A novel approach for scaling vegetation dynamics based on deep learning was published recently in Methods in Ecology and Evolution. The modeling framework developed by Werner Rammer and Rupert Seidl allows a consistent scaling of local vegetation dynamics (with abundant data and high process understading) to much larger spatial extents (think: country to continental level). At the core, the model harnesses deep learning, which is an exciting new branch of machine learning that revolutionized many fields of computer science in the last years.
Environmental issues such as climate change or biodiversity loss are of global concern and addressing them requires policy responses at national to global level. However, the simulation of vegetation development over large spatial scales is methodologically challenging. Simulation approaches with a detailed coverage of processes are typically only available at local to regional scale, and global simulation approaches (e.g. DGVMs) often lack the details with regard to the demographic structure and biotic interactions which are important drivers of e.g. carbon storage.
The new simulation framework "Scaling vegetation dynamics" (SVD) aims at bridging this scale gap: It is a computationally efficient approach to simulate vegetation transitions using deep neural networks (DNNs), allowing the simulation of vegetation transitions at large spatial scales, and the assessment of their consequences regarding important ecosystem attributes such as C storage or biodiversity.
SVD uses deep neural networks as its core engine. DNNs received a lot of attention recently and they were instrumental for almost every major IT break in the last years - from image recognition to automatic translation, from autonomous driving to beating human masters in games such as Go or Starcraft . Deep learning is a field of machine learning and DNNs learn complex relationships from data. That can be the content of an image (e.g. whether a photo shows a cat or a dog) from raw image pixels, or - in the case of SVD - how vegetation changes over time based on a number of predictor variables. In SVD the DNN is responsible for estimating vegetation transitions conditional on the current vegetation of a cell and its neighborhood, but also on site (e.g. soil data), and climate data. Machine learning algorithms need training data, and SVD can use either empirical (e.g. from inventories) or simulated data (e.g. results from detailed vegetation models).
The new paper introduces the concept of SVD and shows results of simulations that demonstrate the utility of the approach and its ability to efficiently simulate large scale (e.g. 25 Mio ha) forest dynamics.
Comparison of SVD simulations with a detailed process based simulation model (PBM). SVD can faithfully reproduce patterns of landscape dynamics. A) Development of forest composition, structure, and functioning over time, B) forest composition after 100 and 500 years of simulation.