10.1007/s11434-011-4870-8 Changes in global potential vegetation distributions from 1911 to 2000 as simulated by the Comprehensive Sequential Classification System

2011 Vegetation classification models play an important role in studying the response of the terrestrial ecosystem to global climate change. In this paper, we study changes in global Potential Natural Vegetation (PNV) distributions using the Comprehensive Sequential Classification System (CSCS) approach, a technique that combines geographic information systems. Results indicate that on a global scale there are good agreements among maps produced by the CSCS method and the globally well-accepted Holdridge Life Zone (HLZ) and BIOME4 PNV models. The potential vegetation simulated by the CSCS approach has 6 major latitudinal zones in the northern hemisphere and 2 in the southern hemisphere. In mountainous areas it has obvious altitudinal distribution characteristics due to topographic effects. The distribution extent for different PNV classes at various periods has different characteristics. It had a decreasing trend for the tundra and alpine steppe, desert, sub-tropical forest and tropical forest categories, and an increasing trend for the temperate forest and grassland vegetation categories. The simulation of global CSCS-based PNV classes helps to understand climate-vegetation relationships and reveals the dynamics of potential vegetation distributions induced by global changes. Compared with existing statistical and equilibrium models, the CSCS approach provides similar mapping results for global PNV and has the advantage of improved simulation of grassland

Global climate and environmental changes brought about by anthropogenic means and their potentially serious impact on global and local ecosystems are receiving enormous attention from scientists, governments and society in general [1]. The study of Potential Natural Vegetation (PNV) has been proposed as a way to examine the impact of climate changes on vegetation distributions [2,3]. To understand the spatial distribution patterns of PNV and their spatial and temporal repeatability is therefore a starting point for studying climate-vegetation relationships.
The separation of anthropogenic and non-anthropogenic influences on climate-vegetation relationships is complicated. Climate-vegetation classification or relationship studies, based on existing patterns of vegetation, can potentially enable the evaluation of societal impacts on relationships between climate and vegetation. Potential vegetation refers to the most stable and mature climax vegetation possible without human interference, and shows the overall trend in regional vegetation development over a certain period. The study of potential vegetation can therefore reflect the influence of climatic conditions on vegetation change [4,5]. For this reason, potential natural vegetation has received much attention in the fields of geography, botany, climatology, and ecology [1], since the concept was proposed in 1956 [2].
In general, potential vegetation research can be divided into two stages: the first stage is a traditional qualitative investigation that is primarily based on experienced observations, and the second stage is a quantitative examination based on PNV forecast models. Quantitative methods utilizing GIS and RS techniques have generally superseded qualitative approaches for potential vegetation mapping, thereby overcoming the obstacles of subjectivity, repeatability, low efficiency, high cost and consumption of time. The simulation procedures and resulting outputs of such quantitative methods depend strongly on digitization and visualization techniques. Simulations result in two types of vegetation classification models: one consists of statistical models based on biogeographical and plant physiology parameters [6,7] and the other is process-based [8][9][10][11]. The latter has particularly improved our ability to understand the response of terrestrial vegetation to past and future environmental variations at global-to-regional scales [1]. However, a process-based model tends to require the integration of more parameters and a great deal of data to adequately consider biogeographical and biogeochemical processes and fully simulate the dynamics of the terrestrial ecosystem (often a shortcoming for the models). There has been little investigation into the level of complexity that is required to capture primary climate-vegetation feedbacks, despite the fact that increasing complexity is raising the computational costs of vegetation classification simulations [12,13]. Additionally, process-based models can have difficulty simulating climate-vegetation relationships from the more distant past, because of the limited availability of historical observations in many parts of the world. This increases the likelihood of errors in plant-types prediction for past time periods [13]. In comparison with process-based models, biogeographic models, which depend on climatic observation data collected over a broad geographical extent and a long time span, are characterized by simple input parameters [11]. In a related field, plant ecologists have been interested in understanding the geographic distribution of vegetation and the consequences of climate change to vegetation dynamics. This aspect has made the biogeographic approach, in which the distribution of vegetation types is predicted from climate variables, more popular [2]. Thus, biogeographic models have an irreplaceable role in the study of the dynamics of potential vegetation distribution on a global scale.
In order to explore the global terrestrial vegetation classes and their areal extent, a natural vegetation classification approach, the Comprehensive Sequential Classification System (CSCS) [14], is used in this study. The CSCS approach is a biogeographic simulation method that was used originally to classify grasslands in China [15]. Using this approach, global potential vegetation is divided into 42 classes based on a biogeographic model; additional detailed grassland vegetation types can be added, which makes this approach significantly different from most other potential vegetation classification methods. This paper documents the first application of the CSCS approach on a global scale, and provides a scientific basis for the temporal and spatial distribution of potential terrestrial vegetation, with a focus on grassland vegetation. The purposes of this study are (1) to validate the CSCS-derived model at a global scale by comparing it with the Holdridge Life Zone (HLZ) and BIOME4 PNV classification models; (2) to examine the spatial distribution characteristics of global potential vegetation in 2000 as generated by the CSCS method; and (3) to analyze the changes in the global potential vegetation distribution over a 30-year temporal scale during the 90 years from 1911 to 2000, based on the CSCS method applied to broad vegetation categories.

The CSCS approach for mapping potential natural vegetation
The CSCS approach is formulated through the grouping or clustering of units with similar moisture and temperature properties [14]. CSCS consists of a 3-class level: Class, Subclass and Type. At the first level, vegetation is grouped into classes according to an index of moisture and temperature. At the second level, vegetation subclasses are differentiated by edaphic conditions. At the third level, vegetation types within a subclass are distinguished by vegetation characteristics.
The class level is the basic unit, and it is identified according to zonal characteristics of biological climate. In practice, the class is determined by combining the quantitative biological climate indices of average annual cumulative temperature above 0°C (Σθ) (i.e., Growing Degree-Days on 0°C base, GDD0) and humidity (K), as calculated by [16].
where MAP is the annual mean precipitation (mm); and 0.1 is a justified coefficient of the model.
Based on decades of studies [14][15][16][17][18][19], 7 thermal zones and 6 humidity zones ( Figure 1) have been identified and used to differentiate vegetation classes. The CSCS recognizes 42 vegetation classes (Table 1), of which all but the tropical desert class (class code 7 of Table 1) are present in China [14]. To more explicitly reflect the spatial distribution of potential vegetation at a large scale, classes are merged into 10 broad vegetation categories (Table 2).

Global climatic data
Two sources of monthly precipitation and mean temperature data with different spatial resolutions for global land areas (excluding Antarctica) were used in this study to produce CSCS-based global potential vegetation distribution maps. The first was the global monthly precipitation and mean temperature dataset gridded at a 30 arc-second (i.e., about 1 km) spatial resolution over the 50 years from 1950 to 2000. This was generated using the thin-plate smoothing spline algorithm implemented in the ANUSPLIN software package [20] using weather stations from a large number of global, regional, national, and local sources [21]. We chose this dataset because the method that created it has been used in other global studies [22,23] and performed well in comparative tests of multiple interpolation techniques [21,24]. The other source of data was the Climate Research Unit (CRU) global climate dataset of CRU_TS 2.1 [25]. It consists of multi-variate mean monthly climatology records at 0.5° resolution for global land areas (excluding Antarctica) for the period 1901-2000. In this study, the monthly precipitation and mean temperature datasets for each decade were used. To compare the agreement between the CSCS and HLZ PNV maps and simulate the change of global potential vegetation distribution using the CSCS approach, the average annual GDD0 and precipitation grid data for different periods using the two datasets were generated using ArcGIS software.

Global DEM and continent boundary data
The global digital elevation model at a 30 arc-second spatial resolution [21], based on the Shuttle Radar Topography Mission (STRM) dataset, was downloaded from http://srtm. csi.cgiar.org. To analyze the areas of different vegetation types, the boundary databases of Europe, Asia, Africa, the Americas and Oceania were downloaded from http://www. diva-bgis.org/Data, and utilized in this study. Using ArcGIS software, all of the related databases are transformed to the Mollweide projection with the WGS_1984 spheroid to calculate the area of each vegetation class.

Comparative maps of global potential vegetation
HLZ is a model that formulates the distribution of potential terrestrial ecosystems in terms of biotemperature, precipitation and potential evapotranspiration [26]. It relates the distribution of major ecosystems (termed "life zones") to the bioclimatic variables. The HLZ classification divides the world into over 38 life zones (Table 3) on the basis of mean annual biotemperature (in °C), average total annual precipitation (in mm) and potential evapotranspiration ratio (PER) logarithmically [6,26,27]. BIOME4 is a process-based equilibrium, biogeographically-and biogeochemically-coupled vegetation model, modified from Biome3 [28], which simulates global vegetation in the form of 13 plant functional types (PFTs) that are combined to form 28 biomes (Table 3) [29][30][31]. BIOME4 has been employed in a number of studies of past, present and potential future vegetation patterns [1,[32][33][34].

The KAPPA statistic for PNV map comparisons
The Kappa statistic is widely used in assessing model-simulated vegetation distributions. The advantage of the Kappa statistic is that it takes chance agreement into account, regardless of the number of categories being compared in the maps [1,8]. In this study, the Kappa statistic was used to evaluate the similarities between the two kinds of PNV maps. For each category, i is the constructed error matrix for two compared PNV maps, the Kappa statistic is calculated by the following equation: where i p ,row is the row total for each category i; j p col, is the column total for each category i; and p ii is the individual entry for the row and column on the main diagonal of constructed error matrix. The overall agreement between two compared maps is estimated by the formula:

Agreement in CSCS, HLZ and BIOME4 PNV maps
From Table 5, it can be seen that the CSCS and HLZ PNV maps are more in agreement than comparisons of maps BIOME4 comparison, and very poor agreements (0.26 and 0.14) for CSCS vs. HLZ and HLZ vs. BIOME4 comparisons.

Spatial distribution characteristics of PNV in year 2000
From Figure 3, 10 broad vegetation categories and 42 classes of the potential vegetation can be identified for the world excluding Antarctica. Statistical analysis indicates that the area of global potential vegetation, excluding regions of permanent snow and ice cover, is 1.289335×10 8 km 2 , which covers 96.07% of the total land area of Earth ( Table 6).
The spatial distributions of classes of potential vegetation are significantly different from region to region. In Asia, the potential vegetation area is about 4.29036×10 7 km 2 , which is the most extensive of the 5 continental regions and covers 33.28% of the total area of global potential vegetation excluding Antarctica ( Table 6). Because of its extent and complicated ecological environment, Asia has the greatest number of vegetation categories in comparison to the other continental regions. Based on the CSCS classification approach (Table 1), there are 42 potential vegetation a) The numbers from 1 to 10 correspond to the broad vegetation categories: 1, tundra and alpine steppe; 2, frigid desert; 3, semi-desert; 4, steppe; 5, temperate humid grassland; 6, temperate forest; 7, sub-tropical forest; 8, tropical forest; 9, warm desert; 10, savanna. b) The percentage of the Earth's land surface covered by the vegetation category. c) The percentage of the global distribution of the vegetation category found in the continent or region.
The Americas form the second largest region and they contain 42 potential vegetation classes. The potential vegetation area is 3.79686×10 7 km 2 , which covers 29.45% of the total area of the global potential vegetation. The dominant vegetation classes are deciduous broad leaved forest iii (39) and rain forest (42), with their individual areas accounting for over 50% of each corresponding global area.
The potential vegetation is characterized by a significant distribution pattern in latitudinal and altitudinal directions (Figure 3). From the Equator to the North Pole, there are 6 major latitudinal zones, which in sequence are: (1) tropical forest, dominated by rain forest (code 42), mainly distributed in the north of South America, Middle Africa and southeastern Asia; (2) Savanna, dominated by tropical desert brush (14) and sub-tropical desert brush (13), mainly distributed in eastern and central Africa, Central America, and the south of southern Asia; (3) warm desert and sub-tropical forest, principally including tropical desert (7), evergreen broad leaved forest ii (41), and mainly distributed in North Africa, North America and a large part of Asia; (4) semi-desert and frigid desert, mixed with steppe in central Eurasia and southwestern North America, as well as tundra and alpine steppe in the Tibetan Plateau, dominated by the classes of temperate semi-desert (10), warm temperate semi-desert (11), warm temperate desert (4), and rain tundra and alpine meadow (36); (5) mixture of temperate humid grassland and temperate forest, mainly including the classes of perhumid taiga forest (37), mixed coniferous broad leaved forest (38) and meadow steppe (24), and distributed over all of Eurasia, extending from the east of North America to the east coast of the Pacific Ocean; and (6) tundra and alpine steppe, dominated by the classes of rain tundra and alpine meadow (36) and tundra and alpine meadow (29), mainly distributed in northern North America, Greenland and most northern Eurasia.
In the southern hemisphere, from the Equator to the most southern edge of Oceania, there are just two major latitudinal zones: (1) The first is dominated by tropical forest, sub-tropical forest and savanna vegetation categories, mixed with a little bit of steppe, and warm desert. Major classes found in this zone include rain forest (42), seasonal rain forest (35), tropic desert brush (14), sub-tropic desert brush (13) and tropical desert (17). This region is mainly distributed in southern and central Africa, northern and central South America, the south of southeastern Asia, and northern Oceania. (2) The second region is dominated by temperate forest and semi-desert, mainly including the classes of perhumid taiga forest (37) and sub-tropical desert brush (13), and mainly is distributed in southern-most America, Africa and Oceania. As a result of interactions among topography, climate and vegetation, there are probably more vegetation categories distributed in eastern South America.
In mountainous areas, spatial vegetation distribution is more complicated and has obvious vertical distribution characteristics because of topographic effects. For example, in the Himalaya mountains and Tibetan plateau in southwestern China and adjacent areas (Figure 4), there is a great deal of tundra and alpine steppe (purple color) scattered with ice and snow (white color) mixed with temperate humid grassland (yellow color) and temperate forest (dark blue color) (Figure 4(a) and (b)). From the top area of Mt. Everest to the southwest through Nepal and India ( Figure  4(c)), the elevation decreases from > 8200 m to about 50 m (Figure 4(d)). The potential vegetation categories change stepwise from tundra and alpine steppe, to temperate forest, to sub-tropical forest (bright green) to tropical forest (red color) in the southern Nepal, and to savanna (dark green) in the north of India.

Change in global PNV distribution from 1911 to 2000
From Table 7, clear decreasing trends can be seen for the area of tundra and alpine steppe and desert vegetation categories. This has amounted to 6.06% and 5.90%, respectively, over the 90 years from 1911 to 2000. Over the same period, the area of forest and grassland vegetation categories has increased by 2.23% and 4.39%, respectively.
In the 3 forest vegetation categories, the total area increased by 3.06% over the period from 1940 ( Figure 5(a)) to 1970 ( Figure 5(b)), then slightly decreased by 0.80% over the period from 1970 to 2000 ( Figure 5(c)). The area of temperate forest increased significantly from 1940 to 1970, but no obvious trend emerged from 1970 to 2000. The area of sub-tropical forest changed slightly in the 90 years. It had a small increase from 1940 to 1970, which was followed by a slight decreasing trend after 1970. This is the same as what was observed for variation in tropical forest cover. In

Discussion
Overall, good agreement was found among the CSCS, HLZ and BIOME4 PNV maps in their evaluations of the extent of tundra, forests and desert, and poor agreement was found for grassland vegetation coverage. However, it is not an easy task to explain why. A first point to consider is that discrepancies exist in the level of detail under which grassland vegetation is classified in the three PNV models. Grassland vegetation accounts for 25% of global land coverage, but it is simplistically represented in the HLZ and BIOME4 models. In the HLZ model, 38 PNV types are recognized, but only one type of grassland (i.e., cool temperate steppe) is specified and no type of savanna [35]. The BIOME4 model focuses more on forest vegetation, including 13 types that are based on thermal grades, and evergreen, deciduous, needleleaf and mixed types from vegetation characteristics (Table 3). However, the BIOME4 model does improve over HLZ in its examination of grasslands. Of the 28 PNV types identified by BIOME4, 5 are grasslands (i.e., tropical grassland, temperate grassland, tropical savanna, cold parkland and temperate deciduous broadleaf savanna). That is, the CSCS PNV approach is even more suitable for classifying grasslands -which it was originally devised to do. Not only does it provide a detailed classification of grassland vegetation, but also produces a detailed classification of the other vegetation types. The CSCS can distinguish 42 classes of vegetation, consisting of 11 classes of grassland, 15 classes of forest, 6 classes of tundra, and 10 classes of desert (Table 1). This means that the CSCS model can produce a more balanced classification of the terrestrial ecosystem than the HLZ and BIOME4 models do. Therefore, the poorer agreements for grassland coverage as assessed by CSCS vs. HLZ and BIOME4 vs. HLZ comparisons could be primarily due to the simplification of grass-land types in the HLZ and BIOME4 models. A second point to consider is that there are significant differences in the constraints applied in constructing the CSCS, HLZ and BIOME4 models. Both HLZ and CSCS are statistical and equilibrium models based on biologically determined climatic constraints, which consider biotemperature and precipitation (MAP) as important classification constraints. However, the model definitions are quite different. In the CSCS model, the biotemperature is defined as growing degree-days on a 0°C base (GDD0), and precipitation is used in the index K (K=MAP/0.1×GDD0) [14]. However, a mean annual biotemperature is defined (i.e., the average value of mean daily temperature over 0°C and below 30°C, BT), and precipitation is not only directly used as a constraint, but generally used to simulate the potential evapotranspiration ratio (PER) (e.g., PER=58.9×BT/P) as well in the HLZ system [6,26,27]. In comparison, BIOME4 is a coupled biogeographic and biogeochemical model that simulates the equilibrium distribution of major PNV types (biomes). Compared to the CSCS-derived model, the BIOME4 model employs not only more bioclimatic constraints (e.g., GDD0, GDD5, temperature of coldest month, mean monthly precipitation, temperature, and percent sunshine) but also other constraints (e.g., PFT, LAI, soil moisture) [29][30][31]. The above discrepancies among the CSCS, HLZ and BIOME4 models ordinarily lead to different definitions for the same vegetation type. Consequently, there are still some deviations in the definition of broad vegetation categories, even after aggregation with the original types. By repeating tests, we found that separation of the current five big categories into finer categories (e.g., the division of the forest class into boreal forest, temperate forest, sub-tropical forest and tropical forest) is difficult. This is because there are big differences in the three sets of classification indices, threshold values and classification approaches.
The Intergovernmental Panel on Climate Change (IPCC) Third Assessment Report [36,37] indicated that during the 20th century, precipitation in much of the Northern Hemisphere land areas increased by 0.2% to 1% per decade, and that no comparable systematic changes were detected in broad latitudinal averages over the Southern Hemisphere. The increase in temperature during the 20th century was the largest of any century during the past 1000 years. As a whole, global temperature had an increasing trend over the 20th century, but most of the warming occurred in the two periods from 1910 to 1945 and 1976 to 2000. Our study shows that over the entire study period, the extent of desert vegetation decreased by 5.90%, and the area of grassland vegetation increased by 4.39%; decreasing trends were found for the extent of sub-tropical and tropical forests, and an increasing trend for temperate forest; the tundra and alpine steppe had a decreasing trend, but more sharply decreased since the late stage of 1970s (T3). Similar results were also found at regional and global scales [1,[38][39][40][41][42].
This suggests that the late 1970s may be a critical period or key starting point when the spatial distribution of potential natural vegetation started to change significantly because of obvious changes related to global climatic warming [36,37].
This study demonstrates the significant role that the CSCS approach can play in the simulation of potential vegetation distributions in relation to climate change at a global scale. This suggests that the CSCS-based models not only have the ability to investigate the effects of climate change on vegetation type and distribution, but also can contribute to balanced predictions of various vegetations, especially detailed classes of grassland vegetation. This is an obvious advantage of the CSCS-based model and a necessary supplement to the other well-known global PNV models.

Conclusions
On a global scale, there is good agreement among the CSCS, HLZ and BIOME4 PNV maps. The CSCS-derived model has the ability to successfully predict the distribution of tundra, desert and forests, and has better abilities than the HLZ classification system for the simulation of grassland. Compared to globally well-accepted PNV models like HLZ and BIOME4, the CSCS model has advantages in the simple input parameters and low computational costs. It can not only systematically classify the extent of known and unknown global terrestrial vegetation, but can recognize more detailed grassland vegetation classes as well. Additionally, a CSCS-derived model can be used to predict potential vegetation classes and their spatial distribution, which is a critical part of research regarding the effects of climate change on vegetation successions, and plays an important role in the management and planning of anthropogenic controls on terrestrial vegetation, especially in grasslands.