There are some limitations to the EWS methods detailed here. These include, but are not limited to, availability of data used to monitor the systems, and the properties of the system assumed to be able to measure the EWS, such as the presence of a tipping point and the underlying timescales of the system.
Most importantly, it is worth noting that, for EWS to be used appropriately as an early warning of a tipping point, we require prior independent evidence that the system in question can actually exhibit tipping behaviour, as opposed to losing resilience with no tipping point or alternative state for the system to tip to. This evidence could come from established theory, models, or palaeo data. A subset of EWS can be used to monitor resilience in systems where a tipping is not necessarily expected, however these are not discussed here.
To make a robust assessment with EWS of a system, there is a requirement for high-quality data. Temporal EWS require a complete time series dataset which is sufficiently long to capture the relevant timescale of the system; infilling missing data points (which can be common in observational records) can interfere with the EWS, while shorter time series may not accurately detect changes in resilience.
Even with suitable amounts of data, there are inherent limitations associated with EWS. While theoretically, AR(1) should equal 1 when a system reaches a tipping point, these realworld systems are exposed to noise and can tip prior to this. Furthermore, the act of detrending the time series in the process calculating EWS changes the absolute value of AR(1). This means that, while EWS might tell us when a system is losing resilience, without a sufficiently dense dataset and knowledge about internal dynamics of the system, they cannot usually give a measure of the distance to a tipping point. However, robustly checking the tendency of these indicators (such as Kendall’s tau) while varying the detrending technique and window length used to calculate the indicator on can provide useful information on the movement towards tipping.
Usually, it is assumed that a system approaching a tipping point is forced slowly towards it, and forced on shorter timescales by perturbations (which can be thought of as like weather in climate systems). It is generally assumed that this short-term noise is independent and identically distributed with a mean (average) of zero. This is unlikely to be the case in reality, with climate systems experiencing extreme weather events, for example, which are likely becoming more prevalent with the changing climate. Furthermore, extreme weather events would also increase the variance of the short-term noise over time, which also hampers the ability to use EWS indicators. To tackle this, we propose measuring EWS on the drivers themselves (e.g. on rainfall for vegetation systems) to check if changes in autocorrelation and variance in these are related to those found in the system being monitored.
Remote sensing products from satellite observations are a great resource of generally freely available data for using EWS on, and can enable complementary analyses to on-the-ground measurements of things that cannot be measured from space. They provide long records of climate systems, allowing us to create a long enough EWS indicator from which to get reliable results. However, due to sensor degradation and upgrades, it can be challenging to get a long time series from a single sensor, and products are often created from combined data sources. This can interfere with the EWS that we have described here, particularly AR(1) and variance, if this merging changes the signal-to-noise ratio (SNR) over time.
Newer sensors will measure with greater accuracy, increasing the SNR and in turn ‘erroneously’ increasing the AR(1) as far as an EWS is concerned, and a decrease in variance would also be expected. Anticorrelation between these two measures can show this is happening, whereas theory dictates that we should see an increase in both for a true EWS. In addition, newer remote sensors will also present shorter revisit times, as well as improved spatial resolutions, imposing the need to carefully consider the way data from different sensors are combined to produce long time series. Recently, we have become more aware of the effects of merging sensors and can prepare our analysis of these accordingly (Smith et al., 2023), such as only using data from a single sensor (Blaschke et al., 2023).
As well as questions around data availability and noise behaviour, the inherent timescale of the system being studied can hinder our ability to predict tipping points. While tipping is by definition a fast process, for slower-moving systems like the AMOC, the tipping event occurs over decades and it could therefore be difficult to detect the tipping point using EWS. Another example of this is the Amazon rainforest, where there is a slow decadal response of the forest based on climate change (1.3.2.1). It could take decades for dieback to occur even under a constant climate, such that a tipping point could be passed long before it is observed.
Part of the assumptions made around the occurrence of these EWS is that the system will approach a ‘bifurcation’ (a mathematically specific and common form of tipping point), rather than alternate forms of tipping. Alternatives include noise-induced tipping, where a system is shifted outside its stable state by a ‘stochastic’ (i.e. random) forcing, or rate-induced tipping, whereby a parameter changes too rapidly for the system to stay in the stable state (Ashwin et al., 2012). Rate-induced tipping can show some EWS (Ritchie and Sieber, 2016), such as threshold exceedance (detailed further in Chapter 2.5), while noise-induced tipping is generally unpredictable. For example, if the system is perturbed by something like an extreme weather event (e.g. a drought in the Amazon rainforest) such that it causes tipping by pushing the system past the ability for restoring feedbacks to return the system back to the previous state, CSD will not occur. However, bifurcation tipping and noise-induced tipping can be linked, whereby a system losing resilience approaching a bifurcation is more likely to be pushed to an alternate state by noise.
A related problem that may hamper EWS detection is that of cascading tipping points (see Chapter 1.5), where a tipping point in one system has a knock-on effect on another system, causing that to also tip. This can make it difficult for EWS to detect these tipping points, especially if the cascade causes instantaneous tipping points (a ‘joint cascade’) or happens soon after the first system tips (a ‘domino cascade’) (Klose et al., 2021).
Box
1.6.1
The EWS detailed here are not limited to use in climate and ecological systems; a recent study identified their use in other fields such as health, social systems and physical sciences (Dakos et al., 2023). Their utility in other domains is considered in later chapters in this report, specifically Section 2.5 – ‘Early warnings of tipping points in socio-economic systems’ and Section 4.5 – ‘Detecting “early opportunity indicators” for positive tipping points’. Particular EWS of note which are used in these chapters include:
Full details of the indicators can be found in the relevant sections.