In recent decades considerable interest has been focused on trends in vegetation phenology. Part of this interest derives from the possibility of using the phenological response of vegetation as an indicator of climate change given the sensitivity of vegetation dynamics to climate variables [1–3]. For instance, there is increasing evidence associating vegetation phenological changes, particularly the earlier flowering of plants in Europe and North America, with warmer temperatures during the last 30 years [4–7]. Another important reason is that changes in vegetation phenology have implications for the survival of species whose patterns of behavior are intricately synchronized to it [8] such as pollinators [7,9] and bird species whose reproductive success depends on timing their reproductive cycle with the peaks of food abundance [10,11]. Field measurements of vegetation phenology are labor intensive. Thus it has been logical to turn to the use of remote sensing instruments on earth observation satellite platforms for phenological measurements. The most well-known example is the use of the Normalized Difference Vegetation Index (NDVI) derived from measurements of the optical reflectance of sunlight in the red and near-infrared wavelengths [12,13]. However, the resolution of satellite-borne remotely sensed imagery is such that picture elements (pixels) commonly represent multiple land covers. In addition, remote sensing instruments sense electromagnetic energy that has travelled through the atmosphere. Thus, the seasonal character of vegetation index measurements over time may exhibit trends that represent changes in land cover and viewing conditions (such as changes in the presence of water vapor and aerosols) as well as true phenological responses. Consequently, it is more accurate to state that such studies are concerned with trends in the seasonality of NDVI. NDVI has been shown to be very sensitive to ecosystem conditions [14–16], particularly when paired with maximum value compositing to reduce the effects of atmospheric contaminants, such as clouds and aerosols [17]. Therefore it is of considerable value as an indicator of environmental change.

The longest image time series of NDVI that has been produced is the archive developed by the Global Inventory Modeling and Mapping Studies (GIMMS) group at NASA’s Goddard Space Flight Center from the Advanced Very High Resolution Radiometer (AVHRR) instruments on the National Oceanic and Atmospheric Administration (NOAA) Polar Operational Environmental Satellite (POES) series. This global archive has now been processed to yield a 30 calendar year record (1981–2011) with a spatial resolution of 1/12° (approximately 8 km at the equator) and a bi-weekly temporal resolution. In this paper we examine trends in the seasonality of NDVI using this archive and a recently-introduced analytical procedure known as Seasonal Trend Analysis (STA) [18]. Additionally, we introduce a means of classifying significant trends.

The purpose of this paper is to document the nature, prevalence and biome distribution of trends in the seasonality of NDVI. While numerous studies have examined trends in NDVI using earlier archives and phenological measurements such as SOS (start of season) or the length of the growing or active season (see for example [3,19–22]), this study is novel in its goal to classify trends in changes in seasonality and to establish their relative global frequency of occurrence over the past 30 years. As will be demonstrated, significant trends in the seasonality of NDVI are dominated by a very small number of highly prevalent trend types that have a preferred biome distribution.

The primary data set used in this analysis is the new GIMMS NDVI version 3 (referred to as NDVI3g). This data set is a reconstructed extension of the previous version [23,24]. It comprises NDVI measurements derived from seven NOAA instruments (7, 9, 11, 14, 16, 17, 18), which have been processed using an adaptive Empirical Mode Decomposition (EMD) [23]. This technique is used to identify and remove artifacts from the NDVI time series including solar zenith angle, trends associated with orbital drift, inconsistencies in the AVHRR data among sensors due to differences in sensor characteristics, and other extraneous factors that introduce nonlinear and non-stationary effects on the data. NDVI3g constitutes a 30-year long term data record (LTDR) and is the first dataset of its kind which is appropriate for long term studies of land surface trends in vegetation, seasonality and coupling between climate variability and vegetation over the last three decades.

As a preliminary step, the archive was modified by collapsing the bi-weekly NDVI data to monthly composites using maximum value compositing in order to further reduce cloud contamination and enhance the vegetation signal [17]. The resulting monthly composites were then projected to a Mollweide equal area projection to facilitate the correct calculation of areas. The monthly images from January 1982 to December 2011 were then selected for analysis (a 30 year record of monthly NDVI). This will be referred to in subsequent discussions simply as the NDVI data set.

The NDVI3g archive already has certain areas masked from consideration because they do not contain vegetation, or measurements are ambiguous due to surface conditions [25]. These include oceans, major lakes, the Greenland icecap, many parts of Arctic regions and Antarctica. In addition, we masked out areas where the mean annual NDVI was less than or equal to 0.15 on the basis that these are extremely arid or correspond to bare or desert regions in which trends may represent variations as much in soil characteristics as they do in vegetation [26–30]. Collectively, all unmasked areas will be referred to as land areas with an NDVI > 0.15.

The second dataset used in the analysis was a global mapping of the world’s biomes produced by the World Wildlife Fund [31]. The intention was to provide a basis for understanding the geographic and climatic distribution of seasonal trends. The map was reprojected to a Mollweide equal area projection and rasterized to match the dimensions of the NDVI archive.

Because we were concerned that the seasonal trends analyzed might relate to anthropogenic land cover change, we also needed a means of determining those areas that had undergone substantial modification. For this we used the GlobCover 2009 dataset [32]. GlobCover is a European Space Agency and partners initiative to produce global land cover maps. The 2009 map is based on 300 m resolution data from the MERIS sensor on ENVISAT for the period January–December 2009. The classification follows the United Nations Land Cover Classification System (LCCS), which is more focused on canopy structure than climate. Therefore, we judged it to be less suitable to our larger purposes. However, the mapping of anthropogenically altered landscapes in GlobCover 2009 was of special value to this study. Because the areas mapped as anthropogenically altered come near the end of the time period considered in this study, we reasoned that if seasonality trends were associated with anthropogenic land cover change, they would be very likely to be shown as anthropogenically altered in this map.

To make the map comparable, we reprojected the GlobCover 2009 map to a Mollweide equal area projection and assigned to each 8 km pixel in the output the predominant class among its corresponding 300 m pixels in the original. We subsequently reclassified the Globcover 2009 into two classes: anthropogenic land use and natural land cover. The anthropogenic class included: “Post-flooding or irrigated croplands (or aquatic)”, “Rain-fed croplands”, “Mosaic cropland (50%–70%)/vegetation (grassland/shrubland/forest) 20%–50%”, and “Artificial surfaces and associated areas (Urban areas > 50%)”, with an average reported accuracy of 91% [32]. The natural land cover class included all other natural and semi-natural terrestrial and aquatic vegetation land cover classes. The resulting map was then crosstabulated with the biome map [31] to yield two subcategories for each biome class—anthropogenic and natural. This modified biome map will subsequently be referred to simply as the BIOME data set.

There are 226 classes of significant changes (Figure S1, in the supporting information), that include different combinations of significant changes in Amplitude 0, Amplitude 1, Amplitude 2, Phase 1 and Phase 2. Figure 1 shows a color composite of all significant seasonal trends, with trends in Amplitude 0 rendered in red, trends in Amplitude 1 in green and trends in Amplitude 2 in blue, with a symmetric stretch around zero. What is striking is the magnitude of areas exhibiting significant trends. More than half (56.30%) of all land areas with an NDVI > 0.15 exhibit a significant trend in seasonality over the 30 year time period measured, producing a total of 226 significantly trending classes of seasonal curves. Thus a trending change in seasonality is not a rare occurrence. Further, significant trends are primarily associated with three major classes of seasonal trends. We therefore focus our attention on describing these three main classes of changes, which account for 46.10% of all significantly trending pixels. The first significant class (Class 1) accounts for 20.43% of all significantly trending areas, while the second (Class 2) and third (Class 3) most frequent classes account for 16.54% and 9.13% of all significantly trending areas, respectively. In addition, these three classes represent 25.95% of all land pixels with an NDVI > 0.15.

What is striking about these results is that not only do these three classes dominate seasonal trends, but also that they are associated with a strongly preferred distribution of biomes. Class 1, the most prevalent, is clearly dominated by forests, and particularly broadleaf forests, followed by conifers. Class 2, on the other hand, is dominated by grasslands and shrublands, while Class 3 is predominantly found in the Tundra and Taiga biomes. Further, they are all more commonly associated with natural landscapes, suggesting that anthropogenic land cover conversion is not the primary agent of change in seasonality within these three classes.

Class 1, dominated by forests, exhibits a significant increase in mean annual NDVI without significant changes in any of the other shape parameters. This implies an increase in productivity year round and a lengthening of the growing season with both an earlier green-up and a later green-down.

Long term studies of phenological changes in Mediterranean vegetation have reported an advance of leaf-out and a delay of leaf abscission, and experimental studies have related these changes to warmer temperatures [48]. Many studies have also shown productivity increases in forested areas. For example, according to the analysis of European forests inventories and timber harvest statistics by Ciais et al. [49], the growth rate and the area of European forests increased between 1950 and 2000 leading to an increase in carbon accumulation in the region. Boisvenue and Running [50] provide additional ground and satellite-based evidence of productivity increases since 1950 in their review of the literature on climate change effects on forest productivity. Out of the 49 studies reviewed, 37 found positive trends. Their review included global and regional studies and mostly focused on the northern hemisphere and the Amazon region (where the bulk of Class 1 instances were located).

Studies focusing on forest systems attribute the increase in productivity to several factors such as global and regional warming, which promotes growth in temperature limited forests, CO₂ fertilization, and nitrogen deposition [19,20,50,51]. Increased productivity over the Amazon and Europe has been attributed to radiation increase, largely associated with a decrease in cloudiness, and CO₂ fertilization [50,51]. Changes in forest productivity have also been attributed to changes in land utilization such as decreased harvesting [49,52].

Class 2, dominated by grasslands and shrublands, exhibits significant increases in the annual amplitude of seasonality with no significant trend in any other parameter. Thus increasing NDVI in the green season is balanced by decreasing NDVI in the brown season. Several studies have noted an increasing trend in NDVI in grassland, shrubland and dryland areas [53–55]. However, since the majority of phenological studies focus only on the growing season, it has been difficult to corroborate the associated decline in brown season phenology associated with Class 2. This may be due in part, to the fact that measuring the end of the growing season is more challenging than the leaf out date in spring [22].

Although dominated by grasslands and shrublands, Class 2 is associated with biomes of substantially different climate characteristics. The Mountain Grassland and Shrublands Biome, is adapted to cool temperatures and intense sunlight with the majority of the areas under this biome experiencing a dry season and some having rainfall all year round. The Flooded Grasslands and Savannas Biome is limited by water, being flooded during the wet season and water stressed during the dry season. Other biomes that experience Class 2 seasonal changes have a characteristic dry season, and therefore water availability is the main factor limiting vegetation growth. The Tropical and Subtropical Grasslands, Savannas and Shrublands Biome is limited by an intense dry season [56]. This ecosystem (with C4 dominated grasslands [57]) is regulated by drought, where deciduous plants drop their leaves as a response to dryness (similar to temperate mid-latitudes where deciduousness is related to temperature), and the degree of leaf drop and herbaceous mortality during this season depends solely on soil moisture availability [56,58]. The Temperate Grasslands, Savannas and Shrublands Biome (with C3 dominated grasslands) corresponds to Dry Mid-latitude Steppe and Semidesert areas in which plant growth is limited by aridity (with soil moisture being deficient year round) and cold temperatures during the winter months [56]. The Tropical and Subtropical Coniferous Forest Biome, and the Tropical and Subtropical Dry Broadleaf Forest Biome also experience a high water deficit during a large part of the year. Although their characteristics vary substantially, what these biomes share is the presence of one or more substantial environmental limitations, most commonly related to water. The seasonal trend associated with these areas is such that NDVI (and by association, productivity) is enhanced during the green season and inhibited during the brown season.

Class 3, primarily associated with high latitude areas of the northern hemisphere, exhibits significant increases in both the mean annual NDVI and the annual amplitude in NDVI. As noted previously, this relates to an increase in productivity in the growing season with no effect on the brown season.

In the Taiga biome, primary production is limited by temperature [56]. There has been reported evidence of increased productivity in the Arctic region due to CO₂ induced temperature increases [19,59,60]. However, increasing trends in productivity are not consistent in the Boreal forest [50,60,61], and some authors argue that the observed responses may be related to the role of forest fires regulating productivity [62,63], with instances of increasing NDVI relating to forest re-growth [61].

In the Tundra Biome, vegetation growth is constrained by temperature and ground snow cover. Plant growth is activated only when soils have thawed, and therefore the growing season depends on the depth of snow cover [56]. Seasonal changes in the Tundra biome can be consistently linked to temperature increases and the impacts of climate change on the composition and structure of vegetation. There is evidence of large declines in snow cover duration in the Boreal forest and Tundra Biomes during the 20th century in relation to increases in temperature [64–66], which extends the period of time of plant photosynthetic activity. Moreover, several observational and experimental studies have reported a linkage between climate change and the modification of the structure of plant communities in the Arctic [67,68], particularly increases in shrub cover [68–73] with declines of lichens, moss and forbs [67,68,70,74], which have been related to increases in summer temperature and nutrient availability [68,73,75] and may have contributed to the increase in primary productivity in this biome [69,76].

It should also be noted that a small, but noticeable percentage of Class 3, is found in the grassland/shrubland areas of the Sahel. This trend is similar in effect in that it is located in areas that exhibit increasing productivity in the green season. During the period measured (1982–2011) there has been a marked increase in precipitation in this region with a well-documented greening [77,78]. Thus while the causes are very different from high latitude areas, the effect on greening is the same.

Finally, it is important to note several issues about these results. First, while these are long term trends over the 30 year period examined, there is no assurance that they are not transient or oscillatory. For example, short-term variability in the NAO will have been effectively removed by the procedure used in this study, but low frequency oscillations on the order of decades may still underlie the trends observed. Furthermore, while this study focuses on overall trends, we note that there is variability of the trends in NDVI over space and time. This variability has been documented by studies at different spatial and/or temporal scales emphasizing the need for longer NDVI series, ground validation, and cross validation of NDVI series between sensors [22,79], as well as the examination of discrepancies in the results obtained applying different methodologies [80].

Second, with an 8-km pixel resolution, land cover contributions are inevitably mixed. Thus the trends observed may be as much related to land cover changes as they are to true phenological responses. While the majority of the trends observed were determined not to be associated with anthropogenic land conversion, a substantial number are, and non-anthropogenic land cover changes may be involved, such as was noted in the Arctic.

Third, the approach taken in this paper represents an important extension of the STA procedure. The STA technique introduced in [18] was intended for the examination of seasonal trends in specific locations. In this paper we have extended the technique to focus on families (classes) of curves with similar trends in shape parameters through categorization based on statistical significance using the Contextual Mann-Kendall procedure.

Fourth, it should be noted that the results are based on the GIMMS NDVI3g dataset in its original form, without further corrections. Our preprocessing consisted of the aggregation of bi-monthly to monthly data only. The dataset used here has already been corrected for orbital drift effects, stratospheric volcanic aerosols, and other artifacts [24,81]. It relates well to vegetation productivity and climate, and it is highly correlated with MODIS NDVI, supporting its use for the study of long term vegetation changes [81]. Due to its recent development, we have not found any comprehensive evaluation of the quality of this dataset identifying issues that may have impacted our analysis, but we recognize that further investigations are needed in this respect.

Generalizing from these results, there are several broad conclusions that can be reached from this analysis of the 30-year NDVI record. (1)

Changes in seasonality are prevalent in the landscape. Over half of all land areas (56.30%) are experiencing a significant trend in seasonality. Given the prevalence of cloud contamination and interannual variability which make the determination of statistical significance difficult, this is substantial.

From this remarkable 30-year archive of satellite imagery, we thus see evidence of a greening trend. Monitoring the evolution of this will be an important task for the next 30 years of earth observation.