Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Recurrent Neural Networks for Modelling Gross Primary Production

David Montero, Miguel D. Mahecha, Francesco Martinuzzi, César Aybar, Anne Klosterhalfen, Alexander Knohl, Franziska Koebsch, Jesús Anaya, Sebastian Wieneke

TL;DR

This study presents a comparative analysis of three architectures: Recurrent Neural Networks, Gated Recurrent Units (GRUs), and Long-Short Term Memory (LSTMs) to reveal comparable performance across all models for full-year and growing season predictions.

Abstract

Accurate quantification of Gross Primary Production (GPP) is crucial for understanding terrestrial carbon dynamics. It represents the largest atmosphere-to-land CO$_2$ flux, especially significant for forests. Eddy Covariance (EC) measurements are widely used for ecosystem-scale GPP quantification but are globally sparse. In areas lacking local EC measurements, remote sensing (RS) data are typically utilised to estimate GPP after statistically relating them to in-situ data. Deep learning offers novel perspectives, and the potential of recurrent neural network architectures for estimating daily GPP remains underexplored. This study presents a comparative analysis of three architectures: Recurrent Neural Networks (RNNs), Gated Recurrent Units (GRUs), and Long-Short Term Memory (LSTMs). Our findings reveal comparable performance across all models for full-year and growing season predictions. Notably, LSTMs outperform in predicting climate-induced GPP extremes. Furthermore, our analysis highlights the importance of incorporating radiation and RS inputs (optical, temperature, and radar) for accurate GPP predictions, particularly during climate extremes.

Recurrent Neural Networks for Modelling Gross Primary Production

TL;DR

This study presents a comparative analysis of three architectures: Recurrent Neural Networks, Gated Recurrent Units (GRUs), and Long-Short Term Memory (LSTMs) to reveal comparable performance across all models for full-year and growing season predictions.

Abstract

Accurate quantification of Gross Primary Production (GPP) is crucial for understanding terrestrial carbon dynamics. It represents the largest atmosphere-to-land CO flux, especially significant for forests. Eddy Covariance (EC) measurements are widely used for ecosystem-scale GPP quantification but are globally sparse. In areas lacking local EC measurements, remote sensing (RS) data are typically utilised to estimate GPP after statistically relating them to in-situ data. Deep learning offers novel perspectives, and the potential of recurrent neural network architectures for estimating daily GPP remains underexplored. This study presents a comparative analysis of three architectures: Recurrent Neural Networks (RNNs), Gated Recurrent Units (GRUs), and Long-Short Term Memory (LSTMs). Our findings reveal comparable performance across all models for full-year and growing season predictions. Notably, LSTMs outperform in predicting climate-induced GPP extremes. Furthermore, our analysis highlights the importance of incorporating radiation and RS inputs (optical, temperature, and radar) for accurate GPP predictions, particularly during climate extremes.
Paper Structure (9 sections, 3 figures)

This paper contains 9 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Data preparation
  4. Modelling and evaluation approach
  5. XAI: Feature Importance
  6. Results and Discussions
  7. Comparison of models
  8. Feature Importance
  9. Conclusions

Figures (3)

  • Figure 1: Comparison of models performance. Median NRMSE per model for a) full period, b) growing season, and c) climate-induced GPP extremes. Black boxes represent the range from Q$_1$ to Q$_3$ while error bars denote the range between the 5th and 95th percentiles.
  • Figure 2: Feature Importances (FIs) per model. Median FIs (expressed in increased units of NRMSE) for a) GRUs, b) LSTMs, and c) RNNs. Features are sorted in decreasing order by model.
  • Figure 3: GPP predictions for three example sites. The first column shows predictions for a) DE-Hai (DBF), c) BE-Vie (MF), and e) FI-Var (ENF). The second column displays zoom-in panels depicting extreme events at each site.