Identifying Dynamic Memory Effects on Vegetation State Using Recurrent Neural Networks.

Area of Science:

• Earth System Science
• Climate Science
• Ecology

Background:

• Vegetation state is influenced by climate, involving complex, non-linear interactions across timescales.
• Temporally lagged dependencies, or memory effects, are increasingly recognized but challenging for traditional data-driven models.
• Existing models often fail to explicitly represent time and complex sequential processes in vegetation dynamics.

Purpose of the Study:

• To develop and apply a novel approach using Recurrent Neural Networks (RNNs) for modeling vegetation dynamics.
• To investigate the role and scale of memory effects in global vegetation state using a Long Short-Term Memory (LSTM) model.
• To predict the Normalized Difference Vegetation Index (NDVI) by integrating climate, soil, and land cover data.

Main Methods:

• Utilized a Long Short-Term Memory (LSTM) network, a type of RNN, to model global NDVI.
• Employed climate time-series and static soil/land cover variables as predictors.
• Conducted permutation experiments and spatio-temporal cross-validation to assess memory effects and model generalizability.

Main Results:

• The LSTM model achieved a high global prediction accuracy for NDVI (R² = 0.943, RMSE = 0.056).
• The temporal model explained 14% more variance globally compared to a non-memory model.
• Memory effects showed the most significant impact in arid, semi-arid, sub-tropical, and transitional water-driven biomes, with improvements up to 25%.

Conclusions:

• Memory effects are crucial for accurately modeling global vegetation dynamics.
• LSTM networks provide a powerful tool for capturing complex temporal dependencies in Earth system science.
• Understanding memory effects is vital for predicting vegetation responses to climate change, particularly in vulnerable arid and semi-arid regions.