Global-scale characterization of turning points in arid and semi-arid ecosystem functioning

Aim: Changes in dryland ecosystem functioning are threatening the well-being of human populations worldwide, and land degradation, exacerbated by climate change, contributes to biodiversity loss and puts pressures on sustainable livelihoods. Here, abrupt changes in ecosystem functioning [so-called turning points (TPs)] were detected using time series of Earth observation data. Hotspot areas of high TP occurrence were identified, observed changes characterized and insights gained on potential drivers for these changes. Location: Arid and semi-arid regions. Time period: 1982–2015. Methods: We


| INTRODUC TI ON
Drylands cover c. 40% of the land surface of the Earth and support roughly one-third of the global human population (Millennium Ecosystem Assessment, 2005). Dryland ecosystems are, however, highly susceptible to changes in their dynamics owing to extreme and/or persistent dry years (Barbier, Couteron, Lejoly, Deblauwe, & Lejeune, 2006;Fensholt et al., 2012;Rice, Matzner, Byer, & Brown, 2004) or human perturbations (Barbier et al., 2006;Reynolds et al., 2007). Persistent pressure on those ecosystems may trigger profound alterations in their functioning (Ponce-Campos et al., 2013), ultimately leading to land degradation and desertification. Recent estimates claim that 75% of the land area of the Earth is already degraded, and >90% could become degraded by 2050 (Scholes et al., 2018), notably as a result of climate change (Easterling et al., 2000).
The term "ecosystem functioning" refers to a state or trajectory of a system that accounts for the totality of complex interactions occurring inside it and caused by internal or external drivers (Jax, 2005). Alterations in the way dryland ecosystems function can take place abruptly, instead of gradually (Berdugo, Kéfi, Soliveres, & Maestre, 2017;Scheffer, Carpenter, Foley, Folke, & Walker, 2001). Those abrupt changes can happen either when a disturbance is strong enough to push the system to another state of functioning or when the system surpasses a threshold of environmental conditions and, consequently, cannot maintain its previous way of functioning (Scheffer et al., 2001). Here, we consider these abrupt changes as tipping points when there are reinforcing feedbacks maintaining the system at a certain functioning state (Lenton et al., 2008) or as turning points (TP) when the functioning changes significantly but not necessarily in an irreversible manner . Following this definition, a tipping point represents an extreme case of a TP.
Both TPs and tipping points in ecosystem functioning may have severe economic and ecological consequences, particularly in drylands (Millennium Ecosystem Assessment, 2005). When a dryland system shifts abruptly from a functional (e.g., being a source of forage) to a degraded state, the livelihoods of the inhabitants are threatened, which can result in increased poverty and food insecurity (Berdugo et al., 2017;Millennium Ecosystem Assessment, 2005). It is therefore essential to monitor such abrupt changes. Although vegetation dynamics (e.g., changes in vegetation activity throughout time) and shifts in vegetation trends have already been addressed in global drylands (de Jong, Verbesselt, Schaepman, & de Bruin, 2012;de Jong, Verbesselt, Zeileis, & Schaepman, 2013;Fensholt et al., 2012), few studies have used an ecosystem functioning approach to do so (e.g., Horion et al., 2016Horion et al., , 2019Ponce-Campos et al., 2013).
Characterization of TPs at large spatial scales is hampered mostly by the lack of long-term, temporally continuous and consistent field observations that can be scaled beyond the plot level. Time series of remotely sensed satellite imagery emerge as a viable solution (de Jong et al., 2012;De Keersmaecker et al., 2015;Tian et al., 2015), which can ideally be combined with groundbased information . However, conventional Earth observation indices [e.g., the normalized difference vegetation index (NDVI); Tucker, 1979] alone may not be optimal for assessing ecosystem functioning in drylands, because productivity in those regions is controlled significantly by precipitation (Herrmann, Anyamba, & Tucker, 2005;Hickler et al., 2005).
Therefore, NDVI trend changes may simply reflect variations in inter-annual rainfall dynamics, bearing little information about changes in ecosystem functioning.
The rain-use efficiency (RUE; Le Houérou, 1984), a key indicator for measuring the response of plant production to precipitation, is derived by calculating the ratio between net primary productivity (NPP) and precipitation. Given that precipitation is the main climate constraint to vegetation growth in arid and semi-arid areas, a basic assumption involved when using RUE for characterization of ecosystem changes is that there is an ecosystem-specific relationship between NPP and precipitation (Fensholt et al., 2013;Prince, Wessels, Tucker, & Nicholson, 2007). As such, RUE has been used as a key proxy for assessing ecosystem functioning Huxman et al., 2004), land degradation (Fensholt et al., 2013;Prince et al., 2007) or ecosystem resilience (Ponce-Campos et al., 2013). Indications of an altered ecosystem response to hydroclimatic conditions may be inferred from the analyses of temporal changes in RUE. For example, when an ecosystem is degraded and species loss happens, its productivity declines, which has severe implications for the functioning of this ecosystem and is reflected by a decrease in RUE (Chapin et al., 2000;Fensholt et al., 2013). Extreme droughts may also cause a decrease in RUE, suggesting that the ecosystem coping mechanism has reached its limit (Ponce-Campos et al., 2013) and that the ecosystem productivity has been impacted significantly (Horion et al., , 2019. The detection of major breakpoints in RUE time series indicates that the ecosystem response to precipitation has changed significantly through time and, ultimately, enables in-depth analysis of TPs in ecosystem functioning in both anthropogenic and natural landscapes .
A first study mapping TPs in drylands was made at the regional scale , but no global-scale analysis has been attempted so far. Thus, the general objective of this study was to provide a general synthesis of changes in the ecosystem functioning of drylands at the global scale by mapping and characterizing TPs. We focused on three specific research questions: 1. Which global dryland regions were hotspots of altered ecosystem functioning during recent decades?
arid regions, ecosystem functioning, human-natural system, land degradation, rain-use efficiency, remote sensing, segmented trend analysis, semi-arid regions, turning point 2. What types of major changes in functioning occurred in those regions, and when did the TP take place?
3. How may human-induced pressure and climate extremes affect the occurrence of TPs?
Considering that abrupt changes in RUE indicate TPs in ecosystem functioning, we assumed that areas with a high spatial concentration of TPs represented hotspot regions of altered functioning.
Here, we used breaks for additive season and trend (BFAST;de Jong et al., 2012de Jong et al., , 2013Verbesselt, Hyndman, Newnham, & Culvenor, 2010), which is a time series segmentation technique able to asses gradual and abrupt changes in RUE. Moreover, explicit information on the rate of change in RUE would allow us to gain a better understanding of changes in functioning over time. Thus, we developed a new typology for categorizing TPs, optimized for characterizing the behaviour of the ecosystem before and after TPs and also taking into account alterations in the rate of change. Lastly, ancillary data on population density and climate extremes were used to provide insights into the determinants of hotspots of change.

| Study area
Drylands encompass regions where the production of crops, forage and other ecosystem services are limited by water (Millennium Ecosystem Assessment, 2005). Köppen's classification (Peel, Finlayson, & McMahon, 2007) was used to define our study area, encompassing drylands with arid and semi-arid climates. Although hyper-arid and dry sub-humid regions might also be considered drylands, these were not included in our analysis. Hyper-arid regions are mainly not vegetated, whereas in dry sub-humid regions precipitation might not be the dominant limiting factor (Nemani et al., 2003), making RUE an inadequate indicator of ecosystem functioning .

| Dataset on ecosystem functioning of global drylands
To characterize temporal changes in ecosystem functioning (expressed here as changes in RUE), we used long-term time series of remotely sensed estimations of NDVI and rainfall. The NDVI, the normalized ratio between red and near-infrared reflectance, derived from satellite imagery is a widely used proxy for vegetation density and/or health (Tucker, 1979). Here, we used the third-generation Global Inventory Modeling and Mapping Studies (GIMMS) NDVI from Advanced Very High Resolution Radiometer (AVHRR) sensors version 1 (NDVI3g.v1) dataset , which provides bi-monthly data at 0.083° spatial resolution ranging from July 1981 to December 2015. A median NDVI value of .1 was used as the threshold to remove areas with sparse/absent vegetation (de Jong et al., 2013). Moreover, areas covered by snow and deserts were excluded using the Moderate Resolution Imaging Spectroradiometer Land Cover product (MCD12C1) for 2012, at 0.05° spatial resolution (Friedl & Sulla-Menashe, 2015). Precipitation data were obtained from the Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) dataset, which provides monthly data spanning from 1981 to 2015 at 0.05° resolution (Funk et al., 2015). We resampled the CHIRPS data to match the spatial resolution of the NDVI3g.v1 dataset using the nearest neighbor method.
Rain-use efficiency is derived by dividing NPP by precipitation.
Here, we used the growing season NDVI small integral values, which represent the sum of the "greenness" during the growing season and can therefore be related directly to vegetation productivity, as a proxy of NPP (Fensholt et al., 2013). The NDVI small integral values in drylands usually range from 0.029 to 1.361 (10th and 90th percentiles, respectively, of the average NDVI small integral values between 1982 and 2015). They were extracted for each year between 1982 and 2015 using TIMESAT (Jönsson & Eklundh, 2004), using the same settings as the ones specified by Horion et al. (2016) (Supporting Information   Table S1). Areas with double growing seasons were included by considering two separate NDVI small integral values for each year.
For estimates of precipitation, we used the sum of precipitation data during the growing season months (as determined by the NDVI analysis). As dryland vegetation may respond to precipitation with variable time-scales (Horion, Cornet, Erpicum, & Tychon, 2013;Vicente-Serrano et al., 2013), multiple time spans for computing the precipitation sum were tested. Specifically, we calculated the precipitation sum over n months before the start of the growing season until the end of the growing season, with n ranging from zero to 12.
For each pixel, the time span that showed the highest correlation between the NDVI integral and the precipitation sum was selected as the optimal time span. Optimal precipitation sums were then estimated per pixel and RUE time series generated. One of the basic assumptions when interpreting temporal changes in RUE as being non-climate related is that the RUE time series is statistically independent of the precipitation inputs (Fensholt et al., 2013). To test this assumption, we regressed (pixel-wise) the RUE data against precipitation sums. Turning points detected in pixels that showed a significant relationship (p < .05) indicated that driver attribution should be interpreted with caution.

| Characterization of changes in ecosystem functioning
To detect and characterize changes in ecosystem functioning, BFAST was applied on RUE time series. BFAST decomposes time series into trend, seasonal and remainder components using a piecewise linear trend and seasonal model, and detects abrupt changes in the trend and seasonal components (Verbesselt et al., 2010). More specifically, BFAST01, a variant of BFAST that allows detection of the major breakpoint (none or one) was implemented (de Jong et al., 2013). The algorithm not only allows the detection of abrupt changes in functioning (i.e., TPs) but also their characterization. The formula argument of BFAST01 was set to "response ~ trend", because we used yearly time series with no seasonality, and several different bandwidth values were used in an incremental fashion, as explained better in the "Confidence of the detected TPs" subsection. Default settings were used for the remaining arguments. The fitted linear trend can be classified, taking into account the existence of an abrupt change in the time series and the slope direction of the detected trends (de Jong et al., 2013).
Application of this classification can already provide information about changes in ecosystem dynamics. Nevertheless, it does not allow (a) the identification of regions where the rate of change varied over time; and (b) ensuring that the trend in functioning of a system has shifted significantly to the opposite direction. Here, we propose an improved typology for BFAST01, which also considers differences in the steepness and significance of the trends before and after the detected breakpoint. This allows for an improved qualification of the rate of change in the ecosystem (i.e., slowing down or accelerating) and the status of shifts in the trend (i.e., transitional or complete).
As a result, the improved typology consists of six types describing the overall trends and trend shifts and four subtypes providing a more detailed view on the rate and significance of changes (Table 1).
Detailed information about the new typology can be found in the Supporting Information (Appendix S1). All analyses were implemented in the R environment (R Core Team, 2018), and the script with the new typology function is already available (see "Data accessibility" section).

| Confidence of the detected turning points
In order to identify regions with high/low confidence in the detected TPs, we developed the turning point occurrence index (TPOI). The bandwidth parameter of the BFAST01 function determines the size of the data window, relative to the sample size, that will be used by the moving sum of residuals test (de Jong et al., 2013). This means that changing the bandwidth value modifies the break detection capacity of the algorithm, meaning that certain values (i.e., intermediate values) will result in easier detection of breaks. We thus applied the BFAST01 to the RUE data with bandwidth values ranging from the minimum to the maximum allowed by the length of the time series (i.e., from 2/34 to 33/34). By summing, pixel-wise, the number of times a TP was detected for all possible bandwidth values and dividing it by the maximum possible number of detected breaks (32), we obtained the TPOI. A high TPOI value indicated that, even when the algorithm specifications made it harder for BFAST01 to detect a break, the TP was detected, meaning that there was a high chance that the TP truly occurred.

| Population density and drought versus turning point occurrence
Population pressure and climate extremes are known to be major drivers of changes in dryland ecosystem functioning (Cherlet et al., 2018). We, therefore, analysed the relationships between the occurrence of TPs, population density and drought incidence. This was achieved by calculating, for each study region, the percentage of Global drylands were separated into focal regions: North America, the Sahel, Asia and Australia. South America was not considered here owing to a lack of large-scale TP events. All pixels were divided into four population density and drought occurrence classes, from very low to very high density/occurrence (Table 2), and a cross-tabulation approach was used. For each region, population density quartiles were used to create classes of equal size (Table 2).
This improved the region-specific characterization of population density pressure over ecosystems, because population density varies greatly among regions. Classes of drought occurrence were assigned using the same rationale, with quartiles of drought frequency estimated for global drylands (and not focal regions) because the SPEI is normalized per pixel (Table 2). Lastly, the percentage of pixels that presented a TP was plotted for the combinations of population and drought classes, providing insights about the relationship between each driver (or combination of drivers) and TPs for each region. Moreover, to characterize the different combinations of population/drought classes further, we calculated the fraction of pixels assigned to major land-cover types (according to the MCD12C1) for each combination. This provided additional information to support the driver attribution analysis. Land-cover types that resulted in <1% cover for each of the possible combinations of population and drought classes were excluded.

| Case studies
Changes in ecosystem functioning were analysed further over four case studies. Within each focal region, we identified a hotspot area of changes and selected a case study based on its spatial extent and the type of TP detected. Results from the new typology were then analysed in scrutiny and based on literature, aiming to confirm the results of our categorization for these areas. Gradual land degradation, such as long-term loss of soil organic matter content (e.g., Horion et al., 2019) Interrupted increase Accelerating Shift in species composition towards a faster increase in ecosystem functioning (e.g., Chapin et al., 2000) Interrupted increase

| Hotspots of turning point occurrence
Slowing down Shift in species composition towards a slower increase in ecosystem functioning (e.g., Chapin et al., 2000) Interrupted decrease Accelerating Extreme drought leads to high RUE, but subsequent plant mortality decreases ecosystem productivity by reducing vegetation density (e.g., Lotsch et al., 2005) (Continues) | 1235 BERNARDINO Et Al.

Illustration of ecological inference
Positive reversal Transition Cropland abandonment, followed by natural regeneration of vegetation (e.g., Horion et al., 2016;present study) Negative reversal Complete Deforestation, leading to a decrease in ecosystem functioning in north-eastern Brazil Note: Examples based on rain-use efficiency (RUE) time series are presented, representing different types of major changes in ecosystem functioning. Examples of the potential ecological inference of the observed changes are given. In addition, the last time series, which shows a negative reversal, was extracted from a region that underwent intense deforestation in the 2000s in Brazil (state of Piauí).
The vegetation states before and after deforestation are presented as true colour satellite images in the Supporting Information ( Figure S4). The grey line represents the RUE time series and the blue line the fitted linear model. The dotted vertical line represents the year of the detected turning point (TP), with the red mark indicating the confidence interval for the estimated timing of the TP. The asterisks indicate that the fitted trend is significant (p < .05). The colours used to illustrate each type and subtype are identical to the colours used in the maps produced. Note that not all possible combinations of types and subtypes are presented here.

TA B L E 1 (Continued)
to precipitation) to detect and categorize TPs in the ecosystem functioning of drylands. The TPOI revealed regions of global drylands with a high (spatial) occurrence of TPs during recent decades and high confidence for the detected TPs (Figure 1). Those regions were centralwestern North America, the Sahel, north-western Central Asia and Australia. Hereafter, we refer to those regions as "focal regions".
It is noteworthy that, by considering the lagged vegetation response to precipitation in the RUE estimation, the number of detected TPs increased by 22.4% when compared with using the growing season precipitation sum without added months. However, some areas still showed a heavy influence of variability in precipitation on RUE data (Supporting Information Figure S1). revealed a dominance of accelerating rates of change regardless of the direction of change (i.e., 71.5% of the pixels classified as interrupted increase and 69.4% as interrupted decrease were assigned to the "accelerating" subtype). In addition, both positive and negative reversals were mostly assigned to the "transition" subtype (81.5% and 75.2%, respectively).

| Characterization of the detected turning points
The relative occurrence of each ecosystem change type and subtype varied greatly between focal regions (Figures 2 and 3a). In North America, the most common TPs were interrupted decrease with acceleration after the TP (23.31% of the pixels with a TP) and negative reversal in a transitional state (27.2%; Figure 3a). On the contrary, the most common pattern of change in the Sahel was positive reversal in RUE (53.72%; Figure 3a). The same predominance was observed in Asia and Australia (40.35% and 47.51%, respectively; Figure 3a). Turning points were detected between 1987 and 2010, with a high frequency of TPs in 1992 observed in the Sahel and Australia (Figures 2 and 3b). Given that the structural change test performed by BFAST01 does not allow for the detection of TPs close to the start or end of the time series (de Jong et al., 2013), no TPs were detected before 1987 or after 2010.

F I G U R E 2
Type, subtype and time of occurrence of turning points (TPs) in ecosystem functioning over the four focal regions, which are North America (a-c), the Sahel (d-f), Central Asia (g-i) and Australia (j-l). Areas used as case studies are highlighted with black rectangles and represented in greater detail in Figure 6 [Colour figure can be viewed at wileyonlinelibrary.com]

| Population density and droughts versus turning point occurrence
In North America, higher drought pressure was related to high TP occurrence (Figure 4a). In addition, TPs did not occur more frequently over a specific land-cover type, because North American drylands present a shared dominance of grasslands and shrublands (Figures 4a and 5a-c). In the Sahel, the highest percentages of TPs also occurred under high to very high drought pressure (Figure 4b).
In addition, fewer TPs were observed where more croplands were present and drought was less frequent (Figures 4b and 5f). In Asia, a particularly high percentage of TPs (34.07%) was observed where population pressure was very high and drought occurrence very low ( Figure 4c). This combination of classes was dominated by croplands

| Case studies
The four selected sites were used to evaluate the new proposed typology in more details (Figure 6; Supporting Information Figure   S3). The sites were located in western Texas (USA; Figure 6a-

| Improved reading of abrupt changes in global drylands
We have provided here the first global-scale assessment of TPs in ecosystem functioning of drylands. Hotspots of altered functioning were identified in central-western North America, eastern Sahel, north-western Central Asia and eastern Australia (Figures 1 and 2).
Our results point to 1992 as a key year for observed abrupt changes in Earth observation data, as already reported by previous studies (de Jong et al., 2012;Tian et al., 2015), which identified the eruption of Mount Pinatubo in June 1991 as a potential driver for residual 2. Regarding the methods, we applied an improved typology of changes in ecosystem functioning, which also allowed for quantification of alterations in the rate of change before and after TPs (i.e., accelerating or decelerating trends; see Supporting Information Appendix S1). In addition, we accounted for the lag in the response of vegetation to precipitation before performing the TP detection analysis and used the BFAST01 algorithm in an iterative fashion with increasing bandwidth values in order to detect and assess the TPs. Those two steps substantially increased the number of detected TPs, potentially reducing the number of false negatives in our analysis. Such improvements represent F I G U R E 6 Sites selected as case studies for the description and interpretation, based on published studies, of the patterns detected by the typology here proposed. Grey lines are international/regional borders, and black lines delimit the site of the case study (locations of sites are indicated in Figure 2). Maps (a-c) show the case study in North America, located in Texas (USA); maps (d-f) are for the case in the Sahel, located in the southern part of Sudan; maps (g-i) show the case in Central Asia, located north-west of the Caspian Sea; and maps (j-l) represent the case in Australia, which is mostly within the Mulga Lands bioregion [Colour figure can be viewed at wileyonlinelibrary.com] important steps towards the comprehensive analysis of pathways of change in ecosystem functioning. For example, it was recently shown that local degradation processes (e.g., decreased vegetation productivity owing to soil erosion) would be signalled by a change in the magnitude of the slope, suggesting a slowing down in ecosystem functioning (Horion et al., 2019). Thanks to our new typology, acceleration and slowing down of ecological processes can now be captured and labelled in a systematic way.

| The new typology under scrutiny
Comprehensive mapping and categorization of changes in ecosystem proxies derived from Earth observation data are important steps towards an improved understanding of changing pressure over dryland ecosystems. Nonetheless, attribution of drivers remains a major challenge in global environmental studies (Rasmussen et al., 2016), notably owing to the diversity of driving forces and the limited number of reference sites suitable for validation (Cherlet et al., 2018). Examples of ecological inference for the types of ecosystem changes are given in Table 1. However, it is important to note that different drivers or combinations of drivers may result in the same type of ecosystem change (Horion et al., 2019). Further research is needed to deepen our knowledge on the variability of ecosystem pathways resulting from single or combined driving forces.
In the present study, we took the first step in that direction by selecting four case studies to assess our new typology further and to investigate potential ecological implications of the observed TPs. Negative reversals in RUE were observed in western Texas ( Figure 6a). This site, together with other areas of the Western High Plains, underwent a cycle of cropland abandonment (during the Conservation Reserve Program established in the 1980s) and re-cultivation or conversion to livestock grazing (Drummond et al., 2012;Sullivan et al., 2004). Such institutional change can lead to abrupt changes towards reduced functioning in areas that were previously preserved, that is, shift to a decreasing trend in RUE owing to loss of biodiversity after conversion from natural vegetation to cropland or pasture. The timing of the TPs detected here (i.e., 1992-1996) aligned with the period of re-cultivation (Drummond et al., 2012;Sullivan et al., 2004).
Interrupted decrease with acceleration (after 2004) was observed in the south of Sudan (Figure 6d-f). The driest year in the decade for this region was reported to be 2004 (Elagib, 2013), and it occurred after a prolonged drought period, which might have triggered an even sharper decrease in functioning in some areas (Elagib, 2013;Elagib & Elhag, 2011). Overall, drought decreases productivity of ecosystems by, for example, reducing vegetation density owing to drought-induced plant mortality (e.g., Lotsch, Friedl, Anderson, & Tucker, 2005; Table 1). However, it has been shown that maximum RUE is observed in dry years (Huxman et al., 2004). The large extent of the hotspot area suggests that the observed accelerated decrease in RUE after 2004 was attributable to significant damage to the ecosystem functioning caused by the long-term and extreme drought conditions (Ponce-Campos et al., 2013). This is also in line with our results (Figure 4b), which showed that drought relates to a large share of TPs in the Sahel.
In western Kazakhstan and adjacent Russian territory, there was a widespread occurrence of positive reversal in ecosystem functioning trends (Figure 6g), pointing to regions that underwent a degradation process in the past but reverted into a functioning recovery in recent years.  (Hölzel, Haub, Ingelfinger, Otte, & Pilipenko, 2002;Zonn, 1995). Also, inappropriate irrigation practices resulted in 1.03 million ha of land suffering from salinization (Zonn, 1995).
These factors might explain the observed decreasing trend in functioning during the first half of the observation period. After cropland abandonment, the original natural steppe vegetation was able to regenerate (Hölzel et al., 2002;Horion et al., 2016), which might have resulted in an increase in productivity (Table 1). The process is, however, not immediate, and signs of regeneration (here, an increase in RUE) might take more or less time to be detectable depending on the degree of preceding degradation (e.g., Castellano & Valone, 2007).
Finally, positive reversals in ecosystem functioning were observed over a site in the eastern Australian drylands (Figure 6j). This area covers a great part (c. 38.8%) of the Mulga Lands. Although the region is not suitable for agriculture, it has typically been used for livestock grazing (Witt, Noël, Bird, Beeton, & Menzies, 2011). As grazing has been persistent in the region since the second half of the 19th century (Witt, Harrington, & Page, 2009), other factors are likely to be responsible for the observed reversal in the direction of trends.
It is well known that drought has had a major impact on Australian ecosystems in recent decades (Rice et al., 2004). Specifically here, increased rainfall was observed at the end of the 1980s, followed by a severe El Niño-induced drought between 1992 and c. 1997 (Rice et al., 2004;Witt et al., 2009). This is in line with our results showing an important share of TPs occurring during these extremely dry years. After 1997, rainfall levels remained above average for several years, which might explain the trend reversal to an increase in functioning observed over the area after the extremely dry years. Indeed, when water availability is not restricting vegetation growth and fire activity (another constraint for vegetation growth in the region) is controlled by fuel removal owing to grazing, the ecosystem is capable of recovering towards an improved functioning (Witt et al., 2009(Witt et al., , 2011. Such large-scale mapping of areas with altered functioning is of great interest for decision-making related to mitigation of and adaptation to land degradation in drylands (Millennium Ecosystem Assessment, 2005;Scholes et al., 2018). The characterization of the ongoing processes in those areas (e.g., land degradation or restoration) can possibly help to set priorities when taking measures for ecosystem conservation. Although developed to characterize changes in dryland ecosystems, the new typology might also be applicable in different ecosystems worldwide, with the pre-condition of applying it to a proxy that reflects the functioning of the ecosystem under study (Horion et al., 2019). However, adequate validation with case studies is still required to ensure the validity of the approach for other ecosystems, because the interpretation of the results might differ from those inferred from our study in dryland ecosystems.

| Conclusion
By implementing the new BFAST01 typology, which allowed for an Zeileis for support with the "bfast" package in R. The authors have no conflict of interests to declare. This work was carried out on the Dutch national e-infrastructure with the support of SURF Cooperative.

DATA ACCE SS I B I LIT Y
The R code (R Core Team, 2018)