the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 27 Oct 2025
Complex Realities, Simple Signals? Global Evaluation of Early Warning Signals for Forest Mortality Events
Abstract. Forests around the world are increasingly experiencing large-scale regional mortality events as a result of droughts and heat waves. Despite their considerable impacts on the material, non-material, and regulatory contributions of forest ecosystems to people, these mortality events remain difficult to predict. Temporal Early Warning Signals (EWS) based on the concept of Critical Slowing Down (CSD) have been applied widely to remotely sensed vegetation indices. These EWS have often been interpreted as indicators of resilience loss. In order to be of practical use in real-world ecosystem management, such EWS must demonstrate the capacity to reliably and robustly predict upcoming forest mortality events. Previous work has applied EWS for local cases of mortality, but to date, there is no global assessment of EWS on remotely sensed vegetation indices of forest mortality events. The objective of this study is threefold: 1) to provide an overview of various types of EWS as applied to forest mortality events in case studies around the world, 2) to empirically assess the effectiveness of different EWS in predicting globally distributed forest mortality events driven by droughts and heat waves using remote sensing time series, and 3) to conduct a driver analysis to evaluate which factors associated with the methodological setup, the characteristics of the mortality event and climatic conditions explain the performance of different EWS. We find that most previous work in predicting forest mortality events is based on tree ring data. In remote sensing applications, there is a significant lack of robust evaluation of CSD-based EWS using control cases. Our empirical analysis indicates that all of the EWS that were evaluated in this study are ineffective and lack the necessary robustness to serve as predictors of drought-induced forest mortality events. The primary factor that determines trends in EWS is the methodological setup employed. We conclude by calling for more caution in the application of system-agnostic CSD-based EWS, increased efforts to assess accuracy and uncertainty, and more consideration of the system characteristics and actual needs of ecosystem managers when assessing forest resilience and early warning systems.
Competing interests: One of the co-authors is an editor on the special issue.
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.- Preprint
(53742 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on egusphere-2025-4902', Anonymous Referee #1, 11 Dec 2025
- AC1: 'Reply on RC1', Nielja Knecht, 04 Mar 2026
-
RC2: 'Comment on egusphere-2025-4902', Anonymous Referee #2, 27 Jan 2026
This manuscript presents a global analysis of commonly used Critical Slowing Down (CSD)–based early warning signals (EWS) applied to remote sensing vegetation indices. The authors combine a literature review with an empirical analysis using forest mortality events and nearby control locations. They conclude that standard CSD-based indicators do not robustly distinguish mortality events from controls and therefore have limited value for predicting forest mortality.
I have several concerns, which I summarize below.
An important assumption behind CSD-based early warning signals is not clearly stated in the introduction. These signals are expected only when there is a slow and gradual change in a particular driver, so that the system gradually loses resilience while remaining fluctuating around its equilibrium. Many forest mortality events, however, are caused by droughts or heatwaves that are abrupt or fast. In such cases, CSD is not always expected. Making this assumption explicit would help to better frame the results.
The manuscript further assumes that large-scale forest mortality events that lead to lasting changes in ecosystem composition or function can be treated as bifurcations that should be preceded by CSD. This assumption is too strong in my view. Persistent ecosystem change can also result from sudden shocks, fast environmental change, or combinations of stressors. In these cases, the absence of CSD is expected and does not mean that early warning signals “fail”.
At several points in the paper, early warning signals are (literally, e.g. line180)) discussed as if they should robustly predict mortality events. However, CSD-based indicators were developed to detect changes in recovery rates or local stability, not to predict specific events or to serve as management tools. It would help to clarify that the study tests whether these indicators can be used for mortality prediction, rather than evaluating them against expectations they were not designed to meet.
Control sites are selected to be similar to mortality sites before the event but without a collapse in the mortality year. It could well be that control sites may show similar loss of resilience as mortality sites but simply did not experience a triggering disturbance, and therefore no mortality event. At the same time, mortality sites may not have experienced a gradual background change at all. This makes it difficult to interpret differences, or lack of differences, between selected mortality and control sites.
The discussion of Smith et al. (2022) and relation of that work to this manuscript draws a strong distinction between engineering resilience and ecological resilience. While these concepts are mathematically different, CSD-based early warning signals are based on the idea that changes in recovery rates (engineering resilience) reflect changes in the size of the basin of attraction (ecological resilience), so they are an indirect measure of recovery rate, and therefore assumed to be an indicator of ecological resilience. Presenting these as largely separate may overstate the difference and weaken the theoretical argument.
In the conclusion, the authors argue that researchers should be cautious when promoting simplified, theory-based resilience indicators for ecosystem management. I agree that there is a clear gap between theory and practice. However, the study itself applies simplified theory to coarse and highly heterogeneous data, where poor performance is not surprising. The paper would benefit from a clearer discussion of the assumptions and limits of CSD theory, clearer alignment between expectations and theory, and a nuanced discussion on when these and other approaches are more appropriate.
Citation: https://doi.org/10.5194/egusphere-2025-4902-RC2 - AC2: 'Reply on RC2', Nielja Knecht, 04 Mar 2026
Status: closed
-
RC1: 'Comment on egusphere-2025-4902', Anonymous Referee #1, 11 Dec 2025
Summary
This study investigates the prediction power of CSD-based Early Warning Signals (EWS) to predict forest dieback events. To this end, the authors first conduct a literature review, during which they find that most of such studies are performed on tree ring data, and that ~33% of the studies are CSD-based. Most studies validate their results by comparing EWS between mortality and control sites. The authors conduct a drivers analysis, training one boosted regression tree model per EWS per vegetation data set, with the goal of predicting the Kendall tau of the EWS in the data set from various variables, comprising choices in the EWS computation, characteristics of the mortality event, and climate characteristics. The data is chosen if there was an occurrence of a forest mortality event, and the authors ensure a balanced training data set by including a non-dieback case for each dieback case, where the first is chosen to lie in the vicinity of the second. The authors claim that “if the EWS work to robustly predict mortality events, we would expect characteristics of the event, i.e., the resistance and reccovery rate of the die-back, to be the key predictors of Kendall tau trends in all eWS, jointly with climate information”. However, they find poor predictability. They find that rather, the most important features to predict the EWS are 1) timeframe, 2) window length, and less consistent, 3) resistance, 4) recovery rate (for most EWS-data combinations). They conclude that EWS based on CSD are thus ineffective and non-robust predictors of forest mortality events.
General comments
I agree that it would be great if we could predict forest mortality events. However, I wonder if the term Early Warning Sign (EWS) was optimally chosen when the scientific community agreed on using it for Critical-Slowing-Down (CSD) indicators. These indicators are – at best – statistical proxies of the recovery rate of the system, which decreases as the system approaches a fold bifurcation type tipping point. By design, they cannot warn of a tipping by themselves, and could probably be used as “EWS” only in combination with some kind of threshold. However, I am not aware of that and doubt it would be robust, as probably do Scheffer et al. (2009) as they write “We note also that most of the signals we have discussed should still be interpreted in a relative sense. For instance, although autocorrelation is predicted to approach unity at a fold bifurcation, measurement noise will tend to reduce correlations.” Hence, I also doubt that Scheffer et al. (2009) claim that “EWS based on the concept of alternative stable states and the theory of CSD have been proposed as a promising alternative approach to predict forest mortality events” (l 48f). Instead, often the change of the proxies is assessed, but all that these changes can tell us is IF the system is loosing stability. I thus believe that even though termed EWS, it is not the CSD-indicators fault that they cannot predict forest mortality, but rather the unfortunate wording that lead to overoptimistic applications in a way the indicators weren’t meant to be applied.However, I agree that the suitability of the CSD based indicators to inform us about tipping should be investigated and I believe that, after consideration of the theory’s possibilities and adapting the study to what can be expected, this study could contribute valuable insights. Namely, I believe the study could reflect some misunderstanding of the CSD-based theory, which I will try to outline
- The sentence “we assume for this study that the forest mortality events can be considered bifurcations, i.e., significant qualitative shifts in the system state” (l 274) highlights the first two of my concerns. First, a bifurcation is induced by some change in external parameters and resembles a structural change in the stability landscape of the system. For example, it could denote the change from a system with two stable and one unstable equilibrium to a system with only one stable equilibrium. This annihilation of a stable and an unstable equilibrium is the typical fold-bifurcation. This type of bifurcation is commonly assumed in “tipping” science, and is visualized e.g. by Box 1 Figure d in Scheffer et al. (2009), where F1 and F2 are denoting the fold bifurcations.
If the system was before in the system state associated with the now disappeared stable equilibrium, it would of course get attracted to the other, now only, stable equilibrium, resembling a qualitative shift in the system state. However, this is not the only way such a qualitative shift could happen. Depending on the distance of the unstable equilibrium to the stable equilibrium the system is currently in, more or less external energy (perturbations) is needed to “kick the system over to the other stable equilibrium”. - Second, as you say, the forest mortality data includes forest mortality from droughts or heat waves, so basically, the assumption “that the forest mortality events can be considered bifurcations” seems difficult to justify. From my understanding you do not investigates bifurcation induced qualitative shifts but rather, perturbation induced ones – which do not only depend on the system state’s resilience, but on how “large” the perturbation is in relation to the resilience. Regarding those, Scheffer et al (2009) already say that “perturbations will often trigger a transition well before a bifurcation point is reached. Thus, although a trend in the indicators may serve as a warning, the actual moment of a transition remains difficult to predict.” If, it would probably be reasonable to try to see if the “EWS” can tell us something about how “big” of a perturbation the system can absorb.
- Moreover, even if the “size” of the perturbation was correctly quantifiable, one would still be remained with the statistics (which is often misleadingly called “EWS”) that are only proxies of the resilience, or more precisely, the recovery rate, for which also one cannot be certain how much of it resembles the vegetation’s intrinsic time scales (we do not assume that boreal forest has the same speed in growing as e.g. a rainforest when both are completely healthy). Hence, I wonder if “Computing recovery rates instead of recovery time in this way allows us to compare different biomes with different ecological dynamics and prevents us from having to make assumptions about which mean vegetation index value the system has to return to” (l 188ff) should be reconsidered.
- As far as I understand, the Kendall tau of the statistic (AC1/var/…) are considered “the early warning signal”. However, the Kendall tau only informs us if the resilience is decreasing (positive trend), and how close this change moves the system towards its “tipping point” is highly dependent on how stable the system was in the beginning. Hence, the trend or Kendall tau is only informative of IF the system is loosing resilience, due to noise I would argue it doesn’t even inform us how much, but for sure not how close to the system is getting to the bifurcation point.
In summary, I believe that the conclusion “that despite their extensive utilization and ease of application, system-agnostic Early Warning Signals based on Critical Slowing Down are ineffective and non-robust predictors of forest mortality events on a global scale” cannot be “attributed to the poor approximation of ecosystem dynamics from optical vegetation indices and the multivariate character of real ecosystems” but rather that the theory was applied beyond the conceptual scope it can reliably represent. Which is, of course, an important notion to make.
Specific comments
- I find that the term “trend”, “Kendall tau”, and “EWS” are often used ambiguously, e.g. what are “Kendall tau trends”?
- “At true mortality locations, only a minority of cases show a statistically significant increase in AC1 before the event (41.7% ± 5.6% as mean±standard deviation as assessed across 1, 3, 5, 10, and all years before the event for NDVI). “ (l 236ff)
This makes me wonder if I understood your approach correctly. How do you get a trend from 1 year only, if your minimum window size for calculating the EWS was 1 year? - I would wish for a bit more of an explanation what “resistance” is supposed to mean. To me, it seems to be only the difference, but then later on you call it the “resistance to climate extremes” (l 253) so I know wonder if there is more to it than I see.
- How do you treat missing values which you might have a lot in the 250m resolution data sets?
Citation: https://doi.org/10.5194/egusphere-2025-4902-RC1 - AC1: 'Reply on RC1', Nielja Knecht, 04 Mar 2026
- The sentence “we assume for this study that the forest mortality events can be considered bifurcations, i.e., significant qualitative shifts in the system state” (l 274) highlights the first two of my concerns. First, a bifurcation is induced by some change in external parameters and resembles a structural change in the stability landscape of the system. For example, it could denote the change from a system with two stable and one unstable equilibrium to a system with only one stable equilibrium. This annihilation of a stable and an unstable equilibrium is the typical fold-bifurcation. This type of bifurcation is commonly assumed in “tipping” science, and is visualized e.g. by Box 1 Figure d in Scheffer et al. (2009), where F1 and F2 are denoting the fold bifurcations.
-
RC2: 'Comment on egusphere-2025-4902', Anonymous Referee #2, 27 Jan 2026
This manuscript presents a global analysis of commonly used Critical Slowing Down (CSD)–based early warning signals (EWS) applied to remote sensing vegetation indices. The authors combine a literature review with an empirical analysis using forest mortality events and nearby control locations. They conclude that standard CSD-based indicators do not robustly distinguish mortality events from controls and therefore have limited value for predicting forest mortality.
I have several concerns, which I summarize below.
An important assumption behind CSD-based early warning signals is not clearly stated in the introduction. These signals are expected only when there is a slow and gradual change in a particular driver, so that the system gradually loses resilience while remaining fluctuating around its equilibrium. Many forest mortality events, however, are caused by droughts or heatwaves that are abrupt or fast. In such cases, CSD is not always expected. Making this assumption explicit would help to better frame the results.
The manuscript further assumes that large-scale forest mortality events that lead to lasting changes in ecosystem composition or function can be treated as bifurcations that should be preceded by CSD. This assumption is too strong in my view. Persistent ecosystem change can also result from sudden shocks, fast environmental change, or combinations of stressors. In these cases, the absence of CSD is expected and does not mean that early warning signals “fail”.
At several points in the paper, early warning signals are (literally, e.g. line180)) discussed as if they should robustly predict mortality events. However, CSD-based indicators were developed to detect changes in recovery rates or local stability, not to predict specific events or to serve as management tools. It would help to clarify that the study tests whether these indicators can be used for mortality prediction, rather than evaluating them against expectations they were not designed to meet.
Control sites are selected to be similar to mortality sites before the event but without a collapse in the mortality year. It could well be that control sites may show similar loss of resilience as mortality sites but simply did not experience a triggering disturbance, and therefore no mortality event. At the same time, mortality sites may not have experienced a gradual background change at all. This makes it difficult to interpret differences, or lack of differences, between selected mortality and control sites.
The discussion of Smith et al. (2022) and relation of that work to this manuscript draws a strong distinction between engineering resilience and ecological resilience. While these concepts are mathematically different, CSD-based early warning signals are based on the idea that changes in recovery rates (engineering resilience) reflect changes in the size of the basin of attraction (ecological resilience), so they are an indirect measure of recovery rate, and therefore assumed to be an indicator of ecological resilience. Presenting these as largely separate may overstate the difference and weaken the theoretical argument.
In the conclusion, the authors argue that researchers should be cautious when promoting simplified, theory-based resilience indicators for ecosystem management. I agree that there is a clear gap between theory and practice. However, the study itself applies simplified theory to coarse and highly heterogeneous data, where poor performance is not surprising. The paper would benefit from a clearer discussion of the assumptions and limits of CSD theory, clearer alignment between expectations and theory, and a nuanced discussion on when these and other approaches are more appropriate.
Citation: https://doi.org/10.5194/egusphere-2025-4902-RC2 - AC2: 'Reply on RC2', Nielja Knecht, 04 Mar 2026
Model code and software
Code to accompany the paper Nielja Knecht https://gitfront.io/r/nielja/FjzWiFTmtTAP/25-09-ews-assessment/
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 2,332 | 3,028 | 205 | 5,565 | 166 | 170 |
- HTML: 2,332
- PDF: 3,028
- XML: 205
- Total: 5,565
- BibTeX: 166
- EndNote: 170
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Summary
This study investigates the prediction power of CSD-based Early Warning Signals (EWS) to predict forest dieback events. To this end, the authors first conduct a literature review, during which they find that most of such studies are performed on tree ring data, and that ~33% of the studies are CSD-based. Most studies validate their results by comparing EWS between mortality and control sites. The authors conduct a drivers analysis, training one boosted regression tree model per EWS per vegetation data set, with the goal of predicting the Kendall tau of the EWS in the data set from various variables, comprising choices in the EWS computation, characteristics of the mortality event, and climate characteristics. The data is chosen if there was an occurrence of a forest mortality event, and the authors ensure a balanced training data set by including a non-dieback case for each dieback case, where the first is chosen to lie in the vicinity of the second. The authors claim that “if the EWS work to robustly predict mortality events, we would expect characteristics of the event, i.e., the resistance and reccovery rate of the die-back, to be the key predictors of Kendall tau trends in all eWS, jointly with climate information”. However, they find poor predictability. They find that rather, the most important features to predict the EWS are 1) timeframe, 2) window length, and less consistent, 3) resistance, 4) recovery rate (for most EWS-data combinations). They conclude that EWS based on CSD are thus ineffective and non-robust predictors of forest mortality events.
General comments
I agree that it would be great if we could predict forest mortality events. However, I wonder if the term Early Warning Sign (EWS) was optimally chosen when the scientific community agreed on using it for Critical-Slowing-Down (CSD) indicators. These indicators are – at best – statistical proxies of the recovery rate of the system, which decreases as the system approaches a fold bifurcation type tipping point. By design, they cannot warn of a tipping by themselves, and could probably be used as “EWS” only in combination with some kind of threshold. However, I am not aware of that and doubt it would be robust, as probably do Scheffer et al. (2009) as they write “We note also that most of the signals we have discussed should still be interpreted in a relative sense. For instance, although autocorrelation is predicted to approach unity at a fold bifurcation, measurement noise will tend to reduce correlations.” Hence, I also doubt that Scheffer et al. (2009) claim that “EWS based on the concept of alternative stable states and the theory of CSD have been proposed as a promising alternative approach to predict forest mortality events” (l 48f). Instead, often the change of the proxies is assessed, but all that these changes can tell us is IF the system is loosing stability. I thus believe that even though termed EWS, it is not the CSD-indicators fault that they cannot predict forest mortality, but rather the unfortunate wording that lead to overoptimistic applications in a way the indicators weren’t meant to be applied.
However, I agree that the suitability of the CSD based indicators to inform us about tipping should be investigated and I believe that, after consideration of the theory’s possibilities and adapting the study to what can be expected, this study could contribute valuable insights. Namely, I believe the study could reflect some misunderstanding of the CSD-based theory, which I will try to outline
If the system was before in the system state associated with the now disappeared stable equilibrium, it would of course get attracted to the other, now only, stable equilibrium, resembling a qualitative shift in the system state. However, this is not the only way such a qualitative shift could happen. Depending on the distance of the unstable equilibrium to the stable equilibrium the system is currently in, more or less external energy (perturbations) is needed to “kick the system over to the other stable equilibrium”.
In summary, I believe that the conclusion “that despite their extensive utilization and ease of application, system-agnostic Early Warning Signals based on Critical Slowing Down are ineffective and non-robust predictors of forest mortality events on a global scale” cannot be “attributed to the poor approximation of ecosystem dynamics from optical vegetation indices and the multivariate character of real ecosystems” but rather that the theory was applied beyond the conceptual scope it can reliably represent. Which is, of course, an important notion to make.
Specific comments
This makes me wonder if I understood your approach correctly. How do you get a trend from 1 year only, if your minimum window size for calculating the EWS was 1 year?