coupling — focus on areas that you care about and/or benefit from the coupling
• Preliminary but systematic reviews: needs, capabilities, priorities
• (Both within and across communities? See extremes example above.)

Breakout 7: Coupling external models to LSMs

​

​

“Issues”

• Are models fit for purpose (see right)
• Legacy problems - models were constructed before we had data.
• Are we considering the correct variables
• Observational uncertainty - biases across datasets (i.e, all burnt area products seem to underestimate fire).

​

Breakout 8: Fire & Humans

​

Challenges

• Fire regimes are becoming more unusual
• How to we use data to:
• construct newer models?
• Understand old models
• Fire suppression isn't accounted for really well in models, or captured well in observations
• How to use models to inform observation collection.

​

​

Opportunities

• New methods to represent uncertainty in processes we cant yet describe (lighting temperature etc)
• Clever stats: Emulations, Bayesian Inference.
• Keen and engaged community through e,g fireMIP

Hantson et al. 2020

Potential collaborative next steps (Do not need to include something for all funnding levels)

​

Please consider breaking down options according to cost

(e.g., no cost, inexpensive (~$5000), medium cost (~$100,000), Super League (~$5m)

​

• Use emulators to understand model sensitivity to drivers vs observations using fireMIP runs already completed.
• New perturbed parameter runs from fireMIP models to assess parameter sensitivity against inferred from observations
• Find ways to prescribe aspects of fire modelling, particularly prescribing fire to vegetation model and prescribe vegetation to fire model, to help understand models better and their differences

​

• Use data, modelling and stats techniques to use recent but short high res fire information to reconstruct longer term records from e.f AVHRR with proper, likelihood-based uncertainty estimates
• Collect high resolution info on all potential variables (e.g. US fire and driver records) and use uncertainty techniques to construct fire probability models in areas of high res and areas of datagaps
• Combien data sources together with models, to assess where to priorities new data collection

Breakout 8: Fire & Humans

​

​

Summary of issues, challenges, and opportunities

​

• Existing relevant data that is not yet coordinated and quality controlled for benchmarking (e.g. site-level data, diurnal cycles, crops, woody demography)
• Lack of process-relevant evaluations
• Coordination of existing benchmarking infrastructure.
• Make it a community effort.
• Better communication with data community and providing incentives

Breakout 9: Land model benchmarking

​

Potential collaborative next steps (Do not need to include something for all funnding levels)

​

​

• Expand every second regular ILAMB meeting to cover all aspects of international benchmarking infrastructure (Email Forrest Hoffman/ILAMB Email list if you like to join), including point-based evaluation
• Survey the community about benchmarking and evaluation data needs across spatial and temporal scales. ILAMB meeting next year likely focus for this (similar to 2016)
• Some kind of coordinated effort to reach out to relevant data collecting communities to involve them in propagation of their data into our community and its utilisation.

Breakout 9: Land model benchmarking

​

​

​

Summary of issues, challenges, and opportunities

​

• Point 1
• Point 2
• …

Breakout 10: Machine learning approaches & LSMs

​

MACHINE LEARNING DISCUSSION

THANK YOU TO THE GROUP!

ML for

... bias correction / downscaling

... improving parameterizations

... generating hypothesis

... processing understanding

... addressing geographical extrapolation

outcome

1. Workgroup on ML->LSM
2. Challenge ML with LSMs (emulation/PPE/attribution/evaluation)
3. “get LSMs to do what we think they should be doing” (improve model parameterizations)
4. Prediction of local scale impacts including extreme events

Potential collaborative next steps (Do not need to include something for all funnding levels)

​

Please consider breaking down options according to cost

(e.g., no cost, inexpensive (~$5000), medium cost (~$100,000), Super League (~$5m)

​

• Option 1
• Option 2
• …

Breakout 10: Machine learning approaches & LSMs

​

​

​

​

Summary of issues / challenges

​

• defining parameter priors
• effectively leveraging observations
• influence of calibration methodology
• differing model architecture

​

Opportunities

• leverage existing work in applied math, CS fields
• elucidate parameter effects
• share workflows, diagnostics, cost functions, best practices

Breakout 11: Parameter estimation & uncertainty

Potential collaborative next steps (Do not need to include something for all funnding levels)

​

Please consider breaking down options according to cost

(e.g., no cost, inexpensive (~$5000), medium cost (~$100,000), Super League (~$5m)

​

• (Find the time to) read other people’s papers!! ($0k)
• Share workflow for a) ensemble generation and SA; b) parameter estimation methods/code ($0k)
• Exchange on defining cost functions → interface with benchmarking community ($0k)
• Ecological working group to define how a PFT should be defined (or if needed if we move more towards VDMs)

​

• (Int’l network of) undergraduate research internships/assistantships ($10-20k) to:
• Collate parameter/trait information
• Cross-check parameters across modeling groups
• Help develop parameter sensitivity database
• Workshops/activities to help w/ above ($0-20k) → Land DA WG and Community can help facilitate

​

• Collaborate on test all configurations of model and DA method ($100k to $5m)
• Cross-model shared parameter estimation system system ($1-5m)
• ​

Breakout 11: Parameter estimation & uncertainty

​

​

​

​