Abstract:
Forests are crucial components of the Earth system, regulating carbon cycles and supporting biodiversity, yet they face significant threats from global change. Understanding the mechanisms and dynamics within forest ecosystems, including the role of disturbances, is therefore paramount. Ecosystem models serve as essential tools for this purpose. High-resolution and process-based models are critical for capturing the complex inter-tree interactions, such as competition, and for accurately simulating the impacts of global change. In this presentation, I will first contextualize the challenges before introducing iLand, an individual-based forest landscape and disturbance model. Subsequently, I will demonstrate how a Deep Neural Network (DNN) based meta-modeling approach can effectively scale this detailed, tree-level perspective to regional and continental extents. Finally, I will present examples that illustrate current research directions and potential avenues for future investigation.
Bio:
Dr. Werner Rammer is a Senior Scientist at the Technical University of Munich. His expertise spans ecosystem modeling, software engineering, and machine learning, with a focus on the development and application of forest ecosystem models, including the models iLand and SVD. His research interests include climate change impacts, ecosystem services, and forest management. He has a strong background in developing and applying computational models for ecological research.
Summary:
Forests are a critical ecosystem: land cover, biodiversity, water holding, carbon sequestration (same amount of carbon as atmosphere)
Climate change is creating a new environment for trees to live and grow up
Powerful disturbance agents
Pest infestation (e.g. bark beetles)
Fires
Storms and wind
(in Europe disturbances increased at 1.5% per year since 1980)
Forest modeling
Started in 1970s
Focus: Landscape scale modeling
Processes
Spatio-temporal interactions
Management decisions
Approach: landscape-scale modeling at the level of individual trees
Climate - vegetation - disturbance
Individual-based: tree-level interaction and competition
Landscape-scale: captures large-scale disturbances
Process-based: account for future climate change signals not observed in past data
iLand model: https://iland-model.org
Hierarchical multi-scale approach
Single tree: location, height, diameter, tree crown shape/size
From carbon perspective can divide into pools: foliage, branches/twigs, stem, roots
Many different physical processes modeled
Modularity: processes can be decomposed into distinct units (same crown repeated)
Repetitiveness
Focus: competition for light process
Model the shade impact of different crown shapes into nearby trees (ray tracing),
Integrating over changing sun position
Compute: light influence pattern over space/time
Can build a library of precompiled light impact patterns
Patterns from multiple trees are superimposed on each other
Focus: demographic processes
Growth
Light
Temperature
Water/nutrients
Predict: annual increment: GPP - respiration/turnover
Carbon distributed into pools: leaves, branches, trunk, roots
Regeneration
Production of seeds
Distribution of seeds on the landscape
Environment filters decide where seeds establish
Mortality
Random tree death (parameterized by tree age)
Stress function of carbon balance
Competition, drought
Stand-level processes (100m x 100m)
Radiation interception
Production of carbohydrates
Environmental modifiers (temperature, water, VPD, nutrients, CO2): 3PG model
Water cycle
Flow of water into/through soil
Evapotranspiration
Disturbances and management
Wind: wind speed at canopy edges, critical wind speeds, damage spread spatially explicit
Bark beetles: spatially explicit dispersal, colonization and tree defense (10m resolution), wind damage can act as trigger for beetles
iLand collaboration
Primary focus: temperate and boreal forests (North America and Europe)
Scaling up using deep neural networks
Facilitating repetitiveness and abstraction
Single tree -> impact pattern -> landscape
Scaling Vegetation Dynamics (SVD): https://github.com/edfm-tum/SVD
Vegetation dynamics: transition between states (probabilistic, take time)
Composition: species abundance
Structure: canopy height
Functioning: density, LAI
Compatible with remote sensing data
Use deep learning to infer probabilities of state transitions
Database of process model runs
Train neural model to predict transitions given environmental drivers
Example: fires in Greater Yellowstone ecosystem
Used neural model to simulate different disturbances: rapid transitions across different states (logging, fire)
Example: combine different simulation types to train DNN approximation that summarizes them all
European forests
Applied empirical disturbances from model trained on historical data, conditional on climate
Fire, Wind
Run at European-continental scale under different climate scenarios
Fire very sensitive to climate scenario
Wind not sensitive
Bark beetle is in the middle
https://www.westernfireforest.org/
Using these models to evaluate fire future