Machine-learning projections of boreal forest gross primary productivity under future climate change

George, Sandip Varkey; Pope, Cameron F. T.; Varghese, Stella Jes; Sunny, Ron

doi:10.5194/egusphere-2026-3281

Preprints

https://doi.org/10.5194/egusphere-2026-3281

Preprints

24 Jun 2026

| 24 Jun 2026

Machine-learning projections of boreal forest gross primary productivity under future climate change

Sandip Varkey George, Cameron F. T. Pope, Stella Jes Varghese, and Ron Sunny

Abstract. Modelling forest productivity under climate variability and change remains a key challenge in terrestrial ecology. The increasing availability of long-term eddy covariance data enables these relationships to be explored using data-driven machine-learning approaches. In this study, we model the relationship between four key environmental variables, namely, photosynthetically active radiation (PAR), atmospheric CO₂ concentration, air temperature, and relative humidity, and gross primary productivity (GPP) using linear regression, eXtreme Gradient Boosting (XGBoost), and deep neural networks. To capture variability across temporal scales, the input variables are decomposed into trend, seasonal, and residual components representing long-term, seasonal, and inter-daily fluctuations. The models are trained on seventeen years (2003–2019) of daily observations from the boreal forest monitoring station at Hyytiälä, Finland, and evaluated on an independent five-year period (2020–2025). The trained models are then used to project GPP under four Shared Socioeconomic Pathway (SSP) scenarios from the Intergovernmental Panel on Climate Change using climate forcings derived from Coupled Model Intercomparison Project Phase 6 (CMIP6) simulations.

All models show strong predictive skill (R²=0.79–0.90; RMSE =1.03–1.49), with the DNN performing best overall. SHAP (SHapley Additive exPlanations) analysis identifies the residual component of PAR as the most influential predictor of GPP across all models. More broadly, residual components across multiple inputs show high predictive importance, suggesting that short-term variability in environmental conditions may play an important role in explaining modeled GPP fluctuations. Projections driven by an ensemble of four CMIP6 climate models suggest relatively stable GPP during the mid-century period across SSP scenarios, followed by an overall increase toward the end of the century, particularly under higher-emission pathways. Overall, this study demonstrates the potential of combining long-term ecosystem observations, engineered environmental features, and climate-model projections to generate localized forecasts of forest productivity. However, we highlight the need for cautious interpretation of these results, since these are data-driven models that are being extrapolated beyond historical climate conditions.

Received: 05 Jun 2026 – Discussion started: 24 Jun 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Sandip Varkey George, Cameron F. T. Pope, Stella Jes Varghese, and Ron Sunny

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-3281', Anonymous Referee #1, 30 Jul 2026
General comments
The manuscript focus is on ML-based modelling and future projections of gross primary productivity at a site in Finland. ML models are trained on the data set of ca 10 years, then tested on a 5 years set, and then projections are done on a monthly scale using CMIP6 simulations as features. Modelling GPP is not a novel topic, specifically GPP (compared to NEE) models provide typically high statistical scores, especially so in boreal forests where variability is moderate from year to year. Also at this specific site, GPP has been modelled using different methods before. So potentially I see novelty in the method, when ML modelling is done separating a trend, seasonal cycle and residuals. However, I have concerns regarding data sets, results and interpretation.
Specific comments
How the measurement data set is treated: in my opinion, the authors provide too little detail on how the data sets were prepared, especially having in mind that some data sets were gapfilled and some sensors were used interchangeably. Please report exactly variables used, measurement heights, and percentages/periods of missing/gapfilled data for all the variables.

Detrending: I would not call moving mean a classical method in time series analysis but simple – yes. Related to how the seasonal cycle was quantified: the authors found average over all the years and subtracted it from the time series every year. I suspect that this is why the seasonal cycle has become less important than residuals. Looking at residuals, there are clearly remainders of the seasonal cycle not only short-term variability; the length and exact dates of the growing season of course change somewhat from year to year. The models might utilize residuals better than the average seasonal cycle, which is the same every year. In view of this, I would urge from interpreting the results as it is done in L220-221: ‘the models primarily respond to short-term fluctuations in the environmental drivers, rather than relying solely on the shared seasonal cycle…’.

I have doubts about increase in winter photosynthesis in the most extreme scenarios. Whatever warming occurs, radiation at this site is very low in winter months (which one can see in Fig. B1). As radiation is the main driver of photosynthesis, a reader needs to be better convinced that this increase in photosynthesis can happen, and at least some explanations of this result are required.

Table 1 caption: ‘GPP is measured’ –> ‘GPP units’ should be used instead, it is not measured. Aslan et al (2024) reported measured annual GPP from this site in the same units (https://www.sciencedirect.com/science/article/pii/S016819232400176X, Fig. 6b) – and they are about twice as large compared to what is reported in the table. Could the authors comment on this difference?

The authors mention thinning, which brings up a question about interannual variability. In the above-mentioned study by Aslan et al, it is shown that GPP is ca 30% lower in thinning year 2020 than in ‘normal’ years. This is not captured by the model, all the years look the same in Fig. 2 (where I would suggest making the blue curve at least partly transparent to better see how close measured and modelled points are). I think that interannual variability is an important point for discussion in this study, which aims to make future projections. How well is interannual variability captured otherwise, excluding the thinning year? Should one trust future projections made by the models which do not capture interannual variability?
Citation: https://doi.org/10.5194/egusphere-2026-3281-RC1
- AC1:
  'Reply on RC1', Sandip George, 09 Sep 2026
  We thank the reviewer for the detailed and constructive comments. We have considered the points raised carefully and outline below our planned revisions to address the concerns raised. A full revised manuscript and detailed response will be provided after considering the comments from both reviewers. The response to each comment is numbered using the same numbering as the original comments
  Thank you for this comment. We will expand the revised version with details of the dataset preparation, particularly focusing on how different datasets were combined. We will provide a table specifying, for each variable, the measurement height, data source/sensor, periods and percentage of missing data, periods and percentage of gap-filled data, and how data from interchangeable sensors were combined.
  
  We thank the reviewer for pointing this out. We agree that the interpretation of the feature importance should be more cautious, particularly given that the residual component may retain aspects of the year-to-year variation in the seasonal cycle. We will revise the text accordingly and avoid attributing the residual component exclusively to short-term variability. We will also replace the moving-mean decomposition with Seasonal-Trend decomposition using LOESS (STL), which allows the seasonal component to vary more flexibly over time. The feature importance analysis will be updated accordingly.
  
  We thank the reviewer for this insightful comment. We agree that the winter predictions reflect the model being applied under low-PAR conditions outside those well represented in the training data. We have therefore incorporated a light-response gate based on a sigmoidal function, with its parameters derived from the empirical PAR–GPP relationship in the training data. This physiologically motivated constraint substantially reduces winter GPP under low-light conditions.
  
  We thank the reviewer for pointing out this reference, which we had missed, and for highlighting the difference in the reported annual GPP. The difference is partly related to our use of median rather than mean annual GPP. On further consideration, we agree that the mean is more appropriate here, as it allows more direct comparison with existing estimates from this site. We will therefore report mean GPP in the revised manuscript. Using the mean, our estimates are much closer to those reported by Aslan et al. (2024).
  
  We thank the reviewer for this important comment. As they correctly point out, the model is unable to capture GPP reductions associated with thinning years, as thinning is not represented in the model inputs. More generally, because the model is instantaneous and has no memory or time-dependent component, it cannot capture lagged effects of previous weather conditions, such as drought or heat stress. It can, however, respond to environmental conditions at the time of prediction and therefore capture interannual differences arising from differences in contemporaneous environmental conditions. We therefore view the projections as estimates of the expected response to changes in the environmental drivers represented in the model, rather than predictions of individual-year GPP. They should consequently not be interpreted as capturing disturbance-driven or lagged interannual variability that is not represented in the predictor variables. In the revised manuscript, we will discuss this limitation more explicitly and provide annual GPP estimates for the testing period (excluding the thinning year) to assess the extent to which interannual variability is captured. We will also discuss how this limitation should be considered when interpreting the future projections.
  
  Citation: https://doi.org/10.5194/egusphere-2026-3281-AC1
RC2:
'Comment on egusphere-2026-3281', Anonymous Referee #2, 27 Sep 2026
This study models daily GPP at a boreal forest site using linear regression, XGBoost and a deep neural network driven by four environmental variables, and then projects GPP under four SSP scenarios using bias-corrected CMIP6 forcings. The main novelty lies in decomposing each driver into trend, seasonal and residual components, which allows the contributions of different timescales to be examined. Although this approach is potentially useful, I have the following concerns, mainly regarding the extrapolation of the models beyond historical climate conditions, and the decomposition procedure.
The authors acknowledged the risks of overfitting (Lines 55-57) and out-of-distribution extrapolation (Lines 292-294). I was therefore expecting an assessment of the model’s extrapolation behaviour, but I couldn’t find any relevant evaluation. The only supporting argument is that the projected GPP increase is “directionally consistent” with previous large-scale studies (Lines 295-300). However, the two cited studies themselves diverge substantially. Lu et al. (2024) projected modest increase in cold-zone GPP by 2100, ranging from ~5% under SSP1-2.6 to ~17.4% under SSP5-8.5 relative to 2021. Knauer et al. (2023) projected a two- to three-fold increase under RCP8.5 relative to the 1976-2005 baseline. These studies highlight the large uncertainty in future boreal GPP projections, so agreement with either one provides limited support for the present results. Also, the comparisons are not made on a common basis. The 126% increase reported here was calculated from mid-century to end-century (please also specify the periods referred to as the mid-century and the end of the century), whereas the above two studies used 2021 and 1976-2005 as their respective baselines. The studies also differ in scenario framework (SSP and RCP), and spatial scale (single site vs. climate zone/latitudinal band). It’s therefore unclear in what sense the present results were “consistent” with previous studies. My concern is that agreement in the direction of the GPP change does not demonstrate that the model extrapolates reliably.

According to figure 1, both observational data and CMIP6 data undergo seasonal decomposition, but the procedure at the monthly scale is not described. Please clarify 1) the window used to estimate the trend component of the monthly time series, for example, a 12-month centered moving average. 2) how the seasonal component was computed and over which period. For CMIP6 projections in particular, it is unclear whether the seasonal component was derived from the historical observations, from the entire CMIP6 time series, or from separate future periods.

The authors stated that “a significant contribution to this divergence arises from increased productivity during the winter months” (Lines 238-239) in figure 6. However, figure 6 is difficult to interpret in several aspects. First, in June-August, GPP under SSP1-2.6 exceeded that under SSP2-4.5 and SSP3-7.0, and was close to SSP5-8.5. This requires an explanation. Does it reflect scenario differences or different responses to summer climates? Second, monthly GPP under SSP1-2.6 was negative in winter, please explain why this occurs, whether an output constraint was applied. Also, the unit of figure 6 (gC m^-2 yr^-1 for monthly values) was a bit confusing.

Lines 217-218: for DNN, the seasonal and residual components of CO2 are among the most influential features. However, the site-measured CO2 concentration is not an independent driver of GPP, because it is strongly influenced by the ecosystem carbon fluxes themselves. The seasonal cycle of CO2 at a boreal site is largely driven by ecosystem photosynthesis and respiration, for example, CO2 concentrations are drawn down during the growing season. The CO2 seasonal component is therefore close to a mirror image of the GPP seasonal cycle. Its high importance may reflect the shared seasonal cycle that the decomposition was intended to separate, rather than a direct effect of CO2 on GPP.

Lines228-229: Please report the exact sample size of this test set, that is, the 10% of the monthly SMEATR II data, and also state whether these months were selected randomly or as a contiguous period.

Figure 4 and Line 255 indicate that “the month of the year” was used as a feature but this is not mentioned in the description of the predictor variables (Lines 83-90). Please state this explicitly and describe how the month was encoded, for example as an integer from 1 to 12. Also, this feature appears redundant with the seasonal components. The seasonal component is the multi-year average of the detrended time series for each day of the year. It is therefore identical in every year and fully determined by the calendar date, so it already contains the information carried by “month” at a finer resolution. I’d suggest that the authors report model performance and feature importance without the month feature and without the seasonal components, respectively, to show how much of the predictive skill depends on calendar information.

Lines 267-269: the authors cited studies that identified VPD and soil temperature, along with PAR and air temperature, as dominant drivers of carbon fluxes at boreal sites. Both variables appear to meet all three criteria in the predictor selections (Lines 83-85). Please explain why these variables were not included. If the choice was constrained by the variable’s availability, this should be stated explicitly, and the implications should be discussed. VPD can be derived from air temperature and relative humidity. I also recommend testing whether replacing RH with VPD, or adding VPD, changes the model performance and the projections.
Citation: https://doi.org/10.5194/egusphere-2026-3281-RC2

Sandip Varkey George, Cameron F. T. Pope, Stella Jes Varghese, and Ron Sunny

Viewed

Total article views: 349 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
217	98	34	349	24	19

HTML: 217
PDF: 98
XML: 34
Total: 349
BibTeX: 24
EndNote: 19

Views and downloads (calculated since 24 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	52	35	10	97
Aug 2026	37	18	7	62
Sep 2026	94	34	11	139
Oct 2026	34	11	6	51

Cumulative views and downloads (calculated since 24 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	52	35	10	97
Aug 2026	37	18	7	62
Sep 2026	94	34	11	139
Oct 2026	34	11	6	51

Viewed (geographical distribution)

Total article views: 310 (including HTML, PDF, and XML) Thereof 310 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 14 Oct 2026

Short summary

Forests play an important role in absorbing carbon dioxide, but their response to climate change is difficult to predict. We used long-term observations from a forest in Finland to train machine learning models of productivity and applied future climate projections to explore changes under different scenarios. We observe that the models capture seasonal influences well and provide useful estimates of future productivity, offering a framework for local predictions using large-scale climate data.


Total:	0
HTML:	0
PDF:	0
XML:	0