1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 Running Head: Terrestrial NPP for 0.50 grid cells Terrestrial net primary production estimates for 0.5 degree grid cells from field observations-a contribution to global biogeochemical modeling Daolan Zheng*, Stephen Prince, and Robb Wright Department of Geography, University of Maryland, College Park, MD 20742-8225, USA Submitted to: Global Change Biology Date of receipt: keywords: net primary production, productivity, biogeochemical model, carbon cycle, biomass, GPPDI, IGBP 2) crop identified from GRASSLAND ARE in the discussion section p. 30 (may need to be moved somewhere else); * Send all correspondence to Daolan Zheng at the above address Email: dzheng@geog.umd.edu, phone: 301-314-2798, fax: 301-314-9299ABSTRACT Net Primary Production (NPP) is an important component of the carbon cycle and, 32 among the pools and fluxes that make up the cycle, it is one of the steps that are most 33 accessible to field measurement. While easier than some other steps to measure, direct 34 measurement of NPP is tedious and not practical for large areas and so models are 35 generally used to study the carbon cycle at a global scale. Nevertheless these models 36 require field measurements of NPP for parameterization, calibration and validation. Most 37 NPP data are for relatively small field plots that cannot represent the 0.50 x 0.50 grid cells 38 that are commonly used in global scale models. Furthermore, technical difficulties 1 1 generally restrict NPP measurements to aboveground parts and sometimes do not even 2 include all components of aboveground NPP. Thus direct inter-comparison between field 3 data obtained in different studies or comparison of these results with coarse resolution 4 model outputs can be misleading. We summarize and present a series of methods that 5 were used by original authors to estimate NPP and how and what we have done to 6 prepare a consistent data set of NPP for 0.50 grid cells for a range of biomes from these 7 studies. The methods used for estimation of NPP include: i) aggregation of fine-scale 8 (plot or stand-level) vegetation inventory data to larger grid cells, ii) mapping of grid 9 cells and area weighting of field NPP observations in each mapped class, iii) direct 10 correlation of extensive data sets of ground measurements with remotely sensed spectral 11 vegetation indices, iv) local modeling of NPP using key independent variables, for which 12 maps are available at the scale of the grid cell, and v) regression analysis to link 13 productivity with controlling environmental variables. For a few grid cells whose NPP 14 were obtained for multiple years, temporal analysis was conducted. The grid cells are 15 grouped to the biome level and are compared with existing compilations of field NPP and 16 the results of the Miami potential NPP model. Mean NPP was similar to the well-known 17 compilation of Whittaker and Likens, except for temperate evergreen needle-leaved 18 forest, woodland, and shrubland. The grid cell datasets are a contribution to the 19 International Geosphere-Biosphere Programme (IGBP) Data and Information System 20 (DIS) Global Primary Production Data Initiative (GPPDI). The full dataset currently 21 contains 3654 cells (including replicate measurements) developed from 15 studies 22 representing NPP in croplands, sparse vegetation, shrub lands, grasslands, and forests 23 worldwide. An edited subset consists of 2335 cells in which outliers were removed and 2 1 all replicate measurements were averaged for each unique geographic location. Most of 2 the data incorporated into GPPDI were wholly or partly developed by participants in the 3 GPPDI, in addition to the present authors. These studies are gathered together here to 4 provide a consistent account of the grid cell component of GPPDI and an analysis of the 5 entire data set. The datasets have been deposited in an IGBP-DIS GPPDI database 6 (http://daacl.esd.ornl.gov/npq/GPPDI/Combined_GPPDI_des.html). 7 Introduction 8 9 The carbon cycle is important to humans from both environmental and economic 10 perspectives. As the focus of research on terrestrial primary production and the carbon 11 cycle has expanded from small plots of land to encompass the Earth as a whole, we need 12 data to assess the magnitude, variation and the controlling factors of the key processes at 13 this larger scale. Net primary production (NPP) represents the amount of carbon that is 14 retained by plants after assimilation through photosynthesis and autotrophic respiration 15 (Clark et al. 2001). Owing to availability and relative ease of measurement, field NPP 16 data are often used as indicators of the patterns and magnitude of variation of natural 17 carbon fixation, also to monitor economic yield of crops, pasture and wood. Field NPP 18 data are also needed to parameterize, calibrate and validate numerical models that are 19 designed to estimate NPP and other aspects of the carbon cycle. 20 21 Among the pools and fluxes that comprise the carbon cycle, NPP is the variable that has 22 been most widely measured. Nevertheless, direct measurement of NPP is only practical 23 for relatively small field plots. For larger areas, up to the entire land surface of the Earth, 3 1 a variety of models have been designed to estimate NPP and other components of the 2 carbon cycle. These models can be separated into several categories (Prince et al. 1995; 3 Cramer et al. 1999), including statistical models that correlate environmental variables 4 with NPP; process models developed for small stands of vegetation that simulate 5 photosynthesis, respiration, water use, assimilate allocation, etc., that are scaled up by 6 using driving variables for large areas; box models that describe only the main processes 7 occurring in the biosphere; production efficiency models that are driven with satellite 8 data alone, or combined with other sources of biophysical variables; and atmospheric 9 CO2 observations with inverse atmospheric transport models, with or without carbon 10 isotope discrimination (Schimel et al. 1995; Clark et al. 2001; Harvey 2000). A 11 consistent NPP dataset that suitable for global NPP model validation is long overdue. 12 13 Although model inter-comparison can highlight model weaknesses and inconsistencies, 14 such a comparison cannot provide validation (Rastetter 1996), particularly between 15 models of the same type. In a recent model inter-comparison for which all models 16 reported results at 0.5° resolution (Lurin et al. 1994; Cramer et al. 1999; Ruimy et al. 17 1999), there were no suitable and consistent field NPP available since most 18 measurements, of necessity, are conducted in small areas ranging from <1 to several ha. 19 The Global Primary Production Data Initiative (GPPDI) was therefore proposed and 20 adopted by IGBP DIS (Prince et al. 1995). Of the several activities accomplished under 21 the auspices of GPPDI, we report here on the first efforts to provide spatial estimates of 22 NPP at the same scale as is normally used in current models that are applied to the entire 23 global land area. 4 1 2 NPP models applied to small field sites and large grid cells may disagree for several 3 reasons. i) NPP is controlled by many biotic and abiotic variables, such as climate, 4 topography, soil, vegetation, and management. The probability that the values of all 5 controlling factors for a single sample point will match the combination of mean values 6 of the factors for the grid cell that the point fell in is very low, especially when cell sizes 7 range from 0.50 to 50 (Henderson-Sellers et al. 1986). ii) Different interactions among 8 variables are to be expected for small sites and for large grid cells. For example, 9 temperature and soil moisture at specific sites normally have wider ranges than their 10 averaged values at 0.5° grid-cell scale, and their interactions vary greatly at different 11 values of temperature and moisture (e.g. increases in temperature simulate plant growth 12 at optimal soil moisture but reduce plant growth in dry conditions) (Aber et al. 2001). iii) 13 The influence of some variables on model outputs (e.g. NPP) are non-linear (Coughlan & 14 Running 1989), for example irradiance and net CO2 exchange, and the attenuation of 15 solar radiation through a plant canopy (Waring & Schlesinger 1985) are often modeled 16 using an exponential decay rate. iv) Unusual environmental conditions, which were not 17 allowed for in the model design, may occur in the grid cell (e.g., waterlogged or toxic 18 soils, bare rock). 19 20 Numerous field NPP measurements have been made in ecological studies and in routine 21 natural resource inventories. These data are useful for conducting non-spatial analyses 22 and for validating models not designed for spatial variation in forcing variables. 23 However, these data can in no way reflect the variations in NPP caused by the complexity 5 1 of the landscape over a large grid cell (Cramer et al. 2001). Such comparisons may result 2 in inappropriate confidence in model performance if the models are tested only at 3 selected field sites. In other words, neither field NPP data nor forcing variables can be 4 used in global biogeochemical models without being explicitly aggregated to the cell size 5 that the models are intended to simulate. 6 7 Compared to point NPP observations, contiguous grids of cells can indicate spatial 8 patterns at regional and global scales. The accuracy and precision of an NPP estimate at 9 a particular point is sometimes of less interest than the geographic patterns, trends, and 10 discontinuities of NPP across ecotones. Tests of the skill of a model in estimation of 11 these geographical features depend critically on spatial data such as that provided by 12 continuous grids of cells. It is practically impossible to collect field data using existing 13 techniques across an entire, large area, such as a 0.50 x 0.50 cell. Thus, datasets that 14 integrate existing field data and natural resource inventory results observed at fine to 15 coarse scales are extremely useful. 16 17 The objectives of this study are to; i) provide a suitable and consistent NPP dataset at 0.50 18 cell size with associated driving variables for parameterizing, calibrating, and validating 19 global terrestrial ecosystem models; ii) assess the data quality of NPP values in 20 relationship with some key environmental variables; iii) discuss some scaling issues 21 encountered during NPP aggregation; and iv) analyze the current GPPDI dataset at the 22 biome level. 23 6 1 Methods 2 3 NPP Data Sources 4 5 The field NPP data used to develop 0.5° grid cell estimates came from 15 sources in 6 several major biomes (Table 1). Most of the data were contributed by participants in the 7 GPPDI program. We summarize here their methods and specify any additional 8 processing undertaken before adding the estimates to the GPPDI database and 9 undertaking quality assessment and biome-scale analyses. The type of data varied from 10 one study to another, ranging from plot, to stand, to large segments of entire countries. 11 Moreover the characteristics of the data varied, because of the different purposes for 12 which the data were originally collected and the particular constraints on measurement 13 presented by different vegetation types. None of the sources used included all 14 components of NPP, most commonly the belowground NPP (BNPP) was not measured. 15 Where possible various techniques were adopted to estimate the missing components so 16 that both aboveground NPP (ANPP) and total NPP (TNPP) could be reported. 17 18 Five categories of sources and subsequent data processing were applied (Table 1), 19 although elements of several were combined in some cases. NPP values were developed 20 i) from natural resource inventory data (e.g. forest, rangeland, and crop) at different 21 scales, from plot to county; ii) based on literature data for the major components of a 22 landscape, and high-resolution maps were used to measure the areas of each component; 23 iii) using relationships between remotely sensed spectral vegetation indices and field 7 1 observations; iv) simulation with models using key independent variables, for which 2 maps were available, and v) by using regression analyses with environmental variables. 3 4 The following procedures were applied to most data sets. To convert NPP in biomass 5 units to carbon units, factors of 0.475 and 0.45 were used for forest and grassland, 6 respectively (Olson pers. comm, 2000). Few of the source data used in this study were 7 originally for 0.50 x 0.50 cells so the average of all the original, fine-scale cells that fell 8 within each 0.50 x 0.50 cell was used to represent the NPP value for that 0.50 cell. Any 9 cells with no data were excluded. For the studies in which the original resolution was 10 coarser than 0.50 x 0.50 (e.g. inventory data at county level), resampling procedures were 11 used. All locations were specified by the latitude and longitude at the centers of the cells. 12 13 Aggregation of finer-scale spatial inventory data 14 15 Yield data for crops, rangelands and forests are routinely collected in many countries for 16 natural resource management purposes. For example, the US Forest Service's (USFS) 17 national Forest Inventory and Analysis (FIA) program has conducted inventories since 18 1928 (Birdsey & Schreuder 1992). These data have great potential for aggregation for 19 large area NPP studies, although data for privately owned land are not accessible at a fine 20 enough scale for our purposes. Whereas inventory data may be incomplete, usually 21 reporting economic yield, not NPP, they have the enormous advantage of having been 22 collected according to an explicit, spatial sampling scheme that aims to account for large- 23 area, spatial variability. 8 1 2 1. Forests in eastern USA 3 Brown et al. (1999) provided annual mean aboveground wood increments for both 4 hardwood and softwood for counties of the 33 eastern states based on the USFS FIA data 5 from 1970s to 1990s at a county level. Reported sampling errors from the FIA program 6 were 5% for growing stock volume estimates for 28.3 x 106 m3 and 3% for forest area 7 estimates for 0.4 x 106 ha (Hansen et al. 1992; Woudenberg & Farrenkopf 1995). Larger 8 errors were expected for county level operation (Brown et al. 1999). 9 10 We selected the counties from Brown et al. (1999) with forest cover 75%. First, the 11 county level wood increment data were resampled to 0.50 x 0.50 cells. Second, total 12 litterfall was calculated using equation 1 (Lonsdale 1988) applied to the centroid of the 13 cells: 14 (r2 = 0.63) log Y = 1.02 - 0.059ELE - 0.012L (1) 15 where Y = leaf litterfall production (Mg ha-1yr-1), ELE = elevation in km, and L = latitude 16 in decimal degrees. Third, leaf litterfall was assumed to be 70% of total litterfall (Bray & 17 Gorham 1964; Meentemeyer 1982; Javis & Leverenz 1984; Martinez-Yrizar & Sarulhan 18 1990). Fine root production was estimated based on leaf litterfall (Raich & Nadelhoffer 19 1989), the same method was used in sources 2 and 3 below and coarse root production 20 was estimated as 22.5% of aboveground woody increments (Krankina et al. 1999). 21 Finally, ANPP and TNPP for those 0.50 grid cells that were entirely located within the 22 selected counties were calculated. 23 9 1 2. Superior National Forest, USA 2 Plot level inventory data were obtained from the Eastwide Forest Inventory Data Base 3 (Hansen et al. 1992). A forest cover type / species association map for the region based 4 on multi-temporal Landsat TM imagery (Wolter et al. 1995) was used. The 20 5 associations in the land cover map were combined by Goetz (pers. comm. 2001) to 16 for 6 which biomass and NPP were available. The NPP values for each of the 16 classes were 7 derived from a combination of measurements of boreal forest stands in the Superior 8 National Forest, Minnesota (Goetz & Prince 1996; Woods et al. 1991) and the average 9 values of statewide FIA plots for approximately a 12-year interval (Jenkins et al. 2001). 10 11 Aboveground wood NPP was calculated as the difference between the two inventories, 12 on a tree-by-tree basis, using species-specific allometric equations based on diameter at 13 breast height (dbh), crown depth and height, and aggregated to the plot level. Coarse root 14 production was calculated using ratios of root biomass to above-stump biomass (Wharton 15 et al. 1997; Wharton & Griffith 1998) for species lacking such equations. TNPP was 16 calculated using equation 2. 17 TNPP = ANPP + coarse root production + fine root production (2) 18 NPP estimates in 0.50 cells were aggregated using proportional area weighting of the 19 forest type associations and other land cover classes. We incorporated the NPP values for 20 these cells, as provided by Goetz, into the GPPDI data set. 21 22 3. Forests in the Mid-Atlantic and Maine, USA 10 1 Jenkins et al. (2001) used 2640 mature, closed-canopy FIA plots to estimate potential 2 NPP (the presumed NPP in the absence of disturbance or human management). 3 Aboveground wood production at the plot level was calculated by these authors, based on 4 dbh and tree species using appropriate regression equations. Belowground wood 5 production was estimated from aboveground production using indices from the literature. 6 Litterfall data were based on vegetation types and were obtained from the database 7 compiled by Post et al. (Jenkins et al. 2001). TNPP was obtained from the sum of wood 8 production (both above- and belowground), fine litterfall, and fine root production. An 9 aggregation was conducted for each 0.5° grid cell using those plots whose centers fell 10 within that cell. Thus the accuracy of spatial NPP estimates depended partly on the 11 number of plots (1-46) falling within a cell. Within-grid-cell CV’s ranged from 0.03% to 12 32.7% for total NPP (Jenkins et al. 2001). We included Jenkins' 0.50 NPP cells into the 13 GPPDI as provided without further processing. 14 15 4. Forests in state of Maine and SE, USA and Porozhski ranger district, Russia 16 Krankina et al. (1999) also used USFS FIA data to develop NPP estimates at 0.50 cell 17 size in SE USA and the state of Maine, USA. The SE USA cells were estimated based on 18 county-level survey, while the cells in Maine were based on plot- and tree-level inventory 19 data following the methods described by Jenkins et al. (Jenkins et al. 2001). 20 21 Krankina et al. (1999) selected a sample of 4 inventory blocks with a total area of 843 ha 22 occupied by closed canopy forest in Porozhski ranger district, Russia (59.750N, 32.250E). 23 The sample contained 124 individual forest stands (110 hardwood, 14 conifer) with a 11 1 wide range of ages (15 to 175 years). Three steps were used for calculating an area-based 2 estimate of NPP. During the processes satellite data were used with the sampled stands as 3 training sites to develop forest age and type maps. Fine scale NPP estimates were 4 aggregated across the area to 0.50 cell size. 5 6 We incorporated the 0.50 NPP cells for USA and Russia into the GPPDI without further 7 processing. 8 9 5. Crops in Midwest USA 10 Prince et al. (2001) used harvest yield statistics provided by the National Agricultural 11 Statistics Service (NASS) and the agricultural census for counties that had 50% of land 12 in agriculture and < 20% in forest. Cropland harvest yield was converted to NPP and the 13 remaining biological yield was estimated using harvest indices and the shoot:root 14 biomass ratio for each crop type. No irrigated crops were included. TNPP estimates for 15 the state of Iowa were means of 1982-96, while the estimates for other Midwest states 16 were based on NASS data in 1992 alone. No further processing of this data set was 17 necessary beyond resampling county-level data to 0.50 cells. Multi-year TNPP data from 18 1982-96 were also incorporated. A sensitivity analysis indicated that even with errors as 19 high as 10% in the values of both the harvest index and the root:shoot ratio, the 20 magnitude and temporal trends in NPP varied more from year to year owing to climatic 21 variations than can be accounted for by errors in the estimation methods or by changes in 22 crop management practice, such as fertilization (Prince et al. 2001). 23 12 1 6. Grasslands in the Great Plains, USA - I 2 In the Great Plains region, USA, thirteen major grassland seasonal land cover classes 3 were studied by Tieszen et al. (1997) with data from three distinct sources. Normalized 4 Difference Vegetation Index (NDVI) data derived from the red and infrared channels of 5 the Advanced Very High Resolution Radiometer (AVHRR) on the NOAA series of 6 satellites, were collected for each pixel (1 km2) over a 5-year period (1989-93) for 7 analyzing quantitative attributes and seasonal relationships, and then aggregated to land 8 cover classes. Data from the State Soil Geographic (STATSGO) database were used to 9 identify dominant plant species contributing to the potential forage production in each 10 map unit. Carbon isotope values were related to STATSGO estimates of potential 11 production (Tieszen et al. 1997). They applied an algorithm derived by Gill et al. (Gill et 12 al. 2001) to calculate the BNPP based on ANPP and hence reported TNPP. 13 14 We aggregated their finest 1 km2 products to 0.50 resolution using a focal mean function 15 (ERDAS 1999). Our aggregated 0.50 estimates were well correlated with Tieszen et al.'s 16 50-km estimates (r2 = 0.98). 17 18 Mapping and area-weighting of field NPP 19 20 Where maps are available of properties of the vegetation that are strongly related to 21 productivity, an efficient method for the estimation of NPP of large areas is to obtain 22 estimates of the NPP of each mapped class and then to apply weighting factors 13 1 proportional to the area of each class. This method is difficult to use where a few, clearly 2 distinct NPP classes cannot be found, or where NPP has not been measured in all classes. 3 4 7. Forest in US Pacific Northwest (PNW) 5 A factor that is strongly related to NPP in the PNW is stand age class (Turner et al. 1995; 6 Turner et al. 2000). Cohen et al. (1995) produced a stand-age class map using Landsat 7 TM (25 m, minimum mapping unit) in an area of predominantly coniferous forest 8 centered on the H.J. Andrews Experimental Forest in western Oregon. About 80% of the 9 study area was covered by mixed western hemlock (Tsuga heterophylla) and Douglas-fir 10 (Pseudotsuga menziesii), and 20% by silver fir (Abies amabalis) (Franklin & Dyrness 11 1990). The NPP estimates for each of the six stand-age classes were based on existing 12 data relevant to the region (Turner & Long 1975; Grier & Logan 1977; Gholz et al. 1985; 13 Vogt 1991). The fine-resolution NPP estimates were aggregated to a 0.50 cell using the 14 area weighting method. Average error for the estimates of biomass in the study was on 15 the order of 29% (Turner et al. 2000). We incorporated the NPP for this single cell as 16 provided into the GPPDI dataset. 17 18 Direct correlation of field NPP with remotely sensed spectral vegetation indices 19 20 Under certain circumstances vegetation indices calculated from remotely sensed spectral 21 reflectances of the vegetation in the red and near infrared are well correlated with NPP 22 (Prince 1991). The widely used NDVI has been shown to be nearly linearly related to the 14 1 proportion of photosynthetically active radiation that is absorbed by a green leaf canopy 2 (Fapar) (Goward & Dye 1996). 3 4 8. Savanna in Senegal 5 Diallo et al. (1991) correlated seasonally integrated NDVI derived from AVHRR, High 6 Resolution Picture Transmission (HRPT) data (mapped to 1 km2), with field 7 measurements of savanna production in Senegal. Homogeneous 10 x 10 km field sites 8 representing the major geomorphologic types within the study area were selected using 9 aerial photographs and soil maps. Measurements of aboveground standing biomass were 10 made at the end of the growing season producing an estimate of ANPP. Regression 11 models between the measured ANPP and NDVI were established for the field sites and 12 were used to estimate ANPP for the entire country in each of 11 years (1987-97) using 13 maps of NDVI. r2 between annual ANPP and NDVI varied from 0.52 to 0.83 during the 14 period. 15 16 We calculated annual ANPP mean in Senegal based on the NDVI maps and regression 17 equations over the 11-year period provided by Diallo (pers. comm. 2000). We then 18 aggregated 1 km2 pixels to 0.50 x 0.50 cells using a focal mean function. The ANPP 19 means for 32 0.50 cells were incorporated into the GPPDI dataset. 20 21 9. Forest in China 22 Jiang et al. (1999) used a database for China's forests assembled by the Forestry Ministry 23 of China (Forest Ministry of China 1994) with biomass and NPP data from over 1000 15 1 plots together with volume growth data from more than 5500 permanent plots, combined 2 with the NDVI images from the AVHRR, to estimate TNPP with a spatial resolution of 6 3 x 6 km for 33 distinct classes of forest in China (Wu 1980). The NDVI images were used 4 to regress NPP on the NDVI using an empirical relationship (equation 3) developed by 5 Cheng & Zhao (1990): 6 NPP = A(1 - ln(1 - b*NDVI)) 7 Where A and b are empirical parameters that vary among vegetation types and were 8 determined for the 33 forest classes by statistical analyses in which each forest type was 9 represented by at least > 30 plots. The NPP estimates from the map were highly (3) 10 correlated with the NPP aggregated from data (r2 = 0.92). Errors of the estimates ranged 11 from 0.16% for Euphrates poplar sparse woods to 41% for tropical rainforest (Jiang et al. 12 1999). 13 14 A NPP map containing more than 6500 polygons for the entire China was provided. We 15 converted the polygon map to a grid file at 1 km2 resolution. The 1 km2 grid cells were 16 then aggregated to 0.50 cell size using the focal mean function. To avoid the extensive 17 areas of forest fragmentation, we included the ninety-nine 0.50 cells within which forest 18 cover accounted for 90% area of the cell, based on Hansen et al.’s (Hansen et al. 2000) 19 1 km2 global land cover map. 20 21 10. Forests in Northern Europe 22 Total plant biomass (TPB) of conifer-dominated boreal forest in Finland was measured 23 and correlated with Landsat TM NDVI data by Hame et al. (1997). It was shown by 16 1 these authors that these correlations could be used to convert calibrated NDVI from 2 AVHRR to biomass for a large area from the west coast of Norway to the Ural Mountains 3 in Russia. The model might underestimate total biomass as large as 20%, compared to 4 literature reports (Hame et al. 1997). 5 6 We extracted TPB for Finland and Sweden from Hame et al.'s (1997) 1 km2 map and 7 developed allometric relationships using 660 plots from the National Forest Inventory 8 (NFI) data for Sweden over a 5-year period (1994-98) (Kempe pers. comm. 2000). We 9 calculated NPP as follows: i) TPB (kg ha-1) converted to annual stem increment (m3 ha-1); 10 ii) total litterfall production estimated from Lonsdale’s (1988) statistical model; iii) 11 ANPP and TNPP mapped for 1 km2 grid cells based on relationships between BNPP and 12 ANPP; iv) 1 km2 cells aggregated to 0.50 size (Zheng et al. 2001). 13 14 Local modeling of NPP using mapped environmental variables 15 16 In many cases highly-parameterized local models are the most effective means of 17 extending local NPP measurements across the landscape, but care must be taken to avoid 18 circularity by using such aggregations to test coarser resolution models based on the same 19 assumptions and data. Usually these models are forced with variables such as climate, 20 soils and vegetation cover, all of which can be estimated with reasonable accuracy for 21 limited areas. Although these are model results, and not direct observations of NPP, they 22 are highly tuned to local conditions and can be used, with due caution, to validate more 23 general models that are applicable to the entire global land surface. 17 1 2 11. Continent of Australia 3 In a continental-scale study in Australia, Barrett (2000) developed a statistical model 4 from georeferenced, relational databases containing documented observations of NPP for 5 sites that did not exhibit rapid change. The independent variables were climate, soil, and 6 vegetation, derived from national data sets. Aboveground NPP measurements were 7 collected from 33 sites based on grass harvest data and visual assessment of growth, 8 together with 76 measurements of aboveground biomass and 91 determinations of fine 9 litter mass (Barrett 2000). 10 11 Climate variables were monthly maximum and minimum temperatures and rainfall. The 12 1:2,000,000 Atlas of Australian soil provided soil depth class and gross nutrient status of 13 the soil (Northcote et al. 1975). A potential vegetation map was obtained from the 14 AUSLIG Digital Vegetation Atlas showing vegetation in 1770, before European 15 settlement and agricultural activities started (AUSLIG 1990). The belowground plant C 16 was estimated as the product of the inferred value of aboveground plant C and a ratio of 17 belowground to aboveground plant C (Barrett 2000). 18 19 We aggregated Barrett's NPP estimates to 0.50 cells using the focal mean function and 20 incorporated cells with land cover homogeneity > 90% into the GPPDI dataset. 21 22 12. Mitchell grasslands, Australia. 18 1 In Queensland, Australia, Day (pers. comm. 1999) used the GRASP model (Day et al. 2 1997) to calculate ANPP for a 50 x 50 km region of Astrebla (Mitchell) grassland (21.00- 3 21.50S, 143.50-144.00E). Although spatially uniform, the cell is characterized by marked 4 year-to-year variation in ANPP. Data from similar pastures within 250 km of the site 5 were used to calibrate and validate the GRASP model. GRASP was used to calculate 6 monthly ANPP from September 1959 to August 1998 (39 years) for 100 5 x 5 km cells. 7 8 We aggregated 100 5 x 5 km cells ANPP estimates spatially to a 50 x 50 km ( 0.50) area 9 and calculated the mean ANPP over the 39-year period. Inter-annual ANPP variation of 10 the cell from 1960-97 was also included in the GPPDI data set. 11 12 Regression models 13 14 This method has commonly been applied in areas where NPP is strongly controlled by a 15 single environmental variable, such as elevation, temperature or precipitation. 16 17 13. Greater Yellowstone region, USA 18 In an area of 9,500 km2 (44.00-45.60N and 110.60-111.70W) Hansen et al. (2000) selected 19 90 samples to represent different cover types and elevations. Tree ANPP was estimated 20 by sampling tree density by species and diameter classes. Shrub ANPP was estimated by 21 calculating current biomass from basal area using BIOPAK (Means et al. 1994) and 22 dividing by the assumed average life span of the shrubs. Multiple regression models 19 1 were developed and the best model was applied to estimate ANPP across the entire study 2 area at a cell resolution of about 130 m. 3 4 We aggregated the finer scale ANPP estimates to produce three 0.50 cells in the region 5 using the focal mean function. 6 7 14. Grassland in the Great Plains, USA - II 8 Sala et al. (1988) grouped the grassland productivity data for 9,498 sites collected by the 9 USDA Soil Conservation Service (SCS) (Joyce et al. 1986) into 100 Major Land 10 Resource Areas (MLRA), within each State. The MLRAs were based on land use, 11 topography and elevation, climate, soils, water, and potential natural vegetation and are 12 usually much larger than 0.50 x 0.50 (USDA 1981). 13 14 The average of the ANPP of all the sites within each MLRA in each State was calculated. 15 Long-term averages of monthly temperature, precipitation and soil water-holding 16 capacity were obtained for locations near the geographical center of each State's MLRAs. 17 Climatic variables were correlated with primary production using multiple regression 18 analysis. As a result, ANPP was estimated using equation 4 with r2 = 0.9: 19 20 ANPP = 0.6 * (APPT-56) (4) Where APPT is annual mean total precipitation in mm. 21 22 We resampled the MLRA estimates of ANPP to 0.50 cells. To calculate TNPP from 23 ANPP for Sala's cells, we applied an empirical equation between ANPP and TNPP 20 1 developed from data provided in the Grassland USA III study (data source 15). The 2 application was justified because i) both studies used the same equation to estimate 3 ANPP; ii) both studies covered the same ecological region; and iii) there were strong 4 linear relationships between the ANPP and TNPP estimates in Grassland III dataset (r2 = 5 0.92). 6 7 15. Grassland in the Great Plains, USA - III 8 Parton et al. (pers. comm. 2001) estimated ANPP based on equation 4, developed from 9 Grassland II data set (Sala et al. 1988). The precipitation inputs were from VEMAP data 10 at 0.50 cell resolution (VEMAP 1995). BNPP estimates were calculated as below (Gill et 11 al. 2001): 12 BNPP = 0.6 * belowground biomass * root turnover rate (5) 13 belowground biomass (g m-2) = 0.79 * ANPP - ((MAT+10)*33.3) + 1289 (6) 14 where MAT = mean annual temperature (0C) and TNPP was the sum of ANPP and 15 BNPP. We incorporated their 0.50 NPP estimates into GPPDI. 16 17 The GPPDI Database 18 19 In addition to the NPP estimates, land cover, climatic variables, NDVI, and soil texture, 20 all at 0.50 cell resolution were included in the database to assist in interpretation. A global 21 land cover database at 1 km2 spatial resolution (Hansen et al. 2000) was used to 22 determine the cover type in a given 0.50 cell if this information was not provided with the 23 original NPP data. All forest types were grouped in a first pass to determine the majority 21 1 physiognomic type; in a second pass, forested cells were reclassified using the majority 1 2 km2 type. The land cover map was also used to determine the degree of homogeneity 3 within a given 0.50 cell. 4 5 In cases where a large number of NPP estimates were available from the same source and 6 methodology, or where extensive human disturbance was evident, a sub-sample of cells 7 were selected for the GPPDI database in order to maximize the quality of NPP estimates. 8 For example, a total of 2,550 NPP 0.50 cells were available in the Australian continent, 9 thus it was possible to impose a criterion by which the majority class was required to 10 occupy > 90% area of a 0.50 cell. The result was that 615 highly homogeneous cells were 11 included in the database. In China, 99 0.50 forest cells with forest cover 90% were 12 incorporated into GPPDI. In Senegal only 68 NPP 0.5° cells were available, therefore a 13 less restrictive criterion (majority cover type >75% of a total cell area) was applied and 14 32 cells were selected. In Finland and Sweden, 82 NPP cells were selected that met the 15 criteria of being within the coniferous forest zone (i.e. south of 66° N latitude), forest 16 cover >80% within each of the cells, and evergreen needleleaf forest >50% area of each 17 0.50 cell. 18 19 The problem of temporal mismatch between NPP and model forcing variables must be 20 recognized. NPP measurements are often made over a single growing season and ought 21 only to be used in comparison with model outputs for the same period. NPP estimates 22 based on long-term measurements should also use long-term mean climate data to 23 account for inter-annual variations. Unfortunately the lack of adequate data often makes 22 1 it necessary to use less than ideal data sets. Consequently, a quantitative analysis of 2 errors for the data aggregation is difficult, if not impossible. 95% of the cells reported in 3 the GPPDI dataset represented either potential or mean NPP values over periods > 10 4 years (Table 1). 5 6 Data quality 7 8 We identified any outliers in the data by calculating the means and standard deviations of 9 NPP for broad biomes. No NPP values were excluded from the data set since; i) our 10 current data may not be representative for each biome, for example, there are no tropical 11 high-productivity NPP cells in the dataset, thus a cell NPP value of 1,235 gCm-2yr-1 that 12 is currently an outlier, may not necessarily be unreasonable if more high-productivity 13 NPP cells are added; ii) some cells had only a single year of observation while others 14 were long-term averages; and iii) the GPPDI dataset includes both actual and potential 15 NPP. 16 17 Further outliers were detected by examining relationships between NPP and other 18 environmental variables based on: i) climate values outside of the 0.01 to 0.99 percentiles 19 for each biome calculated assuming a normal distribution of variables; ii) linear 20 regression analysis between NPP and actual evapotranspiration (AET), precipitation, 21 elevation, and temperature. Cells falling outside of the 0.95 confidence interval about the 22 regression line were flagged; and iii) managed vegetation (e.g. the cells originally 23 reported as crop) were excluded (Olson et al. 2001). After the outlier analysis, averaging 23 1 replicate measurements of the same cell (mainly in the Great Plains Region, USA), and 2 excluding cropland cells, 2335 cells remained. 3 4 To confirm the data quality, we used all available cells that had TNPP values (3618) for 5 comparisons with the Miami model (precipitation based, (Lieth 1975)). 6 NPP = 3000*(1 - e -0.000664*P) (7) 7 NPP is in gm-2yr-1, and precipitation (P) is the annual total in mm. 8 This model has been adjusted empirically to give a good fit to potential NPP, in the 9 absence of disturbance or management. We also compared the grid-cell NPP with other 10 key environmental variables at the same resolution (0.50), such as AET (1961-90) and 11 NDVI (1981-94) to test the consistency of the estimates. The AET was determined from 12 potential evapotranspiration (PET) or precipitation whichever was smaller. The PET was 13 calculated using equations from Thornthwaite and Mather (Thornthwaite & Mather 1955) 14 based on monthly temperature and precipitation (1961-90) (New et al. 2000). 15 16 For data analyses at biome level, we used the full dataset and cells were assigned to 17 biomes through two steps: i) a gridded land cover map with 0.50 cell size, aggregated 18 from Hansen et al.’s 1 km2 map (Hansen et al. 2000); ii) within the forest and grass 19 categories, climate was used to identify sub-classes as follows: (e.g., minimum monthly 20 temperature >15°C for tropical, -15° to 15°C for temperate, and <-15°C for boreal forest 21 respectively (Prentice et al. 2000). Cells classified as water, bare ground, and urban were 22 not included. Biomes containing <60 cells were also excluded. 23 24 1 Results and Discussion 2 3 The GPPDI data set used here contained 3654 0.5o cells, including 36 cells that had 4 ANPP only, 320 cells had TNPP only, and 3298 cells had both TNPP and ANPP (One 5 cell (22.75 oS, 115.75 oE) along the coast line of NW Australia, however, is not included 6 in the global 62483 climate grid cell). Terrestrial boundaries do not match precisely when 7 the data from different sources with difference in original spatial resolution apply to land 8 mask at 0.5o longitude/latitude (Cramer et al. 1999). The NPP cells were unevenly 9 distributed geographically since we depended entirely on existing data sets. The cells 10 were concentrated in five major regions (Fig. 1) with 77.3% of the cells in the USA, 11 followed by 16.8% in Australia, 2.7% in China, 2.2% in Sweden/Finland, and 0.9% in 12 Senegal. Overall, ANPP values ranged from 3 to 890 gCm-2yr-1, and TNPP values 13 ranged from 3 to 1235 gCm-2yr-1. 14 15 The TNPP values were well correlated (r2=0.73, Fig. 2) with the Miami model (Lieth 16 1975) results, but were consistently lower. The Miami model was tuned to potential NPP 17 based on 52 sites worldwide that were close to potential NPP and possibly more 18 productive than average 0.50 grid cells. Thus, it is understandable why NPP for highly 19 developed agriculture (Prince et al. 2001) and some of potential forest cells (Jenkins et al. 20 2001) matched the Miami model estimates better than cells for less fertilized sites and 21 average forest conditions. If crop cells were excluded the overall relationship between 22 TNPP data and the Miami model results improved slightly with r2 = 0.75. 23 25 1 The NPP observations for temperate evergreen needle-leaf, woodland, closed shrubland, 2 savanna, temperate grassland, and cropland had relatively symmetric distributions (Fig. 3 3). Open shrubland NPP had the most skewed distribution to the lower values probably 4 because 83% of the cells came from central-west Australia where low NPP values were 5 reported (Barrett 2000). Temperate deciduous broad-leaf data tended to be skewed to the 6 higher values because more than 75% of cells were reported from Mid-Atlantic region, 7 USA where data representing the potential NPP were included (Jenkins et al. 2001). 8 9 Comparisons of biomes between TNPP ranges, means, and standard deviation that were 10 derived from the GPPDI dataset and those obtained from entire earth (based on the 11 estimates from the GLObal Productivity Efficiency Model (GLO-PEM) (Prince & 12 Goward 1995) revealed that, in general, the GPPDI dataset had narrow ranges (10 of the 13 12 classes) and smaller standard deviation (9 out of 12). The differences in biome means 14 between the GPPDI-based and global-based were within 30%, defined as (global-based 15 - GPPDI-based)/global-based) for 9 of the 12 biomes. Open shrubland, savanna, and crop 16 had larger deviations (Table 2). 17 18 In general the GPPDI cells had lower mean NPP and narrower ranges than those reported 19 in earlier global compilations of point data (Lieth 1975; Whittaker & Likens 1975) (Table 20 3). Previous compilations often used data from homogeneous sites or stands that were 21 more productive than an average 0.50 grid cell (Kohlmaier et al. 1997). Confidence 22 limits for the earlier compilations were not available. The ranges of these earlier 23 compilations were deduced from the authors' compilation of productivity data; the means 26 1 were in some cases averages of published values, but in many cases they were chosen 2 subjectively as reasonable values for a range as indicated by a few field measurements 3 (Lieth 1975; Whittaker & Likens 1975). 4 5 Eight of the nine biomes had NPP values lower than the minimum values previously 6 reported. Cropland had a higher minimum NPP, which may reflect advances in crop 7 technology during the past 40 years. The crop cells that came from the Midwest region of 8 the USA (Prince et al. 2001) had high productivity as a result of intensive management 9 and highly productive crops. It was noted that most of the remaining crop cells 10 (approximate 75%) were not originally reported as cropland by the original authors, 11 instead, were designated as crop cells only by the 0.50 resolution land cover map (Hansen 12 et al. 2000). These cells were mainly from the Great Plains region of the US (Sala et al. 13 1988; Tieszen et al. 1997; Gill et al. 2001) and the grassland designation was retained 14 since the methods used by the original authors applied to grassland only, even in cells 15 where crop might dominate according to the 0.50 land cover map. If the managed crop 16 cells from Prince et al. (2000) were excluded, the TNPP range for crops would be from 17 143 to 684 gC/m2/yr with TNPP mean of 345 gC/m2/yr, about 10% lower than that 18 reported in the Table 3. 19 20 Our minimum TNPP values in shrubland, grassland, and forests were substantially lower 21 than those of the corresponding biomes previously reported in the literature. Although 22 these differences are important since they indicate wider variation than has previously 23 been recognized at this scale, these differences might have been different if the GPPDI 27 1 grid cells included a systematic sample of all biomes. However, the GPPDI data set was 2 not intended to represent global variation of NPP, rather the aim was to provide data to 3 parameterize, calibrate and validate models that would then be used to more accurately 4 estimate global, spatially explicit variation in NPP. Nevertheless, the fact that our 0.50 5 grid cells mostly extended the range of productivity to lower values than previous 6 compilations of field sites may not simply be by chance and lack of systematic sampling, 7 rather the points selected for earlier compilations were often biased by subjective 8 selection for uniformity, low disturbance, and to typify regions that may not equate to the 9 biomes used here. Thus the GPPDI dataset may include some more realistic values. 10 11 Our results indicated that TNPP of forest worldwide (excluding tropical forest) ranged 12 from 70-1,124 gCm-2yr-1 (Table 3), similar to the range of 95-1,188 gCm-2yr-1 reported by 13 (Whittaker & Likens 1975). Forest productivity, on average, increased with decrease in 14 latitudes ranging from about 301 gCm-2yr-1 for boreal forest to over 600 gCm-2yr-1 in 15 temperate forest (Table 3). Among the forest biomes, temperate deciduous broad-leaf 16 (DBL) was the only biome that had a slightly higher (2.8%) mean value (635 gCm-2yr-1) 17 than Whittaker & Likens' value (618 gCm-2yr-1), mainly because 75% of the NPP cells in 18 that biome came from potential TNPP (Jenkins et al. 2001). 19 20 Within the temperate zone, DBL forest tended to have higher TNPP (635 gCm-2yr-1) and 21 ANPP/TNPP ratio (62%) than evergreen needle-leaf (ENL) coniferous forest (431 gCm- 22 2 23 than ENL coniferous forest (301 gCm-2yr-1). The ANPP/TNPP ratio for boreal ENL was yr-1 and 58%). In the boreal forest zone, mixed forest had higher TNPP (425 gCm-2yr-1) 28 1 about 70%. Most of our boreal cells came from northern China and Finland/Sweden 2 where Gower et al. (Gower et al. 2001) reported ANPP/TNPP ratio of 83% and 60% 3 respectively. 4 5 Mean TNPP for temperate grassland (204 gCm-2yr-1) was almost the same as that for cool 6 grassland (203 gCm-2yr-1) mainly because precipitation (not temperature) controls the 7 productivity, and all grassland cells came from the Great Plains region, USA. The ratio 8 of ANPP/TNPP for temperate grassland was about 43% and decreased to 36% in cool 9 grassland, in agreement with Titlyanova et al. (1999) who reported 30% average ratio for 10 10 Siberian grasslands measured over a 15-year period. 11 12 The wide variation in TNPP observed for temperate ENL (162 gCm-2yr-1) was probably 13 caused by variations in both temperature and precipitation, whereas variation for Savanna 14 (110 gCm-2yr-1) was clearly related to differences in precipitation alone. Woodland also 15 showed a high variation in TNPP (142 gCm-2yr-1) probably because it includes either 16 forest with lower canopy cover or degraded forests that can occur at any latitudes. At the 17 biome level, open shrubland had the lowest TNPP (57 gCm-2yr-1). Most of these cells 18 came from Australian central-western deserts where 60% of the cells have annual total 19 precipitation between 158-300 mm. 20 21 At the biome level, cool grassland had the lowest standard deviation of annual mean 22 temperature (1.2o C) and the lowest standard deviation of annual total precipitation (56 23 mm), resulting in lower variation in TNPP. The highest variation in TNPP occurred in 29 1 temperate ENL forest with high standard deviation for both temperature (5.7o C) and 2 precipitation (383 mm). At the biome level, variations in TNPP were generally 3 correlated with variation in climate, but more strongly with variation in precipitation 4 alone (r = 0.94, p=0.001, Fig. 4a) compared with temperature alone (r = 0.56, p=0.1, Fig. 5 4b). 6 7 TNPP in the GPPDI dataset increased with increases in AET, a measure that combines 8 the availability of heat and water, (r2 = 0.34, Fig. 5a). There was a much strong linear 9 relationship between TNPP and annual mean NDVI (Fig. 5b) with r2 = 0.65. A statistical 10 analysis of TNPP values using AET and NDVI accounted for 90% of the variance (data 11 not shown). These comparisons indicate that our NPP dataset is consistent with key 12 climate and ecosystem indicators, suggesting the cell NPP data are reliable for the biomes 13 that are represented. 14 15 The Mitchell grassland in Queensland, Australia (1960-97) had the greatest inter-annual 16 variation in NPP with a coefficient of variation (CV) of 45%, followed by tropical 17 Savanna in Senegal (1987-97) with a CV of 21%, and the cropland in Iowa, USA (1982- 18 96) with a CV of 16% (Fig. 6). Similar CV values were observed for period (1987-96) 19 during which NPP estimates for all 3 studies were available (38%, 22%, and 17%, 20 respectively). Statistical analyses revealed that inter-annual NPP variation trends in 21 Queensland and Iowa were significantly correlated (p=0.05) from 1987 to 1996. Both 22 seem to be correlated with ENSO cycles. Whereas the inter-annual NPP variation pattern 23 in Senegal differed significantly (p=0.05) from either of the other two studies during the 30 1 same period, indicating although ENSO phenomena represent global trends, spatial and 2 temporal variations can differ at regional scales. 3 4 After removing the outliers (including the cells originally reported as cropland) and 5 averaging all replicate measurements for the same cell (Olson et al. 2001), 2335 0.50 x 6 0.50 cells remained. The frequency distributions of data in the full and the edited 7 databases are given in Fig. 7. Statistical analyses indicated that the cumulative frequency 8 curves between the 2 datasets were not significantly different (p=0.05), but the frequency 9 distributions of the 2 datasets did differ significantly. An F test suggests that the 10 variances of TNPP between the 2 datasets do not differ (p>0.05). For the edited NPP 11 data set (n=2335), the general relationships between NPP and estimates from the Miami 12 model, annual AET, and annual mean NDVI were similar to those using the full dataset 13 (n=3618) with slight improvements as expected. For example, r2 values increased for the 14 Miami model from 0.73 to 0.78, for AET from 0.34 to 0.38, and for NDVI from 0.65 to 15 0.77, respectively. 16 17 Several issues were raised in assembling grid cell NPP estimates for global modeling. 18 First, numbers of field observations falling in a given 0.50 cell are generally inadequate 19 for simple aggregation. This is a serious problem especially when land cover types are 20 highly heterogenous and so there may be no measurements at all for some cover types in 21 the cell. Second, some resource management data contain economic products only, thus, 22 provide incomplete NPP measurements. We suggest that when estimating unknown from 23 known components of NPP only equations established under the same environmental 31 1 conditions be used because large errors may be introduced when different regression 2 equations or other site-specific rules are used in regions for which they were not 3 developed. It is also noted that the errors in measured components accumulate in 4 unknown components of NPP that are estimated from measured ones. Third, it is not 5 uncommon that the grid cell vegetation type designated by a map or remotely sensed data 6 may differ from actual field site records that are more accurate. This discrepancy may be 7 caused by heterogeneity of the cell, differences in the vegetation, classification schemes, 8 and other factors. However we suggest that cell identifications should be used because 9 they represent the same entity that models simulate. Fourth, a quantitative analysis of 10 errors in aggregating grid-cell NPP is difficult, if not impossible. Larger errors than those 11 reported from original studies (where available) are to be expected if extra procedures 12 were performed beyond the original studies. 13 14 Conclusions 15 16 This work has identified 3654 specific 0.50 grid cells with NPP (36 cells have ANPP 17 only, 320 cells have TNPP only, and 3298 cells have both TNPP and ANPP) and certain 18 other variables in which the available field measurements have been aggregated using 19 explicit methods. The ranges and means of our results generally agree with previous 20 global NPP summaries and correlate with climate, but they have the added advantage of 21 large numbers of sites and site areas appropriate for application to global biogeochemical 22 modeling. 23 32 1 An edited dataset (2335 cells) was also generated based on the full dataset. In the edited 2 dataset, cells identified as outliers or originally reported as cropland were excluded and 3 each unique geographic location contains a single NPP value. Statistical analyses suggest 4 no significant difference between the two datasets in terms of cumulative frequency 5 distributions. Relationships between the TNPP observations and key environmental 6 variables were slightly improved using the edited dataset, compared to the full dataset. 7 While the full dataset is more suitable for biome level statistical analyses because of 8 larger sample size, the edited dataset is more appropriate for model/data inter- 9 comparison. 10 11 Although not intended to be a comprehensive, globally representative survey, the results 12 show that much of the Earth's vegetation has NPP values well below the potential rates 13 found in rare, favorable conditions. Although these unusual situations have been 14 emphasized in the past, when the focus was on natural mechanisms rather than global 15 consequences of increasing anthropogenic impacts, the GPPDI data set provides a more 16 representative sample of the world's terrestrial vegetation. 17 18 In general the number and quality of field observations of TNPP or ANPP in sites 19 suitable for scaling to 0.50 grid cells, together with climate, soils and vegetation data, fall 20 far short of our current needs for biogeochemical modeling. Short-term measurement 21 campaigns are not ideal, rather the establishment of multi-purpose, regular ecosystem 22 monitoring programs, would be more desirable. Routine forest, crop and rangeland 23 monitoring currently provide the most appropriate NPP measurement for large areas 33 1 because they take explicit account of spatial variability. Relatively minor modifications 2 of these programs would be the most cost-effective method of monitoring field NPP data 3 for large areas. Relatively few additional data are needed, such as root:shoot ratios, 4 harvest indices, and other conversion parameters. Laws and practices that restrict access 5 to inventory data for private land need to be reconsidered urgently, so multiple purposes 6 can be served by these publicly financed programs without infringing privacy. A serious 7 effort is needed to augment the data available in the developed world with equivalent data 8 for less developed countries. 9 10 The effects of land management and utilization on NPP are omitted from many current 11 biogeochemical models. To enable the actual, rather than potential vegetation to be 12 studied, field measurements of soils and land use would significantly increase the value 13 of routine monitoring. Since rapid, routine field techniques rarely measure all 14 components of NPP, further studies on the partition of biomass between pools for 15 different vegetation and climate types are also needed. 16 17 Acknowledgements 18 19 Support for this research was provided by NASA grant NAG 58121 to Dr. Prince. We 20 are grateful to the many individuals and groups cited in the text that have provided 21 original data, and also to the International Geosphere-Biosphere Programme (IGBP), 22 Data and Information System (DIS) for their support and contributions to the GPPDI. 23 We thank Richard Olson at the Oak Ridge National Laboratory for assistance in outlier 24 analysis and data compilation. Part of this work resulted from three workshops 34 1 conducted at the National Center for Ecological Analysis and Synthesis, a Center funded 2 by the U.S. National Science Foundation (Grant # DEB-9421535), the University of 3 California at Santa Barbara, and the State of California.References 4 Aber J, Neilson RP, McNulty S et al. (2001) Forest processes and global environmental changes: predicting 5 6 7 8 9 10 the effects of individual and multiple stressors. BioScience, 51, 735-751. AUSLIG (1990) Atlas of Australian resources: vegetation, volume 6 Australian Survey and land Information Group, Department of Administrative Services, Commonwealth Government Printer Barrett DJ (2000) Stead state net primary productivity, carbon stocks and mean residence time of carbon in the Australian terrestrial biosphere. Global Biogeochemical Cycles, Submitted. Birdsey RA, Schreuder HT (1992) An overview of forest inventory and analysis estimation procedures in 11 the Eastern United States--with an emphasis on the components of change USDA Forest Service, 12 Rocky Mountain Forest and Range Experiment Station GTR RM-214. 13 14 15 16 Bray JR, Gorham E (1964) Litter production in forests of the world. Advances in Ecological Research, 2, 101-157. Brown SL, Schroeder P, Kern JS (1999) Spatial distribution of biomass in forests of the eastern USA. Forest Ecology and Management, 123, 81-90. 17 Cheng S, Zhao Y (1990) Remote sensing and Geo-sciences analysis. Measuring Press, Beijing. 18 Clark DA, Brown S, Kicklighter DW et al. (2001) NPP in tropical forests: an evaluation and synthesis of 19 20 21 22 23 24 the existing field data. Ecological Applications, 11, 371-384. Cohen WB, Spies TA, Fiorella M (1995) Estimating the age and structure of forests in a multi-ownership landscale of western Oregon, USA. International Journal of Remote Sensing, 16, 721-746. Coughlan JC, Running SW (1989) An expert system to aggregate biophysical attributes of a forested landscape within a Geographic Information System. AI Applications, 3, 35-43. Cramer W, Kicklighter DW, Bondeau A et al. (1999) Comparing global models of terrestrial net primary 25 productivity (NPP): overview and key results. Global Change Biology, 5 (suppl. 1), 1-15. 26 Cramer W, Olson RJ, Prince SD (2001) Global productivity: determining present and future. In: Terrestrial 27 global productivity: past, present, and future (Eds, Mooney HA, Saugier B, Roy J) Academic 28 Press, San Diego. 35 1 Day KA, McCaskill MR, McKeon GM (1997) A review of pasture production data in Queensland. In: 2 Evaluating the risk of pasture and land degradation in native pastures in Queensland (Eds, Day 3 KA, McKeon GM, Carter JO). 4 5 Diallo O, Diouf A, Hanan NP et al. (1991) AVHRR mornitoring of savanna primary production in Senegal, West Africa: 1987-1988. International Journal of Remote Sensing, 12, 1259-1279. 6 ERDAS (1999) ERDAS IMAGINE V8.4: Tour Guides. ERDAS Inc., Atlanta, GA. 7 Forest Ministry of China (1994) Statistics of forest resources of China: 1989-93. Forestry Publisher, 8 9 10 11 12 13 14 15 16 17 Beijing. Franklin JF, Dyrness CT (1990) Natural vegetation of Oregon and Washington. Oregon State University Press, Corvallis, OR. Gholz HL, Hawk GM, Cambell A et al. (1985) Early vegetation recovery and element cycles on a clear-cut watershed in western Oregon. Canadian Journal of Forest Research, 15, 400-409. Gill RA, Kelly RH, Parton WJ et al. (2001) Using simple environmental variables to estimate belowground productivity in grasslands. Global Ecology and Biogeography, In Press. Goetz SJ, Prince SD (1996) Remote sensing of net primary production in boreal stands. Agricultural and Forest Meteorology, 78, 149-179. Goward SN, Dye DG (1996) Global biospheric monitoring with remote sensing. In: The use of remote 18 sensing in the modeling of forest productivity (Eds, Gholz HL, Nakane K, Shimoda H) Kluwer 19 Academic Publishers, Dordrecht, pp. 241-272. 20 21 22 Gower ST, Krankina ON, Olson RJ et al. (2001) Net primary production and carbon allocation patterns of boreal forest ecosystems. Ecological Applications, 11, 1395-1411. Grier CC, Logan NS (1977) Old-growth Pseudosuga menziesii communities of a western Oregon 23 watershed: biomass distribution and production budgets. Ecological Monographs, 47, 373-400. 24 Hame T, Salli A, Andersson K et al. (1997) A new methodology for the estimation of biomass of conifer 25 dominated boreal forest using NOAA AVHRR data. International Journal of Remote Sensing, 18, 26 3211-3243. 27 28 Hansen HM, Frieswyk T, Glover JF et al. (1992) The eastwide forest inventory data base: user manual USDA, Forest Service, North Central Forest Experiment Station GTR-NC-151. 36 1 Hansen MC, Defries RS, Townshend JR et al. (2000) Global land cover classification at 1 km spatial 2 resolution using a classification tree approach. International Journal of Remote Sensing, 21, 1331- 3 1364. 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Harvey LDD (2000) Box models of the terrestrial biosphere. In: The carbon cycle (Eds, Wigley TML, Schimel DS) Cambridge University Press, New York, pp. 238-247. Henderson-Sellers A, Wilson MF, Thomas G et al. (1986) Current global land-surface data sets for use in climate-related studies National Center for Atmospheric Research 272. Javis PG, Leverenz JW (1984) Productivity of temperate, decidudous and evergreen forests. In: Physiological Plant Ecology (Eds, Lange O, Nobel P, Osmond C, Ziegler H) Springer-Verlag, New York, pp. 233-280. Jenkins JC, Birdsey RA, Pan Y (2001) Biomass and NPP estimates for the Mid-Atlantic Region (USA) using plot-level forest inventory data. Ecological Applications, 11, 1174-1193. Jiang H, Apps MJ, Zhang Y et al. (1999) Modelling the spatial pattern of net primary productivity in Chinese forest. Ecological Modelling, 122, 275-288. Joyce LA, Chalk DE, Vigil A (1986) Range forage database for 20 Great Plains, Southern, and Western States USDA Forest Service General Technical Report RM-133 RM-133. Kohlmaier GH, Badeck FW, Otto RD et al. (1997) The Frankfurt Biosphere Model: A global process- 18 oriented model of seasonal and long-term CO2 exchange between terrestrial ecosystems and the 19 atmosphere .2. Global results for potential vegetation in an assumed equilibrium state. Climate 20 Research, 8, 61-87. 21 22 23 24 25 26 27 28 Krankina ON, Brown S, Jenkins JC et al. (1999) Use of forest inventory data for estimating NPP. Ecological Applications, Submitted. Lieth H (1975) Modelling the primary productivity of the world. In: Primary productivity of the biosphere (Eds, Lieth H, Whittaker, R.H.) Springer-Verlag, New York, pp. 237-263. Lonsdale WM (1988) Predicting the amount of litterfall in forests of the world. Annuals of Botany, 61, 319324. Lurin B, Cramer W, Moore B et al. (1994) Global terrestrial net primary productivity. Global Change Newsletter (IGBP), 19, 6-8. 37 1 2 3 4 5 6 7 Martinez-Yrizar D, Sarulhan J (1990) Litterfall patterns in a tropical decidudous forest in Mexico over a five-year period. Journal of Tropical Ecology, 6, 433-444. Means JE, Hansen HA, Koerper GJ et al. (1994) Software for computing plant biomass-BIOPAK user guide USDA Forest Service, Pacific Northwest Research Station PNW-GTR-340. Meentemeyer V (1982) World patterns and amounts of terrestrial plant litter production. BioScience, 32, 125-128. New MG, Hulme M, Jones PD (2000) Representing 20th century space-time climate variability II: 8 development of 1901-96 monthly grids of terrestrial surface climate. Journal of Climate, 13, 2217- 9 2238. 10 Northcote KH, Hubble GD, Isbell RF et al. (1975) A description of Australian Soils CSIRO Publishing 11 Olson RJ, Johnson K, Zheng D et al. (2001) Global and regional ecosystem modeling: databases of model 12 drivers and validation measurements Environmental Sciences Division, Oak Ridge National 13 Laboratory TM-196. 14 15 16 Prentice IC, Heimann M, Sitch S (2000) The carbon balance of the terrestrial biosphere: Ecosystem models and atmospheric observations. Ecological Applications, 10, 1553-1573. Prince SD (1991) Satellite remote sensing of primary production: comparison of results for Sahelian 17 grasslands 1981-1988. International Journal of Remote Sensing, 12, 1301-1311. 18 Prince SD, Goward SN (1995) Global primary production: a remote sensing approach. Journal of 19 20 21 22 23 24 25 26 Biogeography, 22, 815-835. Prince SD, Haskett J, Steininger M et al. (2001) Net primary production of U.S. Midwest croplands from agricultural harvest yield data. Ecological Applications, 11, 1194-1205. Prince SD, Olson RJ, Dedieu G et al. (1995) Global primary production data initiative project description IGBP-DIS Working paper No. 12. Raich JW, Nadelhoffer KJ (1989) Belowground carbon allocation in forest ecosystems; global trends. Ecology, 70, 1346-1354. Rastetter EB (1996) Validating models of ecosystem responses to global change. BioScience, 46, 190-198. 38 1 Ruimy A, Kergoat L, Bondeau A et al. (1999) Comparing global models of terrestrial net primary 2 productivity: analysis of differences in light use efficiency. Global Change Biology, 5 (Supp. 1), 3 56-64. 4 5 6 Sala OE, Parton WJ, Joyce LA et al. (1988) Primary production of the central grassland region of the United States. Ecology, 69, 40-45. Schimel D, Enting IG, Heimann M et al. (1995) CO2 and the carbon cycle. In: Climate Change 1994: 7 Radiative Forcing of Climate Changeand an Evaluation of the IPCC IS92 Emission Scenarios 8 (Eds, Houghton J, Meira-Filho, LG, Bruce, J, Lee, H, Callander, BA, Haites, E, Harris, N, 9 Maskell, K) Cambridge University Press, New York, pp. 35-72. 10 Thornthwaite CW, Mather JR (1955) The water balance. Climatology, 8, 1-86. 11 Tieszen L, Reed BC, Dejong DD (1997) NDVI, C3 and C4 production, and distributions in the Great Plains 12 13 14 grassland cover classes. Ecological Applications, 7, 59-78. Titlyanova AA, Romanova LP, Kosykh NP et al. (1999) Patterns and process in above-ground and belowground components of grassland ecosystems. Journal of Vegetation Science, 10, 307-320. 15 Turner DP, Cohen WB, Kennedy RE (2000) Alternative spatial resolutions and estimates of carbon flux 16 over a managed forest landscape in Western Oregon. Landscape Ecology, 15, 441-452. 17 Turner DP, Koerper GJ, Hanmon ME et al. (1995) A carbon budget for forests of the conterminous United 18 19 20 States. Ecological Applications, 5, 421-436. Turner J, Long JN (1975) Accumulation of organic matter in a series of Douglas-fir stands. Canadian Journal of Forest Research, 5, 681-690. 21 USDA (1981) Land resource regions and major land resource areas of the United States USDA 296. 22 VEMAP (1995) Vegetation/ecosystem modeling and analysis project: Comparing biogeography and 23 biogeochemistry models in a continental-scale stuyd of terrestrial ecosystem responses to climate 24 change and CO2 doubling. Global Biogeochemical Cycles, 9, 407-437. 25 Vogt K (1991) Carbon budgets of temperate forest ecosystems. Tree Physiology, 9, 69-86. 26 Waring RH, Schlesinger WH (1985) Forest ecosystems: concept and management. Academic Press, 27 Orlando. 39 1 2 3 4 5 6 Wharton EH, Alerich AL, Drake DA (1997) Estimating total forest biomass in New York, 1993 USDA Forest Service, Northeastern Forest Experimental Station NE-139. Wharton EH, Griffith DM (1998) Estimating total forest biomass in Maine, 1995 USDA Forest Service, Northeastern Forest Experimental Station NE-142. Whittaker RH, Likens GE (1975) The biosphere and man. In: Primary production of the biosphere (Eds, Lieth H, Whittaker RH) Springer-Verlag, New York, pp. 305-328. 7 Wolter PT, Mladenoff DJ, Host GE et al. (1995) Improved forest classification in the northen lake states 8 using multi-temporal Landsat imagery. Photogrammetric Engineering and Remote Sensing, 61, 9 1129-1143. 10 11 12 13 Woods KD, Fieveson AH, Botkin DB (1991) Statistical error analysis for biomass density and leaf area index estimation. Canadian Journal of Forestry Research, 21, 974-989. Woudenberg SW, Farrenkopf TO (1995) The westwide forest inventory data base: users manual. USDA Forest Service Intermountain Research Station INT-GT-317. 14 Wu Z (1980) Chinese vegetation. Science Press, Beijing. 15 Zheng D, Prince SD, Hame T (2001) Estimating gridded net primary production of boreal forests in 16 Finland and Sweden from field data and remotely sensed biomass. Journal of Vegetation Science, 17 Submitted. 18 19 40 1 Figure captions 2 3 Figure 1. Spatial distribution of 0.50 NPP cells in the current GPPDI: a) TNPP in the 4 Great Plains region, USA from 3 different studies; b) crop TNPP in the 5 Mid-West region, USA; c) forest TNPP in the eastern USA, HJ. Andrews 6 OR, Minnesota, and Greater Yellowstone Ecosystem (ANPP only), USA; 7 d) potential forest TNPP in the Mid-Atlantic region and the state of Maine, 8 USA; e) ANPP in Senegal; f) TNPP of conifer-dominated boreal forest in 9 Sweden, Finland, and Russia; g) forest TNPP in China; and h) potential 10 TNPP in Australia. 11 12 Figure 2. Comparison between TNPP values reported in the current GPPDI dataset 13 and the TNPP values estimated from the Miami model (Lieth 1975), 14 precipitation based) at 0.50 cell size. 15 16 Figure 3. Data distributions of 0.50-cell TNPP in different biomes. The central box 17 shows the data between the first (25%) and third (75%) quartiles, with the 18 median represented by a thick line. The dotted lines extend to brackets 19 showing 1.5 times the NPP ranges between the 25% and 75% quartiles 20 or an extreme value, whichever is less. Individual extreme values outside 21 the brackets are shown as plain bars. T=Temperate, ENL=Evergreen 22 Needleleaf, B=Boreal, DBL=Deciduous Broadleaf, MIX=Mixed forest, 23 Cl=Closed, O=Open, and C=Cool. 41 1 2 Figure 4. Relationships between standard deviations (sd) of TNPP and climatic 3 variables at biome level. a) sd of TNPP vs. sd of annual mean total 4 precipitation; and b) sd of TNPP vs. sd of annual mean temperature. Each 5 point represents a biome. 6 7 Figure 5. Comparisons between TNPP values in the GPPDI dataset and 8 environmental variables at 0.50 cell size: a) annual AET, and b) annual 9 mean NDVI at 0.50 cell size. 10 11 Figure 6. Inter-annual NPP variation in 3 studies: a) Mitchell grassland in 12 Queensland, Australia (ANPP, 1960-97); b) Cropland in state of Iowa, 13 USA (TNPP, 1982-96); and c) Tropical Savanna in Senegal (ANPP, 1987- 14 97). The horizontal lines show the mean NPP values over the entire 15 periods. 16 17 Figure 7. Frequency distributions (thin lines) and cumulative frequency distributions 18 (thick lines) for grid-cell TNPP datasets. a) All available cells including 19 replicate measurements and, b) after removing the outliers and averaging 20 replicate TNPP measurements so that each unique geographic cell has a 21 single TNPP value. 42 1 2 3 4 Table 1. Data sets used to develop 0.50 grid cell estimates of NPP. The availability of aboveground NPP (AN) and total NPP (TN) is noted (y=yes, n=no, and e=estimated), and the NPP type (A=actual and P=potential). Studies are grouped according to the field methodology and senior author of each of the 15 sources. Estimation method Vegetation Type AN TN Region Inventory by county Temperate forests e e Eastern USA Inventory, FIA 2711 plots Boreal forest y y Minnesota, USA Inventory, FIA 2640 plots Temperate mixed forest y y Inventory, 124 stands in Russia, by county in USA y y Middle-Atlantic and Maine,USA Porozhski, Russia, SE And NE of USA n y Mid West, USA y n y y Great Plains, USA HJ Andrews, OR, USA y n y n y y Senegal China Finland and Sweden Diallo et al. (1991) Jiang et al. (1999) Zheng et al. (2001) y y y n Australian Continent Queensland, Australia Barrett (2000) Day et al. (pers. comm. 1999) Hansen et al. (2000) Boreal forest, Temperate deciduous. & mixed forests Inventory, NASS2 by county Cropland Inventory, rangeland Literature, RS3 RS and field data RS and field data RS and field data Local modeling Local modeling Grassland Temperate coniferous forest Tropical savanna Forests Conifer-dominated boreal forest Varied Grassland Regression analysis 5 6 Data source NPP Type A Number of cells 62 A 2 P 245 A 8 A 220 P A 922 1 1987-97 Average Average A A A 32 99 82 Potential NPP 1959-98 P A 615 1 1991 A 3 Average Average P P 1262 100 FIA1 latest survey interval (varying with each State) Goetz et al. (1996) FIA latest survey interval Jenkins et al. (2001) Potential NPP Brown et al. (1999) Krankina et al. (1999) Varied Prince et al. (2001) 1992 for Midwest States, mean of 198296 for Iowa Tieszen et al. (1997) Average Turner et al. (1999) 1988 Temperate coniferous y n Yellowstone, USA forest Regression analysis Grassland y y Great Plains, USA Gill et al. (2001) Regression analysis Grassland y e Great Plains, USA Sala et al. (1988) Notes: 1 Forest Inventory and Analysis; 2 National Agricultural Statistics Service, 3 Remote Sensing. 43 Year(s) 1 2 3 4 Table 2. Comparison of total NPP (TNPP) ranges, means, and standard deviation (STD) for biomes between GPPDI grid cells and global estimates obtained from the GLObal Productivity Efficiency Model (GLO-PEM) (Prince & Goward 1995). Biomes based on global land cover map (Hansen et al. 2000) with modification (see text). ----------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------Biome type TNPP (gC/m2/yr) Range Mean STD GPPDI GLO-PEM GPPDI GLO-PEM GPPDI GLO-PEM Temperate ENL1 forest 77-746 0-1359 432 449 162 226 Temperate DBL2 forest 151-1124 289-1298 635 698 107 160 Temperate mixed forest 239-987 168-1135 577 480 120 111 Boreal ENL forest 70-546 144-494 301 310 68 55 Boreal mixed forest 162-690 129-554 425 363 114 50 Temperate grassland 71-412 0.2-1172 204 282 57 173 Cool grassland 90-382 0.1-443 203 193 52 80 Woodland 79-824 0-2106 455 481 142 333 Closed shrubland 73-306 30-913 184 205 48 109 Open shrubland 4-380 0-1968 57 104 52 92 Savann 9-646 0.6-2605 300 809 110 384 143-712 23-1765 398 577 128 342 Crop 5 ----------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------- 6 Notes: 1 = Evergreen Needleleaf; 2 = Deciduous Broadleaf. 7 44 1 2 3 4 5 Table 3. Summary of the observed GPPDI 0.5 0 grid-cell total net primary production (TNPP) and related environmental characteristics with sampling sizes at least > 60 cells. Values in the square brackets are those reported from Lieth (1975) and Whittaker and Likens (1975) at biome level, where available. Values in the curly braces following the TNPP mean are aboveground NPP ratio of the TNPP (%) and number of sites used for the calculation. ----------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------Biome type TNPP (gC/m2/yr) Annual mean temperature (0C) Annual total Precipitation (mm) (Sample size) Range Mean (ratio, No.) STD1 Range Mean STD Range Mean STD 6 Temperate ENL2 forest 77-746 432 {58, 79} 162 1.5-20.0 12.0 5.7 381-2125 1026 383 (91) [285-1188] [618] Temperate DBL3 forest 151-1124 635 {62, 151} 107 5.3-19.5 11.4 2.8 832-1958 1100 171 (159) [285-1188] [618] Temperate mixed forest 239-987 577 {62, 54} 120 4.4--18.4 7.8 2.5 636-1650 1063 176 (63) Boreal ENL forest 70-546 301 {70, 65} 68 -3.2-5.4 2.6 1.5 349-802 600 97 (67) [190-950] [380] Boreal mixed forest 162-690 425 {59, 83} 114 -6.2-6.5 1.0 3.8 427-1325 759 251 (136) Temperate grassland 71-412 204 {43, 525} 57 3.1-20.6 11.7 3.7 302-878 456 101 (525) [90-675] [270] Cool grassland 90-382 203 {36, 376} 52 2.3-8.4 6.3 1.2 288-605 391 56 (376) Woodland 79-824 455 {52, 198} 142 -5.9-20.5 11.9 5.6 398-1607 955 275 (223) [95-475] [285] Closed shrubland 73-306 184 {46, 133} 48 9.7-21.3 15.3 2.5 294-657 468 99 (133) [119-570] [333] Open shrubland 4-380 57 {55, 617} 52 5.0-27.1 21.7 5.2 158-603 304 85 (618) [119-570] [333] Savanna 9-646 300 {57,346} 110 15.1-28.2 20.8 3.6 233-1628 849 236 (346) [90-900] [405] Crop 143-712 398 {47, 630} 128 2.6-23.2 9.6 4.2 338-1568 733 224 (831) [45-1800] [293] ----------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------- 7 Notes: 1 = Standard Deviation; 2 = Evergreen Needleleaf; 3 = Deciduous Broadleaf. 8 45