1. Introduction
Ecosystems provide important services to man and society, such as water purification, regulation of climate, pests and diseases, pollination, provision of wildlife habitat, biodiversity conservation and the delivery of wood and products through ecosystem functions, such as biomass production [
1]. However, changes in average climate conditions, as well as increased frequency of climate extremes, suggested by the latest IPCC scenarios, threaten the stable delivery of these services [
2]. Consequently, understanding the stability of these ecosystem services is of critical importance. Methods that examine the stability of natural and productive systems using remote sensing can therefore be useful for this kind of assessment.
Vegetation or ecosystem stability is commonly measured using four main stability metrics. At the moment a disturbance occurs, the vegetation state may change, where the ability of the ecosystem to withstand the disturbance is referred to as resistance. After the disturbance, the ecosystem may return to its original state. The speed at which the ecosystem returns is denoted by engineering resilience. However, sometimes the disturbance is strong enough or changing conditions have diminished its engineering resilience, and the system might change its regime, i.e., find a new stable state. The magnitude of disturbance needed for the system to switch regime is called ecological resilience. Finally, variance denotes the total variability of the system in response to environmental anomalies [
3].
In order to quantify vegetation stability, the response of a vegetation state indicator, such as biomass, should be assessed in relation with the disturbances. The large-scale quantification of biomass or other vegetation state indicators is however difficult to achieve.The assessment is often labor intensive, costly and sometimes even destructive. Remote sensing time series of vegetation indices provide a means to retrieve indicators of vegetation stability at a global scale. Several indices, such as the Normalized Difference Vegetation Index (NDVI) or Enhanced Vegetation Index (EVI), have been developed, which are related to the biomass and greenness of vegetation [
4]. The characteristics of these remote sensing time series, and more specifically their anomalies, can be used to obtain large-scale spatial patterns of vegetation stability. For example, the anomaly value associated with the moment of disturbance is an indicator of the resistance of the system. The standard deviation can be related to the variance and the auto-correlation at lag one can be related to engineering resilience [
5].
Estimating resistance, engineering resilience or variance using a single metric on a complete anomaly time series, however, poses problems in a situation where the vegetation response changes over time [
6]. This is illustrated in
Figure 1, where two time series are shown with the same overall standard deviation and auto-correlation at lag one. While Time Series B is stationary (i.e., the characteristics of the time series, such as its variability, mean and the relationship between subsequent observations do not change over time), Time Series A shows changing behavior. Its auto-correlation at lag one and standard deviation increases over time, which could be interpreted as a sign of increased vulnerability (i.e., lower resilience and higher variability, respectively).
Assessment of such changes in vegetation response is of major importance for management purposes [
7]. It may provide a warning for increased vulnerability and may serve as an indication for land managers that additional attention is required. Moreover, prior to the critical point of divergence when a regime shift occurs, vegetation response may show an altered response [
8,
9]. Vegetation is assumed to recover more slowly, i.e., critical slowing down, and its variability may increase before the shift [
8,
9,
10]. Quantifying changes in vegetation recovery and variance thus potentially provides an extremely interesting asset for ecosystem management. It may allow one to monitor, predict and potentially intervene with ecosystems before a regime shift has occurred. The latter is of utmost importance, because these shifts may entail large economic and ecological losses, while their reversal may be extremely difficult [
11]. Consequently, if changes in vegetation response are not observed and considered in ecosystem monitoring, important information about vegetation stability may remain unnoticed with severe consequences for production and function.
Many studies have already taken advantage of the availability of long-term and large-scale satellite time series to quantify changes in vegetation response or vegetation dynamics. Several of these studies focused on temporal trends of vegetation greenness indicators (i.e., greening or browning; [
12,
13,
14,
15,
16]), revealing a greening trend in many parts of the world. Other studies quantified changes in phenology (e.g., [
17], rainfall use efficiency [
18], resilience [
10] or the relation between inter-annual temperature variability and northern vegetation activity [
19]. However, a study quantifying and related changes of the different aspects of vegetation stability, i.e., resistance, resilience and variance, has not been performed yet to our knowledge.
In previous studies [
5,
20], the utility of stability metrics such as standard deviation of the NDVI anomaly (SD; indicator of variance), autocorrelation at lag one of the NDVI anomaly (AC; indicator of resilience) and the correlation of NDVI anomaly (CS) with SPEI (Standardized Precipitation and Evaporation Index; indicator related to resistance), has been explored at the global scale. These studies indicated the potential of such metrics for detecting changes in vegetation stability and trends in vegetation cover associated with climate and land use change. Therefore, this study aims to test the usefulness of decomposing the NDVI time series signal into these metrics of vegetation stability for a regional environment subject to major climate fluctuations. More specifically, starting from a stability assessment on the complete time series, we aim to assess: (i) the magnitude and direction of stability changes; and (ii) the similarity in these changes for different stability metrics, which are indicators of resistance, resilience and variance. This study thus does not aim to draw detailed conclusions based on stability changes, but rather to investigate the usefulness of the concept. In order to frame the obtained stability metrics and changes in stability, the magnitude and direction of temporal vegetation stability changes will subsequently be linked to land cover types and changes in climate anomalies.
4. Discussion
Starting from stability metrics derived using the complete time series period, this study quantified the magnitude and direction of temporal vegetation response changes in Australia.
The stability metrics derived over the complete time series length showed that the spatial patterns of the stability metrics are related to those of the land cover types. For example, the tussock grasslands show a large variability and dependency on water availability, but a relatively low memory effect. This could partially be explained by their association with vertisols (i.e., deep cracking clay soils). These clay-rich soils are relatively fertile, but tend to form deep cracks during dry periods. Due to the important changes in soil conditions, vegetation biomass may fluctuate, as well, to a large extent. Tree-based land cover types and hummock grasses in the western part of Australia showed a insignificant relationship with droughts, but also a low memory effect and low to mediocre variability. This could suggest that these areas are one of the most stable of Australia: they recover relatively quickly and show nearly no response to droughts. Yet, other factors may also explain this type of response: (i) the hummock grasslands may be extremely sparsely vegetated, hampering the measurement of vegetation response; (ii) the NDVI response may saturate with relatively high LAI for tree vegetation in combination with (iii) a higher number of clouds in northern areas. These factors decrease the signal to noise ratio of the anomaly signal, complicating its ecological interpretation.
Next to the assessment of stability over the complete time period, temporal changes in stability were obtained. The temporal changes illustrate that the magnitude of temporal vegetation response change is relatively large compared to the stability metric derived over the complete time series period. This relatively large temporal variation has both technical and ecological consequences. First, the relatively large temporal variation in the stability metrics implies that vegetation response should not be considered as a constant, and this variability should be accounted for in vegetation stability assessment. Consequently, methods that are based on techniques implicitly assuming stationarity and using the whole time series period may not be optimal (e.g., [
20]). Modification of these techniques, i.e., through the use of a moving window, the allowance of break-points or the explicit inclusion of changing response may be of interest. Second, the stability metrics, and certainly the response to droughts, will depend on the used time series time frame and time series length. Consequently, the results may differ between sensors or platforms, but also between the first and second part of the time series. The latter was also found in the study of [
5], where the sensitivity of stability metrics to data characteristics and noise were assessed. Third, the result suggests that vegetation response changed over the 1982to 2006 period in Australia. As such, the assessment of these changes may reveal additional information about the vegetation response and may be interesting to monitor vegetation response.
The observed change in vegetation response may be an indicator of altered vegetation stability and may even be a precedent of regime shifts. A rise in SD or AC over time has for example been recognized as a possible indicator for regime shifts in both climatic systems and ecosystems [
8,
9,
11]. Yet, the spatial overview of temporal changes of the three stability indicators indicates that the metrics generally do not change in concordance. For example, the increased standard deviation and correlation with the SPEI index in the western part of Australia suggests that vegetation has become more sensitive and response more variable over time, while the auto-correlation shows a mixed result. The shrub and hummock dominated northern part shows an increased resilience, while the tree and agriculture dominated southern part shows the opposite. Vegetation thus tends to react in a more complicated way, which supports the need to assess multiple stability indicators.
Comparing the increase in standard deviation and auto-correlation with the resilience metric of [
41] in Australia further reveals contrasting patterns. This means that the probability of being in another state is not directly related to the change in stability. The analyses thus warn for a ‘blind’ application of the stability metrics on remote sensing data to assess vulnerability. As we did not test what exactly causes this lack of coherence, it is difficult to pinpoint just one cause. The remote sensing time series could still be too short to capture the increase in SD and AC that is supposed to be associated with ecosystems reaching a tipping point. Next, the states as defined by [
41] are relatively broad (i.e., forest, savanna and treeless). The sensitivity observed in our study could be related to more subtle changes than these broad conversions or ecosystems could behave in a different way than expected.
Furthermore, several factors may influence the observed changes in vegetation response. Both the standard deviation and auto-correlation indicate how the vegetation response changes over time, but not how these changes are related to environmental anomalies. The variance and memory effect of the vegetation may thus increase because the disturbance regime alters, while the vegetation remains inherently the same. This is partly confirmed by the significant positive rank correlation between changes in SD of the NDVI and climate anomalies for most of the pixels. This suggests that changes in vegetation response are partly driven by changes in climate anomalies, which is logical given the large dependency of vegetation in Australia to water availability and climate conditions (e.g., [
42]).
Yet, not all pixels show a simultaneous change in climate and vegetation anomalies, which may also be attributed to other environmental pressures. Australia has experienced a large influence of human activity (e.g., [
21,
43]), which may alter the vegetation response in a different way than expected from natural processes. As such, these conversions may hamper the interpretation of vegetation response changes as an altered sensitivity to climate anomalies. Other pressures than climate may further affect the vegetation state: factors such as altered fire regimes, overgrazing, introduction of invasive species and salinization are important, as well, and may furthermore alter the response of vegetation to climate anomalies [
21,
44,
45,
46]. Moreover, CO
2 fertilization, altered fire and rainfall regimes may change the vegetation composition, which was suggested in the study of [
13]. This may further be aggravated by the low spatial resolution of the GIMMS NDVI dataset. The low resolution results in variable fractional values of the dominant land cover type and increases the multitude of vegetation responses within a pixel, therefore complicating the interpretation. An analysis using higher resolution data in combination with a more complete set of drivers may provide a more detailed insight into the importance of each of these drivers and their relation with spatio-temporal changes in stability metrics. This however fell outside the scope of this study.
Finally, an undesirable and yet unavoidable factor of changes in stability metrics is the presence of noise in the time series [
5]. For example, the presence of clouds, high aerosol contents, image misregistration, Bidirectional Reflectance Distribution Function (BRDF) effects and sensor artifacts may introduce trends, white noise (i.e., random noise due to the combined effect of noise factors) and biased noise (e.g., negative spikes in the time series due to clouds), which may in their turn severely affect the stability metrics. As clouds and aerosol concentrations are not only highly spatially, but also temporally variable, the observed changes in stability may be affected, as well. However, the studies of [
5,
20] already showed that the signal-to-noise factor of the NDVI anomaly time series in Australia is relatively high, thus increasing the probability to reliably detecting changes. Moreover, the highest influence of noise was found in forested, tropical or completely bare areas, whereas semi-arid areas showed the highest signal-to-noise ratio [
5,
20]. The former areas were mostly masked in the CS analysis, as they did not show a significant correlation. Yet, a ground-based assessment of biomass stability would be interesting for further validation.