Vegetation dynamic trends and the main drivers detected using the ensemble empirical mode decomposition method in East Africa
Abstract
Understanding vegetation changes and the underlying causes are critical to predict the future ecosystem dynamics. Water-limited ecosystems in East Africa have been identified as particularly vulnerable to the impacts of increased environmental change, but relatively little is known about the vegetation dynamic trends and the main drivers. Taking the satellite-derived normalized difference vegetation index (NDVI) as a proxy of vegetation growth, we detected the spatiotemporal evolution of NDVI trends and explored the main drivers based on the ensemble empirical mode decomposition method. Results showed that greening (restoration) and browning (degradation) coexisted in East Africa during 1982–2013. However, the early greening stalled or even reversed as the browning continuously intensified over time. Greening-to-browning prevailed in East Africa after 2000. Generally, more than half of the study area (54.77%, 13.77% with a significant decrease) experienced a browning trend during 1982–2013, with an average change rate of (−0.12 ± 1.10) × 10−3. We also found that the percentage of greening area in each land cover type overall decreased. Furthermore, the remarkable spatiotemporal correlations between NDVI and soil water suggested that water was a dominant factor governing the vegetation changes in East Africa. The effectiveness of different types of protected areas (PAs) on vegetation change varied, depending to some extent on management policy (legal laws and on-site patrols). Therefore, a big challenge for sustainable management of protected areas is how to balance conservation and development objectives in East Africa.
1 INTRODUCTION
Vegetation is a fundamental element of the terrestrial ecosystem (de Jong, Schaepman, Furrer, De Bruin, & Verburg, 2013), and is directly related to a variety of ecosystem services, such as water purification, soil retention, and carbon sequestration. Vegetation also alters the exchange of water vapor, energy, and momentum between the land surface and the atmosphere through evapotranspiration, surface albedo, and roughness (Wang et al., 2011). Changes in vegetation activity are commonly considered to be indicators of environmental change. Understanding changes in vegetation growth and their responses to environmental variables is thus a crucial requirement for projecting future vegetation dynamics and ecosystem evolution (Wen, Wu, Chen, & Lu, 2017). In the context of global environmental change, research on vegetation dynamics has become a heated topic (Gang et al., 2014). Many previous studies focused on vegetation dynamics within the Northern Hemisphere or on a global scale (de Jong, Verbesselt, Zeileis, & Schaepman, 2013; Lu et al., 2015; Wang et al., 2011; Zhu et al., 2016); however, relatively little is known about the vegetation dynamic trends and underlying causes in East Africa.
Water-limited ecosystems in East Africa have been identified as particularly vulnerable to the impacts of increased environmental change (Hawinkel et al., 2016). Many protected areas (PAs) have been established for biodiversity conservation and ecosystem management in this region (Figure 1). In recent decades, East Africa has been suffering from rising temperatures and increasing droughts, which has intensified in the 21st century (Busby, Smith, & Krishnan, 2014). Some studies reported that East Africa is one of the most notable regions with vegetation productivity decline and land degradation (de Jong, Verbesselt, et al., 2013; Landmann & Dubovyk, 2014). However, the extent and intensity of such vegetation degradation is still not clear. Furthermore, the vegetation responses to environmental change vary temporally and spatially. Therefore, we need to characterize the spatiotemporal evolution of vegetation growth and identify the main drivers in East Africa, which is a requisite for projecting future ecosystem dynamics, identifying vulnerable areas, and implementing adaptation and mitigation measures (Hawinkel et al., 2016).
Previous studies on vegetation change generally adopt a linear fitting method. Merely by extracting a constant rate to represent the vegetation changes over a long time span, linear fitting could completely mask internal trend shifts and hinder the identification of underlying causes. In recent years, a few algorithms have been proposed to detect the time-varying trends of vegetation growth, including piecewise linear regression (Wang et al., 2011), polynomial fitting (Jamali, Seaquist, Eldundh, & Ardo, 2014), and empirical orthogonal function analysis (Park & Sohn, 2010). These algorithms require setting a priori function form and specifying key parameters subjectively to define the trend with implicit stationary assumptions, however, vegetation dynamics may contain nonstationary variability (Pinzon & Tucker, 2014). In this study, we adopted a noise-assisted ensemble empirical mode decomposition (EEMD) method derived from the signal-processing domain to capture the evolution of vegetation growth. EEMD could adaptively and intrinsically extract the secular trend R(t) of a time-series via removing all oscillatory components with decreasing frequency from raw data. The extracted secular trend, either monotonic or containing only one extremum, does not follow a priori shape and varies with time after the removal of intrinsic variability or the extension of new data, allowing the extracted trend to reveal more underlying information on the nonlinear and nonstationary time series (Ji, Wu, Huang, & Chassignet, 2014). Thus, EEMD has a wide range of applications in climatic research such as analyzing the evolution of land surface air temperature trends (Ji et al., 2014), phenological responses to warming (Guan, 2014), interannual vegetation dynamics (Hawinkel et al., 2016), and nonlinear variations in forest LAI (Yin et al., 2017).
In this study, we used the satellite-derived normalized difference vegetation index (NDVI) as a proxy of vegetation growth. The main objective was to investigate the regional-scale vegetation changes and underlying causes in East Africa during 1982–2013. To achieve this objective, we first characterized the spatiotemporal dynamics of NDVI trends, including instantaneous trend, accumulated trend, and turning point (TP) at which the NDVI trends reversed using EEMD method. Next, we investigated the evolution of NDVI trends at the land cover scale. We then analyzed changes in hydrometeorological variables and their relationships with changes in NDVI. Finally, we studied the comparative effectiveness of different types of PAs on vegetation change, and offered practical advice on the management of PAs.
2 MATERIALS AND METHODS
2.1 Study area
Our study area included Kenya, Tanzania, and Rwanda in East Africa (Figure 1), where climatic elements exhibited spatiotemporal heterogeneity. A large area of East Africa is either arid or semiarid and can be characterized by a bimodal precipitation pattern. The main rainy season occurs from March to May and the shorter rainy periods occur in October and November. Seasonality in East Africa is controlled primarily by the Intertropical Convergence Zone (Anyah & Semazzi, 2007). The interannual fluctuation of rainfall is strongly related to the El Niño Southern Oscillation that often tends to increase rainfall (Wolff et al., 2011). The vegetation types from the northeast to the southwest are in turn shrubland, grassland, and savanna, which is approximately consistent with the gradient of rainfall. Though East Africa is rich in biodiversity, rapid environmental change along with intensive human activities have resulted in biodiversity decline and vegetation degradation.
2.2 Datasets
2.2.1 NDVI data
In this study, we used the Global Inventory Modeling and Mapping Studies (GIMMS NDVI3g.v0) biweekly NDVI dataset from 1982 to 2013 with a spatial resolution of 8 km (https://nex.nasa.gov/nex/projects/1349/). The GIMMS NDVI3g.v0 dataset has been well calibrated through a series of processing steps to reduce NDVI errors arising from cloud cover, intersensor calibration, view geometry, and volcanic aerosols (Wang et al., 2011). To match NDVI data with hydrometeorological data, we resampled NDVI to a resolution of 0.1° × 0.1° using the nearest neighbor method. The monthly NDVI was defined as the maximum value composite of two images each month to further reduce cloud and other noise effects. Areas with very sparse or no vegetation cover (yearly mean NDVI < 0.1) were masked out for each year (de Jong, Schaepman, et al., 2013). The wide spatial coverage and long-term observations by this NDVI dataset can enable trend analysis, and GIMMS NDVI3g.v0 has been widely used to detect vegetation growth change (Fensholt & Proud, 2012).
2.2.2 Hydrometeorological data
The Hydrometeorological data included monthly temperature, precipitation, and soil water (the average soil water content of the surface 120-cm layer). We obtained it from FLDAS Noah Land Surface Model L4 that contains a series of land surface parameters simulated from the Noah 3.3 model in the Famine Early Warning Systems Network Land Data Assimilation System (FLDAS). It is a gridded dataset with a spatial resolution of 0.1° and completely covers our study period (1982–2013; http://disc.gsfc.nasa.gov/datacollection/FLDAS_NOAH01_C_EA_M_001.html).
2.2.3 Land cover data
In this study, land cover data were derived from the MODIS MCD12C1 product, and we adopted the 17-class International Geosphere-Biosphere Programme land cover classification (https://e4ftl01.cr.usgs.gov/MOTA/MCD12C1.051/). To reduce the impacts of classification error and land cover change, only stable pixels—defined as areas with no change in the dominant land cover class during 2001–2012—were used in the land-cover-specific analysis (Friedl et al., 2010). Finally, six land cover classes were considered in this study: open shrublands (OSH), woody savannas (WSA), savannas (SAV), grasslands (GRA), crop/natural vegetation mosaic (CNV), and barren or sparsely vegetated (BSV; Figure 6). The land cover data were also resampled to 0.1° using the nearest neighbor method.
2.2.4 The distribution and boundary of the PAs
For PAs, we used the most recent version of the World Database of Protected Areas (https://www.protectedplanet.net/). We selected the main PA types in each country (Figure 1) and excluded from our analyses all marine PAs and all areas designated only by international conventions (i.e., not nationally gazetted), as well as all PAs that were private or governed by communities (Jenkins, Pimm, & Joppa, 2013). The main PA type in Rwanda was national park. In Kenya, the main PAs include forest reserves, national parks, and national reserves. Nature reserve, game reserve, forest reserve, conservation area, and national park are the main PA types in Tanzania (see Figure 1).
2.3 Methods
2.3.1 Ensemble empirical mode decomposition
(1)
The sifting process does not follow a priori shape and uses only the local extrema information. For more details on the basic EMD method, please refer to Huang et al. (1998). Although powerful, the basic EMD method is highly dependent on the locations and values of local extrema. This makes the decomposition results sensitive to noise, leading to a “mode mixing” problem. By iteratively adding a finite amount of Gaussian white noise to the series, applying EMD to the mixture, and averaging ensembles of the noise-added series, EEMD largely eliminates the mode mixing problem and thus improves the robustness of the decomposition results (Wu & Huang, 2009).
In this study, EEMD was conducted at the pixel level (0.1° × 0.1°) to extract the secular trend of the corresponding NDVI time series. Figure 2 shows a few temporal NDVI profiles and the IMFs using EEMD technique. Taking the computing time and the robustness of decomposition into consideration, the number of the added Gaussian white noise was set at 100, and the amplitude of these noises was set at 0.2 standard deviations of the raw data (Wu & Huang, 2009). The NDVI trends obtained from EEMD vary over time in each pixel, so we defined the instantaneous trend as an annual increment of the NDVI trend, that is, Trendinst(t) = R(t) − R(t − 1), (t > 1982) (Figure 2). The accumulated trend was defined as the increment of the NDVI trend compared with the reference time of 1982. Therefore, Trendaccu(t) = R(t) − R(1982) (Figure 2) facilitates the comparison of the EEMD trend with the corresponding linear trend.
2.3.2 Statistical significance test of EEMD
In this study, we generated 10,000 samples of white noise series of the same temporal data length (32), applying EEMD to each series to determine its secular trend and obtain the empirical probability density function of EEMD trends at any temporal location. To judge whether an EEMD trend of NDVI at a given temporal–spatial location was statistically significant, we (a) divided the EEMD trend of that spatial location by the standard deviation of the corresponding NDVI raw data; (b) calculated the 1.96 standard deviation value of the trends in the 10,000 white noise series; and (c) checked whether the trend value was beyond 1.96 standard deviation (95% confidence interval) of the noise EEMD trend probability density function at the given temporal location. If the trend value exceeded the 1.96 standard deviation level, the EEMD trend was considered statistically significant (Ji et al., 2014).
2.3.3 Classification of trends and identification of TPs
We classified the EEMD trends of NDVI into six classes according to the extremum points and the significance of the trends. If a NDVI trend was monotonically increasing or decreasing, it was assigned to the monotonic greening or browning class, respectively. Under the condition that EEMD trend had one maximum, it was assigned to the significant greening-to-browning class if at least one of its trends was significant, and was assigned to the nonsignificant greening-to-browning class if none of its trends were significant. Similarly, EEMD trend with one minimum was assigned to the significant browning-to-greening and nonsignificant browning-to-greening class depending on whether or not it had significant trends. The transition year was called the TP (Figure 2) when NDVI trend changed from greening to browning or from browning to greening.
2.3.4 Comparative effectiveness analysis on PAs
The status of vegetation change categories (significant greening, browning, and no change) can be a reference to estimate the comparative effectiveness of PAs. The total comparative effectiveness of PAs was defined as ei = ei1 − ei2, where i referred to different PA types, eij = Pij/Nj [j = 1 (greening), j = 2 (browning), and j = 3 (no change)], within which Nj and Pij were the area percentage of the above three vegetation change trends, in the whole study area and in different PA types, respectively (Lu et al., 2015).
3 RESULTS
3.1 The spatiotemporal pattern of vegetation trends
The instantaneous rates of NDVI trends are shown in Figure 3. Before 1988, greening accounted for a large proportion of the vegetated area (62.17%, 11.46% with a significant increase), which mainly located in Kenya and northeastern Tanzania. The browning areas were interspersed in the northwestern corner of Kenya and northwestern Tanzania. By 2003, the spatial pattern of the instantaneous rate had almost reversed: browning was dominant in most of the vegetated area (51.56%, 8.15% with a significant decrease) except for certain minor areas in western Kenya and southeastern Tanzania that experienced a greening trend. The average browning rate in 2013 was (−1.30 ± 3.20) × 10−3, nearly double the average greening rate of (0.66 ± 2.10) × 10−3 in 1982.
The spatial evolution of the accumulated NDVI trends over time is shown in Figure 4. Before 2003, the spatial patterns of NDVI trends changed little: greening occupied most of Kenya and eastern Tanzania, whereas browning occurred mainly in the northern corner of Kenya and the south of Lake Victoria. However, the greening trend stalled or even reversed in most of Kenya after 2003, and the early sporadic browning almost expanded to the whole regions except western Kenya and southeastern Tanzania (Figure 4e,f). The evolution of NDVI trends in each decade also indicated that the early greening stalled or even reversed as the browning trend continuously intensified over time (Figure 5). Overall, more than half of the area (54.77%, 13.77% with a significant decrease) experienced a browning trend during 1982–2013, with an average rate of (−0.12 ± 1.10) × 10−3.
EEMD trends were classified into six types and the spatial pattern is shown in Figure 5d. Greening-to-browning was the most common category, accounting for 49.69% of the vegetated area (20.74% with a significant shift), mainly distributed in central and eastern Kenya, northeastern Tanzania, and the western and southern border of Tanzania. The TP of significant greening-to-browning primarily appeared after the turn of the century (70%; Figure 5e). By contrast, browning-to-greening (20.76%, 7.65% with a significant shift) was often spatially linked with monotonic greening (Figure 5d) and mainly occurred in Rwanda, the northwest of Kenya and the southeast of Tanzania. Significant browning-to-greening mainly appeared after 2000 (77.73%; Figure 5f). Monotonic browning was mainly located in the south of Lake Victoria and in the northern corner of Kenya.
3.2 Evolution of NDVI trends at a land cover scale
For most land cover classes (except for BSV), greening-to-browning accounted for the highest percentage (Figure 6a), as a result of the shift from early greening to later browning. For BSV, more than 70% experienced a browning trend over time (Figure 6), and 47.22% underwent monotonic browning; this percentage was much higher than that of other land cover classes. Moreover, the percentage of greening area in each land cover type overall decreased, although the temporal evolution processes varied substantially (Figure 6b). Before the early 1990s, more than 60% of the study area exhibited a greening trend and this proportion remained relatively stable for all land cover types except SAV (52%) and BSV (18%). After 2000, large expansion of a browning trend occurred in all land cover types except CNV. For CNV, more than 50% of the land cover was dominated by greening and 27.15% underwent monotonic greening (Figure 6).
3.3 Correlations between NDVI and hydrometeorological variables
Before the TP, the spatial patterns of precipitation and soil water trends (Figure 7b,c) were somewhat similar to those in NDVI; however, the spatial pattern of temperature trend had no such high similarity (Figure 7d). A large area (69.46%, 18.36% with a significant increase), mainly in central and eastern Kenya and northeastern Tanzania, experienced a greening trend before the TP. At the same time, an increase in soil water (47.1%, 18.98% with a significant increase) was most evident in those above regions that experienced a greening trend. By contrast, northwestern Tanzania, Rwanda, and the western corner of Kenya experienced a drying trend in soil water before the TP, whereas a decrease or stalling trend was also observed for NDVI. The partial correlation analysis and scatterplot further indicated that NDVI has a significantly positive correlation (p < .05) with soil water in most areas of East Africa (51.78%). After the TP, the NDVI trends ingeniously stalled or reversed, almost mirroring the NDVI trends before the TP (Figure 7a,e). Similarly, the precipitation, soil water, and temperature trends also reversed to some extent after the TP, and the spatial patterns of rainfall and soil water roughly matched that of NDVI (Figure 7). Further, the negative correlation between NDVI and temperature (p < .05) was significant in most of Kenya, and insignificant in other places.
3.4 The effectiveness of PAs on vegetation change
To detect the effects of PAs on vegetation change trends, we compared the vegetation change trends of different PA types with those at the national level (Lu et al., 2015). Rwanda was mainly covered by CNV, where national parks mitigated both vegetation greening and browning, thus exerted marginal positive effects on vegetation change (total effectiveness = 0.064). In Kenya, the main PAs (forest reserve, national park, and national reserve) had negative comparative effectiveness. For these three types of PAs, WSA, OSH, and GRA were the main land cover types. As shown in Figure 6, a large expansion of the browning trend occurred and the TP from greening to browning appeared in the 1990s. Thus, all types of PAs had an especially negative effect on vegetation instead of the expected vegetation protection effects. In Tanzania, the main land cover types within PAs were WSA and SAV. The effectiveness of PAs varied, partly depending on management policy. Nature reserve was most effective at facilitating vegetation greening (total effectiveness = 1.42), whereas game reserve and national park were effective at curbing vegetation browning. The effectiveness of forest reserves and conservation areas were negative, with the forest reserves being least effective at curbing vegetation browning (Table 1).
| Proportion of country in the study area (%) | Type of PAs | Proportion of PAs in the country (%) | Greening (e1) | Browning (e2) | No change (e3) | Total effectiveness (ei) |
|---|---|---|---|---|---|---|
| Rwanda (1.61) | National park | 8.76 | 0.5943 | 0.5304 | 1.1838 | 0.0640 |
| Kenya (37.29) | Forest reserve | 2.51 | 0.3550 | 0.3655 | 1.2644 | −0.0105 |
| National park | 4.67 | 0.0843 | 0.6774 | 1.2320 | −0.5931 | |
| National reserve | 2.56 | 0.2901 | 0.5296 | 1.2345 | −0.2394 | |
| Tanzania (61.10) | Game reserve | 12.51 | 0.7882 | 0.5078 | 1.1572 | 0.2804 |
| Nature reserve | 0.21 | 1.4198 | 0 | 1.1783 | 1.4198 | |
| National park | 5.84 | 0.4858 | 0.2467 | 1.2722 | 0.2392 | |
| Forest reserve | 9.68 | 0.1981 | 1.1096 | 1.1059 | −0.9115 | |
| Conservation area | 0.87 | 0.4840 | 0.7127 | 1.1569 | −0.2287 |
- Note. PA = protected area.
4 DISCUSSION
4.1 The early greening stalled or even reversed
The instantaneous rates of EEMD trends (Figure 3) clearly showed a reversal from greening to browning. For example, the average browning rate in 2013 was (−1.30 ± 3.20) × 10−3, nearly double the greening rate of (0.66 ± 2.10) × 10−3 in 1982. The browning area continuously expanded and the amplitude of browning intensified over time, especially during 2003–2013, leading to the turn-around from the previous greening (Figure 3). As the sum of change from the reference time of 1982, the accumulated trend also showed that the greening stalled or reversed during 2003–2013. From 1982 to 2013, more than half of the study area (54.77%, 13.77% with a significant decrease) experienced a browning trend, with an average rate of (−0.12 ± 1.10) × 10−3, demonstrating that the early greening was offset by later browning.
Additionally, most areas experienced a reversal of vegetation trends (71.45%, 28.4% significantly reversed). The NDVI trends over the entire period resulted from the offset (or reinforce) of trends before and after the TP, confirming that vegetation variation was not linear, but instead, complex and volatile (Park et al., 2015). The TP year of NDVI trends varied across different regions but mainly appeared after the turn of the century (Figure 5e,f), indicating that greening-to-browning prevailed after 2000. Likewise, de Jong, Verbesselt, et al. (2013) reported that signs of greening-to-browning reversals around the millennium transition had been found in many regions, including East Africa.
In addition to EEMD, we used simple linear fitting based on least squares to quantify the vegetation trends. The amplitude and spatial patterns of the accumulated NDVI trends during 1982–2013 using EEMD were slightly different from those using the linear fitting method. However, merely by extracting an average rate over a long time span, the linear method was so oversimplified that it could not reflect how the NDVI trends evolved and reveal the underlying information of changes in trends (Hawinkel et al., 2015; Ji et al., 2014). EEMD could break down an NDVI series into their characteristic time scales and extract the secular NDVI trend through a sifting process, which offered a potentially viable method for nonlinear and nonstationary data analysis (Huang et al., 1998). By removing noises and short-term fluctuations in advance, EEMD method was quite robust when there were noises and disturbances at the interannual scale (Hawinkel et al., 2015). As for any processing tool, the quality of the output is inevitably limited by the quality of the input data. In the case of the EEMD decomposition of NDVI time series, the results may contain uncertainties introduced by the datasets used. Although a series of data processing steps were taken to enhance data accuracy, there may have been errors in GIMMS NDVI3g.v0 data due to the noise effects of cloud cover, intersensor, orbital drift, and other factors that cannot be completely eliminated, especially in extremely cloudy and mountainous areas.
4.2 Possible factors for vegetation change trends
A large area (69.46%, 18.36% with a significant increase), mainly in central and eastern Kenya and northeastern Tanzania, experienced greening before the TP. At the same time, wetting and cooling trends were observed in those greening regions. However, the hydrometeorological trends in those above regions did not last. Instead, the drying and warming trends appeared after the TP, decreasing the water available for vegetation growth and increasing the water consumption by evapotranspiration, which led to water deficits (Hawinkel et al., 2016). Considering such trends, it was not a surprise that these regions experienced a reversal from early greening to later browning (Figure 7). In contrast, a decrease or stalling of NDVI was observed in western Kenya, where drying and warming trends were observed before the TP. After the TP, the NDVI trends over western Kenya also reversed, roughly mirroring the soil water trends. Overall, the spatial patterns of precipitation and soil water trends were somewhat similar to those of NDVI. However, the spatial patterns of trends in temperature did not have such high similarity (Figure 7). Further, the correlation between NDVI and soil water was found to be stronger than that between NDVI and precipitation (Figure 8), probably resulting from the lag effects of precipitation on vegetation growth (Ji & Peters, 2005).
From 1982 to 2013, the spatiotemporal evolution of trends in soil water roughly matched that in NDVI. Also, a significantly positive correlation (p < .05) between NDVI and soil water was observed in most of East Africa (51.78%; Figure 8), implying that water was a dominant factor governing the vegetation change in East Africa. Similar results have been reported that mean annual precipitation explained most variability in vegetation response across ecological zones in East Africa (Hawinkel et al., 2016). Additionally, vegetation dynamics could be affected by human activities, including ecological protection and deforestation. In this study, the opposing effectiveness of forest reserves and game reserves explained the changes in NDVI trends in southeastern and northwestern Tanzania (Figure 7 & Table 1), which partly depended on whether or not the PA was completely managed (protection policy and human disturbance; Landmann & Dubovyk, 2014). Additionally, the general climate conditions in most areas of East Africa changed from vegetation growth-conducive to growth-hostile at the turn of the 21st century (Dai, 2011; Neelin, Münnich, Su, Meyerson, & Holloway, 2006). Thus, it was not a surprise that greening-to-browning prevailed after 2000 (Zhou et al., 2014). Large-scale and long-term increase of browning in East Africa may provide an early warning signal for ecosystems degradation under global environmental change (de Jong, Verbesselt, et al., 2013). Attentions must be paid to these browning regions, referred to as hotspots of vegetation degradation. Plausible adaptation strategies should be employed in venerable areas to cope with climate change (Lu et al., 2015). Long-term ground observations are also needed to contribute to the knowledge of the ecosystem evolution and the driving mechanisms under future climate scenarios of increased variability.
4.3 Effects of management policy on PA's effectiveness
In recent decades, many PAs have been established in East Africa with the objective of conserving biodiversity and managing natural ecosystems (Figure 1). PAs accounted for 29.11% of Tanzania, where forest reserves were explicitly designed for forest protection. However, forest reserves did not perform better than lands without protection (Leverington, Costa, Pavese, Lisle, & Hockings, 2010). Generally, forest reserves (commonly gazetted as multiresource use area) were often located in areas with valuable timber stocks and used for extractive forestry. Incomplete management (devoid of on-site patrols) and wood-fuel consumption led to forest loss in forest reserves (Pfeifer et al., 2012). In contrast, national parks, game reserves, and nature reserves in Tanzania benefitted from complete management practices and strengthened legal powers that were shown to adapt vegetation browning (Table 1).
Generally, PAs with complete protection (guard forces and laws sanctions) were effective in curbing vegetation browning (Caro, Gardner, Stoner, Fitzherbert, & Davenport, 2009; Pelkey, Stoner, & Caro, 2000). But those measures restricted the access of local people to several ecosystem services (mostly provisioning, such as gathering, hunting, and wood collecting). Hence, a big challenge for the sustainable management of PAs in East Africa is to balance conservation and development objectives. Citizen science (Steger & Butt, 2015) and the Continual Engagement Model (Reid et al., 2016) may provide guidance on the allocation of natural resources and PAs management in a social-ecologically sustainable way.
5 CONCLUSIONS
EEMD presents new insights into the nonlinear dynamics of vegetation growth. During the study period (1982–2013), greening (restoration) and browning (degradation) coexisted in East Africa. Generally, more than half of the study area (54.77%, 13.77% with a significant decrease) experienced a browning trend during 1982–2013, with an average rate of (−0.12 ± 1.10) × 10−3. The percentages of greening area in each land cover type overall decreased. The early greening stalled or even reversed as the browning trend continuously intensified over time. Greening-to-browning prevailed in East Africa after 2000. Additionally, the remarkable spatiotemporal correlations between NDVI and soil water suggested that water was a dominant factor governing the vegetation changes in East Africa. Furthermore, the effectiveness of PAs on vegetation change varied, depending to some extent on management policy (legal laws and on-site patrols).
ACKNOWLEDGMENTS
This work was funded by the National Key Research and Development Program of China (2017YFA0604701) and the National Natural Science Foundation of China (31361140360, 41761144064).
