SEARCH

SEARCH BY CITATION

Keywords:

  • Bayesian dynamic models;
  • browning;
  • climate change;
  • greening;
  • Normalized Difference Vegetation Index;
  • phenology;
  • tropical mountains

Abstract

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

Global climate change has emerged as a major driver of ecosystem change. Here, we present evidence for globally consistent responses in vegetation dynamics to recent climate change in the world's mountain ecosystems located in the pan-tropical belt (30°N–30°S). We analyzed decadal-scale trends and seasonal cycles of vegetation greenness using monthly time series of satellite greenness (Normalized Difference Vegetation Index) and climate data for the period 1982–2006 for 47 mountain protected areas in five biodiversity hotspots. The time series of annual maximum NDVI for each of five continental regions shows mild greening trends followed by reversal to stronger browning trends around the mid-1990s. During the same period we found increasing trends in temperature but only marginal change in precipitation. The amplitude of the annual greenness cycle increased with time, and was strongly associated with the observed increase in temperature amplitude. We applied dynamic models with time-dependent regression parameters to study the time evolution of NDVI–climate relationships. We found that the relationship between vegetation greenness and temperature weakened over time or was negative. Such loss of positive temperature sensitivity has been documented in other regions as a response to temperature-induced moisture stress. We also used dynamic models to extract the trends in vegetation greenness that remain after accounting for the effects of temperature and precipitation. We found residual browning and greening trends in all regions, which indicate that factors other than temperature and precipitation also influence vegetation dynamics. Browning rates became progressively weaker with increase in elevation as indicated by quantile regression models. Tropical mountain vegetation is considered sensitive to climatic changes, so these consistent vegetation responses across widespread regions indicate persistent global-scale effects of climate warming and associated moisture stresses.


Introduction

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

Anthropogenic climate change in recent decades has affected the characteristics of vegetation worldwide (Root et al., 2003; Parmesan, 2006). Local, regional, and global-scale assessments all show that plant species ranges (Parmesan & Yohe, 2003), vegetation phenology (Cleland et al., 2007), primary productivity (Nemani et al., 2003; Piao et al., 2005), biomass (Chen et al., 2010), and even vegetation–climate relationships (D'Arrigo et al., 2004) have been significantly altered. These climate change impacts are not restricted to the world's flora alone, and have had widespread impacts on faunal assemblages through species extinctions and altered community structure (Pounds et al., 1999, 2006).

Despite mounting evidence for the impacts of climate change on flora, fauna, and ecosystem function at all scales, many regions of the globe remain poorly studied. Furthermore, climatic changes vary with location (Hansen et al., 2006) and warming effects are altered by climatic phenomena such as the El Nino-Southern Oscillation (Anchukaitis & Evans, 2010), and regional hydrology (Shindell et al., 2006) and topography (Daly et al., 2010). For example, warming rates in tropical latitudes appear to have been lower than those at temperate or arctic latitudes (IPCC, 2007a, b). Correspondingly, biological impacts of climate warming such as shifts in species' geographic ranges, changes in seasonal phenology, and species extinctions have mostly been observed at mid- to high latitudes (Parmesan, 2007).

Similarly, mountain regions may have experienced greater warming rates at higher elevations and such elevation-dependent warming may be more important in tropical latitudes, but the evidence for such patterns remains equivocal (Rangwala & Miller, 2012). Climate models indicate that rising tropical sea surface temperature can enhance oceanic evaporation and cause increased warming in tropical mountains, but empirical evidence is lacking. Early observations of increases in freezing heights (Diaz & Graham, 1996), the lifting cloud base (Still et al., 1999), and temperature-induced moisture stress (Pounds et al., 1999) are now matched by recent observations on the retreat of alpine glaciers (Bradley et al., 2009). However, direct investigations on changes in climatic patterns and biological impacts in tropical mountains have rarely been undertaken.

The decline of biodiversity in species-rich mountain ecosystems (Körner, 2004) (Spehn et al., 2002) could have adverse impacts on ecosystem function and decrease ecosystem services on which millions of people are dependent (Viviroli et al., 2007). Mountain ecosystems are thus of high priority in global conservation strategies (Brooks et al., 2006). Because they are often isolated and strongly structured by climatic gradients, mountain ecosystems, particularly those in tropical latitudes, are considered ecologically sensitive and vulnerable to climate change (La Sorte & Jetz, 2010). Although the empirical evidence remains scanty, modeling studies predict greater impacts on community structure in tropical mountains compared to those at temperate latitudes (Sheldon et al., 2011). Arguments from metabolic theory also suggest that climate change impacts may be greater in tropical ectotherms (Dillon et al., 2010). Furthermore, projections of future climate distributions show that tropical mountains are among the primary regions in the world with fast-disappearing climates (Williams et al., 2007). Collectively these studies imply that tropical mountain ecosystems are likely to experience considerable change in the near future.

In this study, we examine patterns and trends in greenness of vegetation in tropical mountain ecosystems in response to recent climatic change. To do this, we use the Advanced Very High Resolution Radiometer (AVHRR) 8-km NDVI (Normalized Difference Vegetation Index) data from 1982 to 2006 but focus only on vegetation in Protected Areas (PAs) located in the pantropical belt for the elevation range 1000–<6000 m above mean sea level (a.m.s.l.). We restrict our study to PAs to minimize the confounding effects of direct human activity on changes in vegetation. Our objective here was to examine vegetation greening trends and monthly cycles of greenness, and to test whether these patterns are correlated with corresponding trends in climate. We further test whether greening rates were elevation dependent as would be expected if warming rates were dependent on elevation. Our study also demonstrates the utility of Bayesian and state-space dynamic models in studying the time evolution of vegetation response to climate change. Such models have been used to study biogeophysical processes (Krishnaswamy et al., 2000, 2001) and monsoon dynamics (Maity & Nagesh Kumar, 2006), and their utility in analyzing NDVI-climate dynamics is demonstrated by this study.

Materials and methods

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

NDVI data

We derived the vegetation greenness time series for the period 1982–2006 using the Normalized Difference Vegetation Index (NDVI) obtained from 8 km resolution satellite data. These data were obtained from the Global Inventory, Monitoring, and Modeling Studies (GIMMS) database, distributed through the Global Land Cover Facility (http://www.landcover.org/). A global assessment of these NDVI data using 73 ground-control stations has reported strong correspondence with independent ground-based assessments of changes in land cover/land use (Sobrino & Julien, 2011). Furthermore, ground studies and data from independent satellite sensors have corroborated patterns and trends inferred from GIMMS NDVI data (Brown et al., 2006; Fensholt et al., 2009; Song et al., 2010; Berner et al., 2011; Luo et al., 2013). We confined our study sites to Protected Areas (PAs) with elevations >1000 m a.m.s.l. and identified 47 PAs located in five continental regions in Africa, Central America, South America, South Asia, and Southeast Asia (Fig. 1). We obtained a total of 801 NDVI pixels representing these study sites, which accounted for a total area of 51 264 km2 and were all located in World Biodiversity Hotspots (Table 1).

Table 1. Numbers of National Parks and NDVI Pixels sampled for each of the five continental regions
RegionNo. of National parks sampledNo. of NDVI pixelsTotal area (km2)
Central America5402560
South America1929018 560
Africa81137232
South Asia101559920
Southeast Asia520312 992
Total4780151 264
image

Figure 1. Locations of 47 higher elevation National Parks that occur in the pantropical latitudinal belt 30°N–30°S included in our study. The overall elevation range for all sites was 1000–5887 m above m.s.l.

Download figure to PowerPoint

Elevation data

The average elevation of each 8 km pixel was computed using a 30 m resolution Global ASTER Digital Elevation Model (http://www.ersdac.or.jp/GDEM/E/3.html). We then divided each mountain region into three elevation zones: Low (1000–2500 m), Medium (2500–4000 m), and High (4000–5887 m).

Climate and NDVI data

We obtained corresponding time series of monthly rainfall data at 1.0° resolution (ca. 110 × 110 km) from GLDAS, NASA, USA (http://disc.gsfc.nasa.gov/SSW/), and mean monthly temperature data at 0.5° (ca. 55 × 55 km) available through GHCN, Earth System Research Lab (ESRL), USA (http://www.esrl.noaa.gov/psd/). The higher resolution temperature data were appropriately combined and averaged to 1.0° resolution. We overlaid the 801 NDVI pixels on the 1.0° resolution climate pixels and obtained 55 distinct climate pixels covering all the 47 protected areas. These constitute the climate data obtained as monthly time series for years 1982–2006 for our study sites. We aggregated the 801 spatially explicit monthly NDVI time series into 55 distinct monthly NDVI time series by averaging across NDVI pixels that were coincidental in each of the 55 (coarser) climate pixels. This resulted in 11 distinct pixels (hereafter sites) for Africa, 5 for Central America, 18 for South America, 11 for South Asia, and 10 for Southeast Asia. These matched monthly NDVI and climate time series for 55 distinct sites distributed among five continental regions in the pantropical belt were analyzed individually for NDVI trends and NDVI–climate relationships. In presenting results, we averaged across sites within regions, with averaging done on parameter estimates and not on the data. Spatial variability within each region for the key NDVI and NDVI–climate parameters were quantified as standard errors.

Trend analyses

To measure the rate of monotonic change in univariate time series, we used the Sen slope (Sen, 1968), which is a robust nonparametric statistic of the rate of change, and is widely used in trend analyses. This is a rank-based regression technique in which the slope is calculated from a series with n independent and dependent observations as the median of all-possible (n (n-1)/2) pair-wise slopes. Trends in maximum annual NDVI and climate annual time series were analyzed using the Sen slope corrected for potential autocorrelation (Yue et al., 2002). For trend analyses using annual NDVI data, we first computed mean annual monthly NDVI time series for each of the five regions by averaging across NDVI pixels located in each region. We then computed the maximum annual NDVI time series for each region from these region-averaged monthly time series. Similarly, monthly climate averages for each region were also defined for trend analyses, while the individual spatially explicit time series in each region were used in NDVI-climate analyses (described below). We computed Sen slopes on annual climate time series by first using dynamic models (see below) to decompose the monthly time series into level and cyclic components. The annual medians were computed from the level component and Sen slope was estimated for these annual median time series. We tested elevation dependence of NDVI greening or browning rates by regressing the Sen slope of the full NDVI time series against elevation of the pixel. Here, we used data for all 801 AVHRR NDVI pixels and performed quantile regressions, for the 50th, 75th, and 95th percentiles of observed greening/browning rates against pixel elevation. While linear regression measures the conditional expectation (or mean) of the response variable as a function of observed values of the covariate(s), quantile regression measures all parts (quantiles) of the distribution of the response variable. This is a particularly useful regression technique when the variability in response is high at given values of the most important (measured) covariate in the presence of other unmeasured covariates or limiting factors (Cade & Noon, 2003).

NDVI–climate relationships

We first tested temporal change in NDVI–climate relationships by computing spatially explicit monthly time series of NDVI, temperature, and precipitation, and then carrying out dynamic linear regressions of NDVI as a function of temperature and precipitation for each of the 55 sites of our study area. In a static model, parameters do not change over time. A dynamic model on the other hand allows us to capture time-dependent variation in parameters such as the intercept and regression slopes, and cyclic components (Petris et al., 2009). The time-varying regression parameters track the changing response of NDVI to climate, and the intercept includes variability in NDVI that remains after climatic influences are accounted for. The dynamic models of monthly NDVI as a response to monthly temperature and precipitation yield estimates of parameters such as the intercept (NDVI level) and regression slopes with respect to temperature and precipitation that vary with time, and in which the distributions of these parameters are updated at each time step, conditional on the data available using Bayes' theorem. We also used another type of dynamic model for decomposing the monthly NDVI and climate time series into a dynamic level and a cyclic component from which the amplitude of the annual cycle for each year was derived. These two models were specified as follows:

Dynamic linear regression models

To model temporal variability in slope of the NDVI–climate relationship, we specified the dynamic regression model as: NDVIt = β0t + β1tTempt + β2t Precipt + et. This is called the observation equation, where β0t is the time-varying intercept and β1t and β2t are time-varying slope parameters, and et is the noise or error, with et ~ N(0,V) at time t. V is the observation variance, and is obtained from the residual variance from the ordinary static regression model, in which parameters are invariant with time. Next, system dynamics are modeled as: β0t = β0t−1 + w, β1t = β1t−1 + w, and β2t = β2t−1 + w, where w ~ N (0, W). These are called the system equations. A quantity W which controls the dynamics of the system parameters was specified a priori by setting W/V, the signal-to-noise ratio. As V is known, the ratio W/V specifies W. We set this ratio at a conservative value of 0.1. Starting from initial prior values obtained from an ordinary regression model summary (the β coefficients above), the regression parameters (say θ) perform a random walk about their previous states with uncertainty specified by W, with their probability distributions updated according to Bayes' theorem using data available (Yt) at time t−1: P(θt ∣ Yt) ∝ P(Yt ∣ θt)P(θt ∣ Y(t−1)). We focus on estimation of these parameters conditional on all the data being observed, by applying the Bayes' theorem in a recursive manner, estimating parameters at each time step retrospectively by conditioning on all data being observed. We estimated 55 such models for each distinct block of NDVI and climate time series, which yielded 55 monthly time series for all three parameters – regression coefficients for temperature and precipitation, and the intercept. We then computed the mean monthly time series and the standard error time series of these parameters for the five regions by averaging across blocks within each region. These time series constitute the final results for NDVI–climate relationships for each region.

Dynamic state-space model

Dynamic state-space models allow analyses of cyclic variation in a time series by resolving seasonal deviation from the dynamic level (or mean) of the time series. Our model was specified as NDVIt = Levelt + Cyclict + et, where et is the error or noise et ~N(0,V). A good estimate of the cyclic component requires one to estimate the time-varying level and vice versa. We computed the time series of the annual NDVI cycle for each of the 55 blocks using the StrucTS (Petris, 2010) function in R (R Development Core Team, 2011). Structural time series models are (linear Gaussian) state-space models for (univariate) time series based on a decomposition of the series into a number of components. A similar decomposition was done for the monthly temperature and precipitation time series to estimate time-varying level and time-varying amplitude of the annual cycle for each of the 55 sites. The time-varying monthly level time series was converted into an annual time series using medians and these were used for climate trend analyses (described above). Using these 55 time series of amplitudes for NDVI, temperature, and precipitation, we computed mean time series and standard error within each region at each time step by averaging across blocks within each region.

Results

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

We first examined trends in annual maximum NDVI values and found a globally consistent vegetation response in all mountain regions. We found mild increasing trends in NDVI (greening) until the mid-1990s that reversed thereafter and became a stronger decreasing trend in NDVI (browning) (Fig. 2). Segmented regression analyses reveal that the timing of trend reversal was different among regions, varying over a 7-year interval from the early- to the late-1990s (Fig. 2; Table 2). Using breakpoints obtained from segmented regressions, we divided each time series into greening and browning phases and computed Sen slopes for each phase. The Sen slopes were consistent across regions, showing positive values in the greening phase and pronounced negative values in the browning phase in all cases (Table 2). These greening and browning trends are quite distinct even when we consider spatial variability within each region (Fig. S1).

Table 2. Segmented regression analyses of annual maximum NDVI time series for each region. We computed a single annual NDVI time series for each region by averaging NDVI values across sites within a region. The breakpoints indicate number of years starting 1982 at which trend reversals were seen, and SE is the standard error of this estimate. Sen slopes of the greening and browning segments for each region and the associated P-values are also shown
RegionBreakpoint, SEGreening Sen slope, P-valueBrowning Sen slope, P-value
Central America20.0, 1.11.128, 0.09−21.720, 0.089
South America15.6, 1.72.861, 0.06−7.683, 0.035
Africa19.1, 2.10.780, 0.36−6.197, 0.086
South Asia13.0, 2.72.134, 0.54−4.324, 0.001
Southeast Asia13.9, 1.72.645, 0.13−8.091, 0.107
image

Figure 2. (a) Trends in annual maximum NDVI greenness of mountain vegetation over the period 1982–2006 for five continental regions located in pantropical latitudes. The time series of NDVI for each region was obtained by averaging across time series for all sites within each region. (b–f) Segmented regressions of annual maximum NDVI. Breakpoints and corresponding standard error bars are indicated on the abscissa in each panel.

Download figure to PowerPoint

We examined corresponding trends for temperature and precipitation during this time period and found that systematic trends were present mainly for temperature (Fig. 3). Thus, mean annual temperature showed significant warming in three regions (P < 0.05 for Central America, Africa, and South Asia) and significant cooling in one (P < 0.05, South America). In contrast to temperature trends, there were no discernible trends in precipitation, except for Africa, where a significantly negative trend was found (P < 0.05).

image

Figure 3. Trends in mean annual temperature (top row) and annual precipitation (bottom row) for five tropical mountain regions (a),(b),(c),(d),(e) for the period 1982–2006. Each series is constructed from median annual values, which were obtained from one average monthly time series per region, computed across all sites within each region. The monthly time series were first decomposed into level and cyclic components and monthly level component was averaged across sites for each region, from which the annual median value was computed. The Sen slope was then estimated for these annual time series. Sen slopes are indicated by straight lines where they are statistically significant (P < 0.05).

Download figure to PowerPoint

The correspondence between NDVI and climate is evident in the relationship between amplitudinal variation in NDVI and climatic parameters. Using a dynamic model that extracts the cyclic component in the NDVI and climate time series after removing trends in each, we found that the amplitude of the annual NDVI cycle and of temperature and precipitation increased with time in all pan-tropical mountains (Fig. 4). During the same period we found increasing trends in temperature but smaller change in precipitation. Multiple linear regressions of NDVI amplitude vs. temperature and precipitation amplitudes were highly significant for temperature in all regions (P < 0.01 for all), and significant for precipitation in three regions (P < 0.05 for Central America, South Asia, and Southeast Asia, Table S1) (R2 values > 0.8 in four regions and 0.54 in one). Furthermore, the amplitude for NDVI takes a dip after year 2000, which is coincidental with decrease in precipitation amplitude (Fig. 4). These very trends and patterns remain clearly evident even when spatial variability among sites in each region is quantified (Fig. S2).

image

Figure 4. Amplitude of seasonal NDVI cycles (a), and amplitudinal variation in mean monthly temperature (b) and monthly precipitation (c) in each year during the period 1982–2006. The amplitude for each year was calculated after decomposing the time series into level and cyclic-oscillation components using a dynamic model. Each time series is an average for each region, obtained after averaging across individual amplitude estimations for sites within each region.

Download figure to PowerPoint

Our dynamic regression model analyses on monthly NDVI and climate data yielded time-varying (monthly) coefficients of NDVI response to changing temperature and precipitation. We found that the regression coefficient of NDVI response to temperature was mostly positive for Africa and Southeast Asia (Fig. 5a, top). While the values of this coefficient remained positive for Africa, it decreased over time for Southeast Asia (Fig. 5d and f, top). On the other hand, the temperature coefficient was negative throughout for Central America and predominantly negative for South America and South Asia (Fig. 5a, top). In Central America and South Asia, the regression coefficient for temperature tended to decrease over time (Fig. 5b and e, top), and although an increasing trend was evident for South America, the coefficient remained predominantly negative (Fig. 5c, top). In four regions (Central America, South America, Africa, and South Asia) there appears to be a tendency for increase in the NDVI temperature coefficient in the early part of the period followed by a decrease in the mid-1990s, whereas for Southeast Asia there is a more consistent decrease over time. Compared with temperature coefficients, the magnitudes of the regression coefficients of NDVI response to precipitation display weaker influence, with the median annual values that are close to zero or slightly negative in all regions (Fig. 5, bottom). The same trends and patterns are evident when looking at annual parameter averages along with spatial variability quantified within each region (Fig. S3).

image

Figure 5. (a) Time evolution of NDVI–climate relationships. Boxplots of the monthly time-varying slope coefficients from the dynamic linear regression analyses of monthly NDVI against monthly temperature and precipitation. The monthly time series for each parameter was obtained by averaging across parameter time series among all sites within a region (a). Each boxplot represents 300 monthly values covering the period 1982–2006. ((b),(c),(d),(e),(f)). The same data from Fig. 5a but plotted region wise for each year during 1982–2006, using 12 monthly values per year. The boxes indicate the interquartile range, and encompasses 50% of the data values and the open circles indicate data points that are located at least 1.5 times the interquartile range away on each side of the box. Whiskers join the box to the farthest values smaller than this threshold.

Download figure to PowerPoint

The dynamic regression models of NDVI vs. temperature and precipitation also yielded a time-dependent intercept or NDVI level for each region. This quantifies the NDVI trends that remain after accounting for the dynamic effects of temperature and precipitation. We compared the overall trend in the time-varying intercept or level by computing the Sen slope of the intercept time series for each site in each region, with individual sites contributing to spatial variability about the median. A positive value of Sen slope for the time series of the intercept indicates residual greening whereas a negative Sen slope indicates residual browning. Our results show evidence for both residual browning and residual greening in different regions (Fig. 6). We found mostly residual browning in Central America, South America, and South Asia, where the median values (among blocks within each region) of the Sen slope were negative, and residual greening in Africa and Southeast Asia where median values of the Sen slope were positive (Fig. 6).

image

Figure 6. Time-varying NDVI level or intercept obtained from dynamic linear models for each continental region. The response variable was the monthly NDVI and the covariates were monthly temperature and precipitation. Sen slope of the annual intercept time series for each site was computed and plotted as boxplots grouped by region. The boxplots display spatial variation within each region. The boxes indicate the interquartile range, and encompasses 50% of the data values and the open circles indicate data points that are located at least 1.5 times the interquartile range away on each side of the box. Whiskers join the box to the farthest values smaller than this threshold.

Download figure to PowerPoint

Finally, we tested elevation dependence in these vegetation greening/browning rates using quantile regression models and again found a globally consistent pattern. We computed Sen slopes of the monthly NDVI time series for the entire period 1982–2006 for all 801 AVHRR pixels and found that most pixels in each region show browning. Quantile regressions of the Sen slope of the NDVI time series against pixel elevation show that browning rates became progressively weaker with increase in elevation in all the mountain regions. Although most pixels have negative Sen slopes (browning), pixels located at higher elevations tend to show smaller negative values, whereas some pixels show positive Sen slopes (greening) (Fig. 7). Furthermore, this elevation dependence was in general stronger for quantiles higher than the median (Fig. 7; Table 3).

Table 3. Rate of greening in mountain protected sites as a function of elevation estimated using quantile regression. The P-values corresponding to the higher quantiles of the response were all <0.1 in all regions, and less than 0.05 in four of five regions. Also, the elevation effect is positive at all quantiles and increases with elevation for Africa, Southeast Asia, South Asia, and Central America
RegionNo. of pixelsQuantileElevation coefficientSEt valueP(>|t|)
Central America400.500.00050.00070.82510.4145
0.750.00060.00060.92170.3625
0.950.00070.00041.71750.0940
South America2900.500.00140.00027.85980.0000
0.750.00110.00018.14340.0000
0.950.00060.00014.23990.0000
Africa1130.500.00040.00060.73530.4637
0.750.00070.00051.43070.1553
0.950.00260.00102.69440.0082
South Asia1550.500.00060.00032.26750.0248
0.750.00060.00022.97960.0034
0.950.00080.00024.18990.0001
Southeast Asia2030.500.00090.00042.29860.0226
0.750.00110.00042.51640.0126
0.950.00120.00042.89360.0042
image

Figure 7. Elevation dependence of trends in vegetation greenness. Quantile regressions of the Sen slope of the annual maximum NDVI time series (greening/browning rate) vs. elevation of individual pixels in each region ((a),(b),(c),(d),(e)). Regression lines at quantiles 50% (lower/solid), 75% (dashed/middle), and 95% (dotted/upper) are shown. Regression slopes for all quantiles in all regions were positive, and these were statistically significant for 95% quantiles for all regions (P < 0.01) except Central America (P = 0.09).

Download figure to PowerPoint

Discussion

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

We used satellite-derived data on vegetation greenness for the period 1982–2006 to understand the response of mountain vegetation to recent climatic changes. The observed mild greening trends with reversal to browning around the mid-1990s are not only consistent among tropical mountain regions but are also similar to patterns that have been reported for regions in temperate latitudes. In North America, for example, climatic data reveal that April–May temperatures increased until the late 1980s to early 1990s, and decreased thereafter. Corresponding patterns of spring-vegetation greening and browning that match these temperature changes have been found (Wang et al., 2011). Similar trend reversals are known from other biomes and regions around the world, including the greening to browning shift of the mid-1990s (Nemani et al., 2003; Beck et al., 2011; Piao et al., 2011; Verbyla, 2011). The browning phase in particular has been reported for large areas of the Northern Hemisphere (Lotsch et al., 2005) and tropical regions (Sobrino & Julien, 2011), and these changes have being attributed to drought-induced reduction in global net primary production after year 2000 (Zhao & Running, 2010). These and other shorter term global assessments (Xiao & Moody, 2005) indicate that vegetation browning was associated with rising temperature and/or decreasing precipitation (combined moisture stress), whereas greening has been generally attributed to simple warming and/or increased precipitation. A strong and matching signal of reversal has also been reported for canopy tree growth rates in a lowland tropical forest during years 1984–2000 (Clark et al., 2003). Comparing annual tree growth rates to the 16-year mean they found negative anomalies starting during the early 1990s and persisting until year 2000, a pattern driven by increasing annual minimum temperature and strong El Niño warming. Our results for tropical mountain vegetation are consistent with these reported patterns of vegetation response at lowland sites in temperate and tropical latitudes.

The role of temperature-induced moisture stress on vegetation browning is evident from other observations as well. For example, tree-ring data obtained at tree-line ecotones show that increased temperature combined with no gain in precipitation had a negative influence on tree growth (D'Arrigo et al., 2004). Also, pertinent to our study are observations of increase in the dry period length caused by reduced dry season mist frequency and decreased input of nonprecipitating moisture in tropical mountains. This mechanism can account for the relative insensitivity of vegetation greenness to regional precipitation averages that we found in our dynamic model analyses. It also complicates assessments of the importance of moisture on vegetation dynamics where data on inputs of nonprecipitating moisture are unavailable. Such decreased input of nonprecipitating moisture in tropical mountains has led indirectly to amphibian species extinctions, disruption of faunal assemblages (Pounds et al., 1999), and possible decrease in primary productivity, similar to that observed in Northern latitudes (Zhao & Running, 2010). The globally consistent browning trends that we observe indicate that such phenomena are probably more widespread in tropical mountains, with implications for primary production and species diversity at all levels.

The lifting-cloud-base and related mechanisms that cause changes in dry season moisture availability in tropical mountains predict elevation-dependent changes in climatic parameters. Our tests show that rates of change in vegetation dynamics were dependent on elevation in a consistent manner across all mountain regions. The observed decreases in browning rates with increase in elevation may be driven primarily by a weakening of the effects that cause decreased mist frequency. But if warming rates were more pronounced at higher elevations, moisture stress could intensify and cause browning (Beniston, 2003). Contrarily, elevated temperatures could also ease constraints to plant growth and result in increased productivity and greening, if moisture were not limiting (Nemani et al., 2003). Therefore, multiple factors can potentially interact and influence the vegetation dynamics that we observe. Although we do expect to see differences in vegetation responses among regions based on vegetation characteristics such as substrate-dependent drought tolerance (Aiba & Kitayama, 2002), the synchronous responses and trend reversals across far-flung continental regions of the globe indicate the presence of strong global drivers that dominate regional differences in vegetation and the environment. Nevertheless, our findings from quantile regressions of NDVI vs. elevation, which show strong influence at high quantiles, indicate that explanatory variables not included in our models (e.g., dry period length and drought tolerance) are important.

Observed climatic changes have not only altered vegetation characteristics but also appear to have affected the nature of vegetation–climate relationships (Barber et al., 2000). Dynamic regression models, which enabled us to capture the time evolution of these relationships, show that NDVI–climate relationships changed over time in all regions. The dynamic regression coefficient of NDVI vs. temperature showed both positive and negative values, except for Central America, where it remained negative throughout the time period. Two important patterns appear to emerge from these results: (i) there appears to be some evidence for trend reversals in the NDVI–temperature relationships (Fig. 4), although the patterns are not as clear as for the greening/browning trends that we found for annual maximum NDVI (Fig. 2) and (ii) the NDVI temperature coefficient obtained using monthly values decreases over time after the mid-1990s. This latter pattern in trends indicates a consistent loss of positive temperature sensitivity in at least four regions (all except Africa), and perhaps evidence for a complete reversal in Central America. Such loss of positive temperature sensitivity has been reported earlier for temperate latitudes as constituting a dramatic shift in vegetation–climate relationships due to recent climatic changes (Barber et al., 2000). Our dynamic regression model analyses, however, show that monthly precipitation had only a very weak influence as the NDVI-precipitation regression coefficient values were close to zero or mildly negative in all regions. However, the importance of nonprecipitating moisture inputs remains to be evaluated across these tropical mountains. Nevertheless, the loss of positive temperature sensitivity of vegetation greenness in the face of rising temperature and no corresponding increase in moisture constitutes strong evidence for changes in moisture regimes that cause drought-like conditions. In combination with increasing temperature such changes in moisture regimes appear to be driving the vegetation responses that we observe in both temperate latitudes and tropical mountains.

The presence of significant patterns in the time-dependent intercept or NDVI level is a clear indication that there are other climatic or nonclimatic drivers of vegetation change in tropical mountains. The role of processes such as reactive nitrogen deposition (Townsend et al., 1996), CO2 fertilization (Friedlingstein et al., 1995), and forest regrowth (Pan et al., 2011) are known from several studies, even from lowland tropics (Ichii et al., 2002), but their importance in tropical mountains has not been evaluated. We found residual greening trends and browning trends in almost equal measure among regions and, but within regions, either residual greening or browning tended to dominate. These regional differences need to be investigated further, but distinct regional scale phenomena such as the Asian Brown Cloud (Ramanathan et al., 2005) could influence climatic conditions in particular tropical mountain regions (Bonasoni et al., 2010), and thus vegetation. As all unmodeled effects, including satellite sensor performance, will contribute to the time-varying intercept, attribution to any specific driver in any region is difficult without detailed data from other sources.

The changing dynamism of NDVI–climate relationships is further evident in the observed changes in amplitudinal variation in the seasonal NDVI cycle. Future climate change scenarios predict increases in seasonal climatic variability and frequency of extremes (IPCC, 2007a), and evidence for such changes has been found in one mountain region (Liu et al., 2006). Using a dynamic model that extracted the cyclic component of NDVI after removing trends, we found a systematic increase in the amplitude of the seasonal NDVI cycle in all regions. Although this amplitudinal increase in NDVI was strongly correlated with amplitudinal variation in temperature, and to a lesser extent with precipitation during the same period, a causal link cannot be inferred from these analyses alone. In any case, the observed systematic amplification of the seasonal climatic cycle has the potential to intensify temperature-induced moisture stress in these ecosystems. Such stress could operate both in the dry season and during dry episodes in the wet season, when increasing frequency of high-intensity precipitation events in the tropics due to warming could increase runoff and reduce recharge of soil moisture (Bonell et al., 2010; Krishnaswamy et al., 2012; O'Gorman, 2012). Finally, changes in the timing of specific events can impact mountain vegetation, as reported in a temperate mountain region where early snow melt led to increased wildfire activity and tree mortality (Trujillo et al., 2012).

Ideally, the evaluation and attribution of vegetation responses to climate change would require long-term ecological observations. However, such data are limited, so analyzing long time series of satellite-derived data on vegetation change is a valuable and complementary approach to documenting and understanding the biological impacts of climate change on biota (Pettorelli et al., 2005). The GIMMS-AVHRR data that form the basis of our study have been widely used to examine vegetation responses to climate change at all spatial and temporal scales for which these data are available (de Jong et al., 2012). Although some questions have been raised about GIMMS-AVHRR data quality in some studies (e.g., Zhang et al., 2013), these data have been evaluated extensively in comparison with both satellite-derived data from independent sensors (Fensholt et al., 2006) as well as ground-based data from ecological observations (Sobrino & Julien, 2011; Luo et al., 2013) and found to be robust for studying vegetation characteristics.

Proper attribution of observed changes in vegetation is, however, difficult without matching ground-based measurements. We designed our study carefully to minimize confounding effects of land-cover or land-use change with the effects of climate change. Restricting our study to well-protected areas, we found that increasing temperatures associated with climate change in tropical mountains are influencing vegetation phenology in multiple ways. We found consistent decadal-scale trends and trend reversals in vegetation dynamics in all regions. These rates of vegetation changes were dependent on elevation, which support recent assertions that climate warming rates and temperature-induced moisture stress are important in tropical mountains. Furthermore, the amplitude of the seasonal NDVI cycle has been increasing steadily with time in all tropical mountain regions. These changes have already led to significant changes in the nature of vegetation–climate relationships. Although ecosystem responses to recent climate change are being reported from specific mountain sites (Shrestha et al., 2012), we studied sites located across the pantropical belt, spanning different elevations within each region. Our findings support a dominant role for global drivers of vegetation, but important regional differences in climatic effects on tropical mountain vegetation remain and need to be urgently investigated.

Acknowledgements

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

The datasets were obtained from World Database on Protected Areas (UNEP-WCMC), Earth Remote Sensing Data Analysis Center (Japan), Global Land Cover Facility (University of Maryland), GLDAS, NASA Goddard Earth Sciences Data and Information Services Center (GES DISC), and Global Historical Climate Network, Earth System Research Lab (ESRL). We thank the Department of Biotechnology, Government of India, for full financial support (BT/01/NE/PS/NCBS/09). We gratefully acknowledge Samantha Ryder for extensive help in organizing and processing data, Hualan Rui and Stefan Kern for help with data queries, and Nathan Phillips, Brian Cade, Mahesh Sankaran, and Harini Nagendra for comments on an earlier version of this manuscript. Finally, we thank three anonymous reviewers for incisive comments that helped us greatly to improve this report.

References

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information

Supporting Information

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  9. Supporting Information
FilenameFormatSizeDescription
gcb12362-sup-0001-FigS1-S3-TabbleS1.pdfapplication/PDF1592K

Figure S1. Maximum annual NDVI with spatial error bars (one SE centered on the mean) for the five tropical mountain regions.

Figure S2. Amplitude of NDVI, Temperature, and Precipitation with spatial error bars at each time step (one standard error is centered on the means).

Figure S3. Time series of annual averages of time-varying regression coefficients from the dynamic linear models with spatial error bars for each region (one SE is centered on the mean).

Table S1. Results of multiple regressions of NDVI (detrended) amplitude in response to (detrended) Temperature and Precipitation amplitudes as covariates.

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.