the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 17 Oct 2023
Learning Extreme Vegetation Response to Climate Forcing: A Comparison of Recurrent Neural Network Architectures
Abstract. Vegetation state variables are key indicators of land-atmosphere interactions characterized by long-term trends, seasonal fluctuations, and responses to weather anomalies. This study investigates the potential of neural networks in capturing vegetation state responses, including extreme behavior driven by atmospheric conditions. While machine learning methods, particularly neural networks, have significantly advanced in modeling nonlinear dynamics, it has become standard practice to approach the problem using recurrent architectures capable of capturing nonlinear effects and accommodating both long and short-term memory. We compare four recurrence-based learning models, which differ in their training and architecture: 1) recurrent neural networks (RNNs), 2) long short-term memory-based networks (LSTMs), 3) gated recurrent unit-based networks (GRUs), and 4) echo state networks (ESNs). While our results show minimal quantitative differences in their performances, ESNs exhibit slightly superior results across various metrics. Overall, we show that recurrent network architectures prove generally suitable for vegetation state prediction yet exhibit limitations under extreme conditions. This study highlights the potential of recurrent network architectures for vegetation state prediction, emphasizing the need for further research to address limitations in modeling extreme conditions within ecosystem dynamics.
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(1647 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(1647 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2023-2368', Anonymous Referee #1, 23 Jan 2024
The authors conducted a thorough comparison of recurrent neural networks on NDVI prediction. The manuscript is well-written and easy to follow. It would be a good fit to be published in Nonlinear Processes in Geophysics. Here are several comments to be addressed before publication.
Title. Considering that the models applied to all NDVI time series and extreme cases are analyzed for performance comparison, I would suggest removing “extreme” from the title to better reflect the broader application of the model
Section 2.1. The models are trained separately for each site. Is it feasible to train a universal model with all the stations by incorporating some static features (such as elevation, latitude, longitude, etc.,)? There have been studies showing the LSTM model benefits from diverse training datasets. How might this approach impact conclusions, especially regarding whether gated models benefit more from augmented training data given their complex architectures?
DL features. The input features need to be clarified, particularly mean temperature and sea level pressure. Are these area means over the European continent, or are they in-situ measurements? If the latter, how do the gridded dataset correlate with in-situ NDVI values? Additionally, clarify the input time window size for the model.
Figure2, I believe subfigure a is for conventional RNN, LSTM and b is for ESN. The figure caption is mismatched.
Figure4, is the black line for the real/target value? I would suggest adding it in the legend.
Figure5, the legend color in the right panel is missing. Please revise it for completeness.
Table 1. Clarify whether the standard deviation is derived from the 20 selected sites.
I would suggest adding a comparison of training and inference speed of the selected networks for a more comprehensive evaluation.
Citation: https://doi.org/10.5194/egusphere-2023-2368-RC1 - AC1: 'Reply on RC1', Francesco Martinuzzi, 21 May 2024
-
RC2: 'Comment on egusphere-2023-2368', Anonymous Referee #2, 10 May 2024
In general, the manuscript is well-written and organized. It would have an impactful contribution to literature. Here are my comments that need to be addressed.
I advise the authors to change the places where "temperature" is written to "air temperature". And “global radiation” to “global solar radiation”.
Figure 1 a: The scale and north arrow should be added to the map.
Description of Fig 1. A typo “radii” needs to be corrected.
L115-116: Did you have in situ NDVI measurements to compare MODIS values?
Figure 2: It would be good to show differences in network architectures.
Section 2.4: What is the ratio between training and testing data?
Figure 3e: Please explain the meaning of the red-shaded area here?
Section 2.6.1: Please add a statement indicating at which values these performance indicators work well.
L246: Please use another abbreviation for entropy-complexity. In line 117, EC stands for "eddy covariance".
Figure 5b: The legend should be corrected.
Appendix B: Please add computational resources (GPU, etc.)?
Data Availability section: You should add MODIS data source link.
Citation: https://doi.org/10.5194/egusphere-2023-2368-RC2 - AC2: 'Reply on RC2', Francesco Martinuzzi, 21 May 2024
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2023-2368', Anonymous Referee #1, 23 Jan 2024
The authors conducted a thorough comparison of recurrent neural networks on NDVI prediction. The manuscript is well-written and easy to follow. It would be a good fit to be published in Nonlinear Processes in Geophysics. Here are several comments to be addressed before publication.
Title. Considering that the models applied to all NDVI time series and extreme cases are analyzed for performance comparison, I would suggest removing “extreme” from the title to better reflect the broader application of the model
Section 2.1. The models are trained separately for each site. Is it feasible to train a universal model with all the stations by incorporating some static features (such as elevation, latitude, longitude, etc.,)? There have been studies showing the LSTM model benefits from diverse training datasets. How might this approach impact conclusions, especially regarding whether gated models benefit more from augmented training data given their complex architectures?
DL features. The input features need to be clarified, particularly mean temperature and sea level pressure. Are these area means over the European continent, or are they in-situ measurements? If the latter, how do the gridded dataset correlate with in-situ NDVI values? Additionally, clarify the input time window size for the model.
Figure2, I believe subfigure a is for conventional RNN, LSTM and b is for ESN. The figure caption is mismatched.
Figure4, is the black line for the real/target value? I would suggest adding it in the legend.
Figure5, the legend color in the right panel is missing. Please revise it for completeness.
Table 1. Clarify whether the standard deviation is derived from the 20 selected sites.
I would suggest adding a comparison of training and inference speed of the selected networks for a more comprehensive evaluation.
Citation: https://doi.org/10.5194/egusphere-2023-2368-RC1 - AC1: 'Reply on RC1', Francesco Martinuzzi, 21 May 2024
-
RC2: 'Comment on egusphere-2023-2368', Anonymous Referee #2, 10 May 2024
In general, the manuscript is well-written and organized. It would have an impactful contribution to literature. Here are my comments that need to be addressed.
I advise the authors to change the places where "temperature" is written to "air temperature". And “global radiation” to “global solar radiation”.
Figure 1 a: The scale and north arrow should be added to the map.
Description of Fig 1. A typo “radii” needs to be corrected.
L115-116: Did you have in situ NDVI measurements to compare MODIS values?
Figure 2: It would be good to show differences in network architectures.
Section 2.4: What is the ratio between training and testing data?
Figure 3e: Please explain the meaning of the red-shaded area here?
Section 2.6.1: Please add a statement indicating at which values these performance indicators work well.
L246: Please use another abbreviation for entropy-complexity. In line 117, EC stands for "eddy covariance".
Figure 5b: The legend should be corrected.
Appendix B: Please add computational resources (GPU, etc.)?
Data Availability section: You should add MODIS data source link.
Citation: https://doi.org/10.5194/egusphere-2023-2368-RC2 - AC2: 'Reply on RC2', Francesco Martinuzzi, 21 May 2024
Peer review completion
Journal article(s) based on this preprint
Model code and software
Code for machine learning models Francesco Martinuzzi https://github.com/MartinuzziFrancesco/rnn-ndvi
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 451 | 219 | 42 | 712 | 27 | 22 |
- HTML: 451
- PDF: 219
- XML: 42
- Total: 712
- BibTeX: 27
- EndNote: 22
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Francesco Martinuzzi
Miguel D. Mahecha
Gustau Camps-Valls
David Montero
Tristan Williams
Karin Mora
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(1647 KB) - Metadata XML