the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 08 Apr 2026
The growing seasons of global forest ecosystems from 1850 to 2100 estimated with a probabilistic temperature-based model
Abstract. Global climate warming has significantly altered forest phenology in the past decades, with measurable shifts in the timing and duration of the growing season (GS). These changes are expected to intensify in the future, potentially affecting both ecosystem productivity and land-atmosphere interactions. Accurately representing GS dynamics is therefore essential for assessing ecosystem vulnerability and improving the representation of vegetation processes in Earth system models. Here, we introduce GS-P, a probabilistic, temperature-based model developed within a machine-learning framework to estimate the start and end of the growing season (SGS and EGS) for global forest ecosystems over the period 1850–2100.
Results show stable GS timing until the 1970s, followed by significant shifts characterized by earlier SGS and later EGS, leading to a global extension of the GS. Under future climate scenarios, GS duration is projected to increase by approximately one month under low-emission conditions and up to two months under high-emission scenarios, with stronger responses in the Northern Hemisphere. Compared to alternative models, GS-P achieves comparable or improved predictive accuracy while exhibiting greater extrapolation capabilities and providing explicit uncertainty estimates. Furthermore, the model effectively represents key ecological features, such as stronger temperature control and greater spatial heterogeneity in spring than autumn phenology, and detection of regions where temperature alone provides limited explanatory power, suggesting a stronger role of additional drivers. Additionally, GS-P enables the identification of regions characterized by transitional states and high prediction uncertainty, potentially reflecting climate–ecosystem disequilibrium and enhanced ecosystem vulnerability. This model provides a flexible and interpretable framework for simulating GS dynamics at the global scale, offering improved constraints for carbon cycle modelling and supporting the assessment of ecosystem responses to future climate change.
- Preprint
(1709 KB) - Metadata XML
-
Supplement
(3508 KB) - BibTeX
- EndNote
Status: open (until 15 Aug 2026)
-
RC1: 'Comment on egusphere-2026-1710', Anonymous Referee #1, 06 Jul 2026
reply
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-1710/egusphere-2026-1710-RC1-supplement.pdfReplyCitation: https://doi.org/
10.5194/egusphere-2026-1710-RC1 -
RC2: 'Comment on egusphere-2026-1710', Anonymous Referee #2, 22 Jul 2026
reply
The paper presents a very simple phenology model that is only based on temperature. They calibrate the model on modeled phenophases based on modis observations and then calculate start and end of the growing season based on the temperature simulated by the UKESM. They compare there model to two very simple semi-parametric phenology models both of which were not developed for global coverage. The presented model does not clearly outperform these models.
In the introduction, the authors claim (lines 78 - 88):
- Several reasons motivate this effort. First, such a product can provide continuous and consistent phenological estimates for forest ecosystems, allowing us to infer changes in phenological transitions during the considered two and a half centuries. This product may help reduce uncertainties in terrestrial carbon cycle models and could support adaptive forest management
strategies in a changing climate.
But this would only be helpful if the estimations are faithful. The UKESM has an LAI estimate itself. I do not see how producing an additional estimate based only on temperature is helpful especially if it does not outperform very simple models.
- Second, we seek to improve the capability of phenological modelling in predicting the SGS and EGS, and to assess the potential advantages of machine-learning approaches over more traditional methods. Namely, we aim to overcome previous limitations in spatial extent, species specificity, and extrapolation skills under climate change scenarios, by providing a robust, generalizable framework for prediction of phenological stages.
I also do not see this point. First, other work using machine learning to predict phenology (phenophases, LAI, greenness, NDVI or EVI) exists and second, I do not see from the results how this work overcomes these limitations as the authors do not compare to state of the art models and do not beat models clearly behind the state of the art.
- Third, using temperature as the sole predictor allows us to clarify its role as a driver of phenological transitions at global scale. In this context, model
uncertainty is not only a measure of predictive performance but may also help to identify regions where additional controls
on phenological timing (e.g., precipitation seasonality) are likely to be more important than temperature alone, and where
future climate change may influence forest phenology and, more broadly, forest ecosystem functioning.
I disagree about this point as well. Using a model that predicts only from temperature ignores that temperature may co-vary with other inputs. Predictability from temperature is not the same as reaction to an intervention. Ignoring the other factors does exactly the opposite. It turns a blind eye and does not allow for understanding the importance of temperature.
To solve these problems the main thing the authors need to do is compare their results to the state of the art and demonstrate a more specific gap they want to close. I understand that the authors claim that their model is using only temperature and hence they compare only to other temperature based models but, as discussed above, I do not see how this is advancing the state-of-the-art.
In addition to the chosen baseline model, also the data-split is not chosen well. Since the target is to generalize in time and to unknown climate, the authors should also reflect this in the data-split. They should ideally cut a test-set from the end of the time series and not from the middle as well as cutting out a region with a different climate to check the generalization. In addition, the data split is not well described as it is not clear what constitutes an observation.
As described above, I do not think the interpretation of the model parameters for this very simple and under performing model is meaningful.
I also think the results for the growing season in the southern hemisphere are believable. Since they go against the current literature, they should be discussed thoroughly.
In the Discussion, the authors mention multiple different models that according to some picked results compare bad to their model. I think incorporating these models into the main analysis or at least picking the best model from these reviews and compare to them instead of the two random models the authors selected.
In Summary, I would recommend to either demonstrate that we can learn something from the temperature only model or to compare to real baselines.
Citation: https://doi.org/10.5194/egusphere-2026-1710-RC2 - Several reasons motivate this effort. First, such a product can provide continuous and consistent phenological estimates for forest ecosystems, allowing us to infer changes in phenological transitions during the considered two and a half centuries. This product may help reduce uncertainties in terrestrial carbon cycle models and could support adaptive forest management
Data sets
GS-P output: the growing seasons of global forest ecosystems from 1850 to 2100 estimated with a probabilistic temperature-based model P. R. Guaita et al. https://doi.org/10.5281/zenodo.19224585
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 235 | 71 | 19 | 325 | 28 | 12 | 22 |
- HTML: 235
- PDF: 71
- XML: 19
- Total: 325
- Supplement: 28
- BibTeX: 12
- EndNote: 22
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1