Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Global Vegetation Modeling with Pre-Trained Weather Transformers

Pascal Janetzky, Florian Gallusser, Simon Hentschel, Andreas Hotho, Anna Krause

TL;DR

The paper addresses predicting vegetation activity (NDVI) from high-resolution meteorological data by transferring knowledge from a pre-trained weather Transformer, FourCastNet, to a global NDVI task at $0.25^\circ$ resolution. It adapts FCN by replacing the weather head with a dense, tanh-activated layer and employing an Adaptive Fourier Neural Operator, comparing finetuning against training from scratch and evaluating against CNN, LSTM, and local state-space baselines. Finetuning the pre-trained atmospheric representation yields substantial gains (global $R^2$ ≈ $0.633$, RMSE ≈ $0.040$) over scratch and CNN baselines, though a long-memory LSTM achieves higher fidelity at coarser resolution ($R^2$ ≈ $0.904$, RMSE ≈ $0.017$). The results demonstrate the viability of transferring weather-model representations to vegetation modeling, quantify data and training requirements through ablations, and point to future work incorporating additional drivers and explainability to enhance ecological insight.

Abstract

Accurate vegetation models can produce further insights into the complex interaction between vegetation activity and ecosystem processes. Previous research has established that long-term trends and short-term variability of temperature and precipitation affect vegetation activity. Motivated by the recent success of Transformer-based Deep Learning models for medium-range weather forecasting, we adapt the publicly available pre-trained FourCastNet to model vegetation activity while accounting for the short-term dynamics of climate variability. We investigate how the learned global representation of the atmosphere's state can be transferred to model the normalized difference vegetation index (NDVI). Our model globally estimates vegetation activity at a resolution of \SI{0.25}{\degree} while relying only on meteorological data. We demonstrate that leveraging pre-trained weather models improves the NDVI estimates compared to learning an NDVI model from scratch. Additionally, we compare our results to other recent data-driven NDVI modeling approaches from machine learning and ecology literature. We further provide experimental evidence on how much data and training time is necessary to turn FourCastNet into an effective vegetation model. Code and models will be made available upon publication.

Global Vegetation Modeling with Pre-Trained Weather Transformers

TL;DR

The paper addresses predicting vegetation activity (NDVI) from high-resolution meteorological data by transferring knowledge from a pre-trained weather Transformer, FourCastNet, to a global NDVI task at resolution. It adapts FCN by replacing the weather head with a dense, tanh-activated layer and employing an Adaptive Fourier Neural Operator, comparing finetuning against training from scratch and evaluating against CNN, LSTM, and local state-space baselines. Finetuning the pre-trained atmospheric representation yields substantial gains (global ≈ , RMSE ≈ ) over scratch and CNN baselines, though a long-memory LSTM achieves higher fidelity at coarser resolution ( ≈ , RMSE ≈ ). The results demonstrate the viability of transferring weather-model representations to vegetation modeling, quantify data and training requirements through ablations, and point to future work incorporating additional drivers and explainability to enhance ecological insight.

Abstract

Accurate vegetation models can produce further insights into the complex interaction between vegetation activity and ecosystem processes. Previous research has established that long-term trends and short-term variability of temperature and precipitation affect vegetation activity. Motivated by the recent success of Transformer-based Deep Learning models for medium-range weather forecasting, we adapt the publicly available pre-trained FourCastNet to model vegetation activity while accounting for the short-term dynamics of climate variability. We investigate how the learned global representation of the atmosphere's state can be transferred to model the normalized difference vegetation index (NDVI). Our model globally estimates vegetation activity at a resolution of \SI{0.25}{\degree} while relying only on meteorological data. We demonstrate that leveraging pre-trained weather models improves the NDVI estimates compared to learning an NDVI model from scratch. Additionally, we compare our results to other recent data-driven NDVI modeling approaches from machine learning and ecology literature. We further provide experimental evidence on how much data and training time is necessary to turn FourCastNet into an effective vegetation model. Code and models will be made available upon publication.
Paper Structure (16 sections, 4 equations, 5 figures, 3 tables)

This paper contains 16 sections, 4 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Pre-trained weather models for vegetation modeling
  3. Dataset
  4. Method
  5. Comparison models
  6. Experimental setup
  7. Evaluation
  8. Results and discussion
  9. Conclusion
  10. Extended results and supplementary figures
  11. FourCastNet training details
  12. Baseline and comparison models details
  13. Evaluation setting
  14. Global evaluation
  15. Local evaluation
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Results for ablation studies I & II. Left: varying number of finetuning epochs. Right: Varying amount of training data. Results reported in \ref{['tab:results']} are highlighted in both plots.
  • Figure 2: Results for ablation study III: varying number of frozen Transformer blocks during finetuning. Runtimes are averaged over five epochs.
  • Figure 3: Overview of the used architecture based on the FourCastNet model pathak2022fourcastnet. We initialize the model from pre-trained weights fcn_weights and replace the weather-specific head with a linear head for modelling the normalized difference vegetation index (NDVI).
  • Figure 4: Global visualization of the $\text{R}^2$ score on the entire test set. $\text{R}^2$ scores below 0.0 are clipped to 0 for ease of visualization. Performance is the strongest for continental Europe and North America and decreases towards higher latitude regions.
  • Figure 5: Distribution of the normalized difference vegetation index (NDVI) data from 1982 to 2023. NDVI data was only used until 2013 due to the noticeable data shift afterwards.