Background & Summary

Forests reduce the impact of climate change, which is the greatest environmental challenge of the 21st century1, by capturing CO2 from the atmosphere, and acting as guardians of a carbon store in their wood, roots, and soils. However, there are major uncertainties as to the amount of carbon stored in forests, and how this is changing through time2. The carbon stored in forests changes as a combination of the processes of deforestation, degradation and growth driven by both natural and anthropogenic disturbances, and climate change itself. This leads to fluxes in stored carbon, making forest carbon stock one of the greatest sources of uncertainty in the global carbon cycle2,3,4. It is important to reduce these uncertainties to improve our understanding of the carbon cycle and thus various elements of Earth system and climate models. This reduction is also crucial to enable policies to conserve and increase forest carbon storage to be designed and monitored.

Modern remote sensing data, cloud computing, and machine learning capacity, enable the production of high quality maps of AGB, the largest carbon pool in most forests, and a carbon pool that can change substantially over time5. This increasing capacity to generate high quality maps of AGB has been shown by the recent release of a large number of global products6,7,8,9. However, such global products can have high uncertainties at a local or regional level, and often perform poorly when compared to field plots9,10,11. Given this, there is clearly a place for regional or country-specific maps where datasets and methods are tuned to local conditions, and that perform well against independent validation datasets12.

China covers 6.3% of the world’s land surface, has 18% of its people (The World Bank, 2021), and an estimated 5.4% of its forest13. As forests play such an important role in the global climate system and in terrestrial ecosystems, what is happening in China’s forests is important to the global climate system. The area of China’s forest was 102 million hectares (ha) in 1949, and it remained at 106 million ha until 1994–199814, and then increased dramatically to 220 million ha in 2020, accounting for 23% of China’s area13. Typically, substantial increases in forest area lead to increases in total AGB. However in China’s case, it is possible that higher-biomass, old-growth forests are being replaced with new, lower-biomass forests or plantations15,16.

Currently, information of forest AGB can be derived from field inventory data and Earth Observation (EO) data. Conventionally, forest AGB estimation requires inventory forest variables such as tree species, diameter at breast height (DBH), and tree height, measured in fixed area forest inventory plots, which are converted into estimates of tree biomass through allometric equations17. Although conventional inventory techniques generally provide accurate estimates of forest AGB at plot level18, it is time-consuming, only possible to collect data covering a very small area, requiring good statistical stratification and placement of a large number of field plots in order for them to represent the wider forest, and sometimes impractical due to inaccessibility. It is obviously impossible in the present time to add sampling plots or locations retroactively for historical data collection. On the other hand, EO data, from satellites or aircraft, has shown promising capabilities for consistent and systematic observations of the dynamics of forest ecosystems19,20,21,22. However, producing accurate EO-based results relies on consideration of local factors, ideally using local field calibration data as well as appropriate EO data sources.

Some studies of AGB in China have been conducted using satellite data. However, these studies produce very different estimates, due to differences in their prediction models and dataset sources (Table 1). The total carbon estimates from the map of Su23 are four times higher than that of GEOCARBON6. While these maps use different forest cover maps to mask non-forest areas, resulting in variations in forest area, and the maps are for slightly different periods, this disparity in area and time cannot explain much of the observed difference.

Table 1 Total carbon stored in forest AGB in China from a number of studies, along with the area covered by the maps.
Full size table

Therefore, our study aims to produce a reliable 2007 AGB map for China. We chose to map 2007 for a number of reasons, summarised as (i) it is a useful baseline year to compare against forest policy and (ii) a variety of remote sensing and field validation datasets is available around that date. The Asia-Pacific Network for Sustainable Forest Management and Rehabilitation (APFNet) was proposed in 2007, and it further marks the acceleration of the China’s reforestation/afforestation programmes, as well as the implementation of forest protection laws24. From the remote sensing standpoint, 2007 is where spaceborne LiDAR and suitable optical and radar satellite data are available. LiDAR data from satellites provide a direct estimate of height, which is closely related to biomass, unlike other satellite sensors which rely on indirect estimation, with the shape of relationships often varying from place to place25. The Ice, Cloud, and land Elevation Satellite (ICESAT) Geoscience Laser Altimeter System (GLAS) operated throughout the 2000s, including collecting lots of data in 2007, and provides a large number of ~65 m circular footprints with estimates of height, used to train and test the wall-to-wall maps of AGB we produce. In order to extrapolate across the landscape, we need Synthetic Aperture Radar (SAR) and optical datasets. The C-band Environmental Satellite (Envisat) Advanced Synthetic Aperture Radar (ASAR) and Advanced Land Observing Satellite (ALOS) Phased Array L-band SAR (PALSAR) were available around 2007. C-band SAR is capable of penetrating into the canopy, and L-band SAR can penetrate deeper into the canopy and thus obtain information of larger element of trees such stems26. Landsat-5 TM data was also used in the prediction, as optical data can capture detailed spectral information of forest canopy.

In this study, we introduced a new 2007 AGB map for China using high resolution multisource remote sensing data. Our model was trained specifically for China, considering its unique allometric equation and terrain factors. This dataset can serve as the historical baseline dataset for forest AGB change studies, as it outperforms existing maps when compared to two unseen field validation datasets.

Methods

We generated point estimates of AGB from ICESat GLAS Lorey’s height data27, using local allometric relationships developed independently for northern and southern China using field data we collected in 40 forest inventory plots in China. We then predict AGB spatially to create a wall-to-wall map at a resolution of 50 m using a random forest model with L-band SAR from ALOS PALSAR, C-band SAR from Envisat ASAR, and optical satellite data from Landsat-5, all acquired around 2007, in addition to a Digital Elevation Model from the Shuttle Radar Topography Mission (SRTM) (collected in 2000). Both forest structure and remote sensing responses are different in sloped compared to flat areas28,29, so we generated continuous AGB maps for flat regions and rugged regions separately, to enable different prediction models for the two and thus reduce the errors in AGB estimates. Due to gaps in the coverage by Envisat ASAR (1.2% of China was missed), we generated a second AGB map with slightly lower accuracy to use in the regions without Envisat ASAR data (Fig. 1). As a last step, we combined the maps for flat region (slope ≤10°), rugged region (slope >10°) and region without Envisat data, to produce a final map for validation (Fig. 2).

Fig. 1
figure 1

The spatial distribution of the three forested regions.

Full size image
Fig. 2
figure 2

Overall workflow of forest AGB estimation and validation.

Full size image

Further detail about all these steps is given below, and code is available.

Field data

Field data collected in 2021

26 field plots in northeastern China and 14 field plots in Southwestern China were measured in 2021 (Fig. 3). Each of the field plot was circular with a radius of 12.5 m (~0.05 ha). We measured tree height using Vertex IV and Transponder T3, and measured DBH of all the trees with DBH ≥5 cm. These measurements were used to estimate the AGB of each tree, utilizing the following biomass equation for China30:

$$W=0.1355\times {\left({D}^{2}H\right)}^{0.817}$$
(1)

Where, W is the AGB of each tree (kg), D is DBH (cm) and H is height of the tree (m). We summed the AGB of every tree to estimate total AGB per plot, and converted the total biomass to AGB density at the 1 ha scale (Mg ha−1).

Fig. 3
figure 3

Spatial distribution of field data collected for this study in 2021 and compiled from other studies from 2005 to 2015.

Full size image

We also calculated Lorey’s height of each plot using Eq. (2). Lorey’s height is an index of height whereby individual trees are weighted in proportion to their basal area, which enhances the significance of larger trees in the forest, and is closely correlated with AGB25,31,32.

$$Lorey=\frac{{\sum }_{i=1}^{N}B{A}_{i}{h}_{i}}{{\sum }_{i=1}^{N}B{A}_{i}}$$
(2)

where, BAi represents the basal areas, and hi represents the canopy height of the individual trees.

We used these plot-level measurements to develop allometric equations between the field measured Lorey’s height and AGB in northern China and southern China independently (Fig. 4).

$$AG{B}_{n}=0.5014\times Lore{y}^{1.8762}$$
(3)
$$AG{B}_{s}=0.1237\times Lore{y}^{2.3281}$$
(4)

where, AGBn is AGB density (Mg ha−1) of field plots in northern China, AGBs is AGB density of field plots in southern China and Lorey is Lorey’s height. The allometric equation for northern China showed an R2 of 0.75 and an RMSE of 35.58 Mg ha−1, while for southern China, the R2 was 0.50 with an RMSE of 22.96 Mg ha−1.

Fig. 4
figure 4

Allometric equations between field measured Lorey’s height and AGB. The solid lines are fitted lines. Both fitted lines were extended to the full range to provide a more comprehensive representation of the potential trends of the allometric equations.

Full size image

The field data were collected in 2021, therefore they cannot be used to validate the accuracy of a ~2007 AGB map. Given the diverse forest types and vast geographical expanse of China, using these two equations to convert Lorey’s height to forest AGB across China could lead to biases when used in different forest types, and in different areas of China. The fact that the two datasets are very close together does add some confidence that the forest height: biomass relationship of China does not change greatly from location to location. There is precedent in the literature with single equations being used to derive biomass across large regions, with Saatchi, et al.25 using a single equation per continent in their pantropical biomass map, Baccini, et al.21 using a single equation across the whole tropics, and the Global Ecosystem Dynamics Investigation (GEDI) LiDAR mission’s biomass product itself using a single equation for all of Asia33. In order to assess the impact of this decision, and other errors in the biomass modelling, two independent field datasets collected closer to 2007 from across the woody areas of China were used for validation (Fig. 3).

Field data from Zhu, et al

One of the datasets consisted of field inventory measurements collected between 2011 and 2016. It included 189 sites with three 20 m × 20 m plots at each site. The longitude and latitude of each site were the averaged values of the three plots within the site (the three plots were usually close to each other, e.g., within 50 m). In order to reduce the uncertainty of biomass change from 2007 to 2016 and the uncertainty of geolocations, 83 field sites were used to validate our map. Field sites were used in analysis if they met the following three criteria34.

  1. 1)

    At least half of the area in the buffered region 100 m around the field sites were forests according to the Hansen forest cover data35. We created 100 m buffers to reduce the implication of the uncertainty of the geolocations, considering the geolocations of the sites were averaged values of the three plots in each site.

  2. 2)

    In order to show there had been no major changes in the forest in that area between the time of the creation of the maps and the date of the measurement of the field plots, we confirmed that:

    • The maximum NDVI of each year calculated from Landsat were stable from 2007 to 2016 (Fluctuations of NDVI less than 0.15).

    • The values of ALOS PALSAR and PALSAR-2 HH and HV polarizations were stable from 2007 to 2016 (Fluctuations of HH band and HV band less than 2 in decibel unit).

  1. 3)

    To exclude the field sites that are highly spatial heterogeneous, which may introduce large errors, we removed field sites that have large ratio of standard deviation of the AGB in the 100 m buffers and the field AGB (ratio > 0.25).

Field data from National Science & Technology Infrastructure (NSTI)

The other dataset from NSTI (http://www.cnern.org.cn/) was collected from 2005 to 2015 in 29 sites (Fig. 3). There were 5 to 100 plots (10 m × 10 m) in each site. The dataset was collected by nine field research stations, and was measured every few year (normally every five years). We used the average values of the measurements of AGB which were collected around 2007. We refrained from implementing the above filters due to several reasons. Firstly, the geolocation uncertainty of each site varied, making it challenging to apply a uniform filtering approach. Secondly, the size of each field site in this dataset was potentially varied, further complicating the filtering process. In addition, the dataset available for analysis was relatively limited, and applying filters could potentially result in insufficient field sites for validation purposes.

Satellite data

ICESat GLAS Lorey’s height

GLAS which was launched aboard the NASA Ice Cloud and Land Elevation Satellite (ICESat) satellite on January 12, 2003, was the first laser-ranging (LiDAR) instrument for global observations of the Earth. It was designed to measure ice-sheet topography and associated temporal changes, cloud and atmospheric properties, to obtain information on the height and thickness of radiatively important cloud layers, and as a tertiary aim obtain information on tree heights globally.

GLAS provided data with 65 m diameter circular footprints every 170 m along orbits between −86° to 86° latitude. Since we aimed to produce a forest AGB map in ~2007, we selected 287,399 GLAS footprints located on forests35 acquired during the operating periods L3H and L3I (Table 2). In this study, we directly used the Lorey’s height values generated by Lefsky27, who developed distinct equations for estimating Lorey’s height for needleleaf, broadleaf, and mixed forests globally, based on waveform extent indices from GLAS.

Table 2 Reference orbit tracks acquired during ICESat observation periods.
Full size table

We removed the GLAS footprints over hills with slope > 15°, because of potential overestimates or underestimates in Lorey’s height in high slope areas28. We also excluded GLAS footprints in forests with slopes between 12° and 15° and where AGB ≤ 40 Mg ha−1 (roughly 11–12 m in height). This is because deriving tree height becomes highly uncertain when the topographic relief within a footprint is large compared to the tree height (Fig. 5), resulting in lower AGB estimates being more susceptible to potential impacts from slope36.

Fig. 5
figure 5

The effects of slope on trees of different heights.

Full size image

ALOS PALSAR annual mosaic data

PALSAR was an L-band SAR instrument on ALOS that operated by the Japan Aerospace Exploration Agency (JAXA) between 2006 and 2011, that has proved very useful in mapping forest biomass across a range of ecosystems37,38,39,40. L-band SAR is the longest wavelength SAR (~24 cm) currently available from space. JAXA produced annual seamless mosaics of the radar backscattered intensity from the PALSAR observations at HH or HV polarization, with terrain and slope correction already applied41. We accessed the 2007 mosaic at the resolution of 25 m and did pre-processing through Google Earth Engine (GEE). We applied a focal mean smoothing filter with a radius of 150 m to reduce speckle noise. The digital number (DN) values were converted to gamma naught values in decibel unit (dB) using the equation provided by JAXA:

$${\gamma }^{0}=1{0\log }_{10}(D{N}^{2})+CF$$
(5)

where, γ0 is the backscattering coefficient in dB, and CF is a calibration factor equal to −83.0 dB42 for the PALSAR mosaic. HV, HH polarizations and the ratio of HV and HH were used in the estimation model, as the polarization ratio has a higher saturation point43,44.

Envisat ASAR

Envisat was a satellite mission operated by the European Space Agency (ESA). The main objective of the Envisat programme was to study and monitor the Earth’s environment on various scales, from local through regional to global45. Envisat was equipped with ten instruments, including an active C-band sensor ASAR. C-band is a shorter wavelength (~6 cm) than L-band, making it suitable for assisting with lower biomass areas; additionally its frequent observations in the Wide Swath mode at a resolution of 150 m allow the generation of the standard deviation of the backscatter through the year, which further helps to understand characteristics of the forests that link to biomass46. Envisat ASAR measured the radar backscatter of the Earth’s surface at HH or VV polarization, both of which were collected in our study area.

We calculated the mean values and standard deviation of combined HH and VV images together over 2007 to 2008, as C-band SAR backscatter and forest biomass has similar relationship with both polarizations47,48, and this combination reduced artefacts due to differing look angles and number of observations. In order to calculate the mean values and standard deviation, we only used pixels where there were at least two images of each pixel (gaps correspond to 1.2% of the total area of China, see Fig. 6b, where a separate AGB retrieval model not based on Envisat data was used46). This method of selecting pixels with sufficient multi-temporal coverage was aimed at ensuring more stable and reliable observations. In total we used 113,698 ASAR images tiled in 1 degree × 1 degree grid cells across China.

Fig. 6
figure 6

Earth observations used in this study. (a) False color composite of ALOS PALSAR annual mosaic data for the year 2007 (HV-polarized backscatter in red, HH-polarized backscatter in green, and the ratio of the backscatter HV/HH in blue). (b) False color composite of the Envisat ASAR dataset covering China acquired over 2007 to 2008 (mean of HH and VV in red, minimum of HH and VV in green, and standard deviation of HH and VV in blue). The white areas are where no observations were collected. (c) False color composite of temporal mean values computed from Landsat-5 TM bands 2, 3 and 4 acquired from 1st January 2006 to 30th December 2008. Band 4 (Near-infrared) in red, Band 3 (Red) in green, Band 2 (Green) in blue. (d) Filtered grid cells of ICESAT GLAS data used to create and test the model.

Full size image

Landsat-5

Since the first Landsat satellite launched in 1972, Landsat data have contributed to our understanding of Earth, especially in land use and land use change. We used Thematic Mapper (TM) images here, which consist of seven spectral bands with a spatial resolution of 30 meters for Bands 1 to 5 and 7, and a temporal resolution of 16 days. We extracted all the images acquired from 1st January 2006 to 30th December 2008, and used Quality Assessment (QA) band, generated from the CFMASK algorithm49, to mask cloud and cloud shadow. We then stacked the cloud-free Landsat-5 bands and computed the mean value of the spectral reflectance in the stack. The maximum, minimum and the differences of maximum and minimum values of Normalized Difference Vegetation Index (NDVI) of each pixel were also calculated because of the strong predictive power when related to AGB50,51,52,53.

Topographic data

NASADEM was generated by reprocessing the SRTM data with other DEM data, such as ICESat GLAS, ASTER GDEM and ALOS PRISM DEM data54. NASADEM is more accurate than SRTM, has no holes, and is globally available at the finer resolution of 30 m55. In this study, we used the elevation in the NASADEM product as well as terrain slope derived from NASADEM.

Forest AGB estimation

We converted Lorey’s height derived from GLAS into forest AGB using the allometric equations developed for northern China and southern China independently with the field data. Each individual Lorey’s height estimate from GLAS comes with considerable random errors in height estimation (RMSE was 5.9 m and R2 was 0.6727), in addition to geographic uncertainty. In order to reduce the influence of these errors, we averaged the Lorey’s height within 0.01 degrees × 0.01 degrees grid cells and obtained 8,981 cells with at least two GLAS footprints (mean 4.14 footprints). We limited Lorey’s height values to 25 m, as our review of the literature showed nearly no plots with Lorey’s height over 25 m in China56,57, especially when averaged in 0.01 degrees × 0.01 degrees grid cells.

All remote sensing imagery was resampled to a spatial resolution of 50 meters in GEE. We then used Random Forest (RF) regression58 to extrapolate AGB grid cells from GLAS data to a continuous map at 50 m resolution. We trained the RF regression using the variables derived from Earth Observation datasets listed above as well as layers giving latitude and longitude, to enable the model to take account of the differing relationships across the varied ecosystems of China. As we were focused only on forest AGB, we used the 30 m resolution Hansen et al. forest cover data35 to mask all the input layers to forest only, preventing any influence of non-forest areas on the RF regression. The 2007 forest cover map was obtained by removing pixels flagged as lost between 2000 and 2007, from baseline forest canopy cover >10% in the year 2000 canopy cover layer. However, due to the lack of forest gains between 2000 and 2007, the 2007 forest cover map may underestimate the actual extent of forest cover. Since the importance of the variables differs in flat region and rugged region (Fig. 7), we modelled the AGB in flat region and rugged region separately to make full use of the remote sensing layers.

Fig. 7
figure 7

Importance rank of selected variables in RF regression (importance of layers in flat regions from high to low). The six least important variables for all the three regions are not shown in the figure. LBx represents Landsat-5 Band x, dif_NDVI represents the differences of maximum and minimum values of NDVI. EnvisatMean and EnvisatStd represent the mean values and standard deviation of Envisat data.

Full size image

We eliminated redundant information and thus improved the computational efficiency of prediction by analysing the importance of all input layers on the AGB map. We considered the node purity, which is equivalent to Mean Decrease Gini, of each variable (Fig. 7). The higher the value of node purity, the higher the importance of the variable to our model. This is because node purity increased along with a decrease in the residual sum of squares, indicating a decrease in the Gini coefficient in regression analysis59,60. We removed the 10 variables with smallest node purity in each of the three models.

Accuracy and bias correction

To evaluate the performance of the AGB estimation model, 8 981 grid cells of 0.01 degrees × 0.01 degrees AGB were randomly split in training set (60% of the data) and validation set (40% of the data). The performance of the model was assessed by aggregating the pixel-wise AGB estimates to the same size of grid cells at 0.01 degrees × 0.01 degrees. The coefficient of determination (R2) and Root Mean Square Error (RMSE) were used to quantify the performance of the model.

The model estimated AGB corresponded well with the AGB derived from GLAS, and the R2 and RMSE between the estimated AGB and AGB derived from GLAS were 0.61 and 28.6 Mg ha−1, respectively (Fig. 8a). Our method tended to overestimate forest AGB at the forests with low AGB (AGB <60 Mg ha−1), and underestimate in forests with large AGB, as is common with RF models. To correct for this bias, a linear regression bias correction was performed to the estimated AGB using the slope of the fitting line of 10 Mg ha−1 intervals61,62,63. This improved, but did not solve, the bias issue, with increased scatter in the output (RMSE = 40 Mg ha−1 compared to 28.6 Mg ha−1 pre-correction), but with the range of values better matching the range from the GLAS predictions (maximum predicted value 251 Mg ha−1 after correction, compared to 182 Mg ha−1 pre-correction, compared to the maximum GLAS prediction of 222 Mg ha−1).

Fig. 8
figure 8

Evaluation of the model. (a) Estimated AGB from our model against AGB of averaged grid cells derived from GLAS data (grey points). (b) Bias corrected AGB against AGB of averaged grid cells derived from GLAS data (grey points). The dotted line is the 1:1 line. The black solid line is the regression of the whole set of points. The colour dots are the average values of 10 Mg ha–1 interval of AGB derived from GLAS, and the colour bar represents the number of grid cells in a given AGB interval. The blue line is a standard regression of colour dots with the intercept forced to be zero; in the bias corrected figure (b) this line overlaps exactly with the 1:1 line and is thus not visible.

Full size image

Data Records

The new forest AGB map was produced using multisource remote sensing datasets, including ICESat GLAS data acquired in 2007, ALOS PALSAR data acquired in 2007, Envisat ASAR data acquired in 2007 and 2008, Landsat-5 data acquired from 2006 to 2008, and NASADEM data. Due to the remote sensing datasets all acquired around 2007, it is a circa 2007 forest AGB map. The AGB map is stored in GeoTIFF (.tif) files on Edinburgh DataShare64. The map provides the values of AGB in Mg ha−1, and the values have been rounded to integers. It can be converted to carbon by multiplying by 0.4765.

Technical Validation

Our estimates of total carbon stored in forest AGB in China is 9.52 Pg C. To assess the accuracy of our AGB map, we validated our map using one field datasets from a previous study34 and the other independent dataset from NSTI. The coefficient of determination (R2) and Root Mean Square Error (RMSE) were used to quantify the performance of the maps. In addition, we compared this map to the GEOCARBON map6, Su map23 and CCI map9 and attempt to explain the differences.

Validation using field data

Comparison to Field data from Zhu, et al

We used 83 sites from Zhu, et al.34 to validate our map at a pixel level. Overall, the estimated AGB of our map demonstrated a reasonably good correspondence with field measured AGB, with an R2 of 0.62 and an RMSE of 57.15 Mg ha−1; this exceeds considerably the metrics obtained from the other datasets (Fig. 9). Compared with the field data, all the four maps exhibited a tendency to underestimate AGB values in areas with large AGB (>200 Mg ha−1). The other national scale AGB map, the Su map23, exhibited a lower RMSE than the two global AGB maps.

Fig. 9
figure 9

Assessment of AGB map against independent field data from Zhu, et al. The dotted line is the 1:1 line.

Full size image

To assess the similarities and differences in the AGB distribution, we compared field data from Zhu, et al.34 and the pixel values of the four AGB maps within 1 km × 1 km grid cells, which represents the resolution of the coarsest map (Fig. 10). The field data hosted the highest median and mean AGB, and a half of the field sites had an AGB density of 95–160 Mg ha−1. The map from this study had a similar low end of the interquartile range with the field data, but was more homogeneously distributed with 75% of the AGB below 130 Mg ha−1. GEOCARBON map hosted the smallest median and mean AGB and was most homogeneously distributed. The AGB of Su map had the largest dispersion, with 50% of the AGB were between 80 Mg ha−1 and 190 Mg ha−1. The CCI map had the second smallest median AGB of 60 Mg ha−1. It can be seen that the two national scale AGB maps are closer match to these field data in this comparison.

Fig. 10
figure 10

Boxplot of field data and forest AGB of the four AGB maps within 1 km × 1 km grid cells of field sites. The orange horizontal bar indicates the median and the boxes show the interquartile range, and the “+” indicates the mean forest AGB.

Full size image

Comparison to Field data from NSTI

We also validated our map, and the GEOCARBON, Su and CCI maps using field data from NSTI. The uncertainty of the geolocations of the data from NSTI were from 0.0001 degrees to 0.01 degrees (~10 m–1000 m). In order to reduce the effect of errors caused by the coarse geolocation, we aggregated our estimated forest AGB map and the field data to 0.01 degrees × 0.01 degrees grid cells. 15 grid cells aggregated from 29 field sites were used to validate the four AGB maps. The Su map had the strongest correlation with the field data from NSTI, which can be attributed to the incorporation of part of the field data in their training dataset (so this is not an independent test of the Su map). Our map demonstrated the second best correlation with this field data, with an R2 of 0.57 and an RMSE of 103.2 Mg ha−1. All four maps showed relatively high RMSE values when compared to the field data (Fig.11). This can be attributed to the limited coverage of field measurements within the 0.01 × 0.01 degrees grid cell. In fact, even within the grid cell with the highest number of field plots, the area of the field measurements cover ~1% of the total area of the grid cell.

Fig. 11
figure 11

Assessment of AGB map against independent field data from NSTI. The dotted line is the 1:1 line.

Full size image

The field data from NSTI had a large range of forest AGB, with 50% of the field data between 80 Mg ha−1 and 230 Mg ha−1 (Fig. 12). The field data had the largest median and mean AGB. 50% of our map had an AGB of 100–150 Mg ha−1, and both of median and mean AGB were around 120 Mg ha−1. Both of GEOCARBON map and CCI map had 50% AGB between 40 Mg ha−1 and 100 Mg ha−1. The two maps had similar mean AGB of 80 Mg ha−1, but the median AGB of GEOCARBON map is lower. The Su map had the second largest median AGB and mean AGB of 140 Mg ha−1and 130 Mg ha−1, respectively.

Fig. 12
figure 12

Boxplot of field data and forest AGB of the four AGB maps within 1 km × 1 km grid cells of field sites. The orange horizontal bar indicates the median and the boxes show the interquartile range, and the “+” indicates the mean forest AGB.

Full size image

Comparison to existing AGB maps

The bias corrected AGB map from this study was compared with AGB maps by Avitabile, et al.6, Su, et al.23, and Santoro and Cartus9. Some of the maps are generated globally, while others are for specific smaller regions. We masked the non-forest area of Santoro and Cartus9 using the Hansen forest cover map in 200735, as this map expresses woody biomass and we focused on forest AGB in this study. In order to compare the maps at pixel level, we reprojected and resampled the AGB maps to the same projection and resolution as our map using a nearest neighbour resampling method, avoiding uncertainty which can be introduced during warping, and appropriate as our map is much higher resolution than the maps compared. Forest AGB density were converted to above-ground carbon using a default conversion factor of 0.4765. We compared the AGB maps across China and for three specific regions: southern, northeastern and middle part of China (Fig. 13). These regions were chosen due to their extensive forest coverage and were derived by combining small vegetation zones66. Compared to the three AGB maps, our results show large differences in total carbon and the distribution of AGB, in particular in southern China, where there are dense forest and steep slopes which introduce large uncertainties in AGB estimation.

Fig. 13
figure 13

Forest zones we used to compare the forest AGB maps.

Full size image

Across China

All four forest AGB maps showed considerable higher AGB in southwestern China, and the lowest AGB in northern China. However, significant differences were visible when the four AGB maps were compared (Fig. 14).

Fig. 14
figure 14

Spatial distribution of forest AGB across China.

Full size image

Among the four AGB maps, average AGB from the Su map23 was higher than the other three maps, while the average AGB from this study was the second highest and differed greatly from the values from GEOCARBON map6 and CCI Biomass map9.

Some of the differences in average AGB and total carbon storage in Table 3, may relate to the differences in the total area of forest used in the calculations. According to assessment report by the Food and Agriculture Organization (FAO), the forest area in China was 193.0 million ha in 200567. The forest area from this study, which used the Hansen forest cover map35 to mask the non-forest area, was close to the forest area by FAO. The explanation for a smaller forest area from CCI map9, which we also used the Hansen forest cover map35 to mask non-forest area, was that we only counted the pixel values higher than zero as forest, leading a small difference in the forest area of the two maps. The GEOCARBON map6 and Su map23 showed significant differences in the estimated forest area compared to this study. This is because the two AGB maps used different forest cover maps, the GLC2000 map68 and the 2000 land use map69. The total above-ground carbon stocks for China were estimated by summing the carbon of all pixels in each AGB map. The total carbon stored in China’s forests was estimated 9.52 Pg C in this study, which was almost two and a half times as much as that in GEOCARBON map6, 60% higher than that in CCI map9, but less than 60% of that in the Su map23 (Table 3).

Table 3 Average AGB, forest area and total carbon in China.
Full size table

The histograms of the four AGB maps across China in 10 Mg ha−1 bins showed that the peak of frequency of AGB in CCI Biomass9 occurred at a lower AGB of 0 to 40 Mg ha−1, while the peak value in Su map was the highest (Fig. 15). The frequency of AGB in this study peaked at 100 to 110 Mg ha−1, and more than half of the AGB ranged from 70 to 130 Mg ha−1. More than 80% of AGB in GEOCARBON map6 ranged from 20 to 90 Mg ha−1, with nearly no AGB higher than 150 Mg ha−1. The AGB in CCI Biomass map9 had no AGB in 10 Mg ha−1 bins greater than 10%, and the frequency of AGB decreased from 40 Mg ha−1, and increased at the >200 Mg ha−1 interval. The AGB histogram for southern China is similar to that of the entire country, with the exception that the Su map contains a greater frequency of AGB values exceeding 200 Mg ha−1. In the northeastern and middle China, the AGB is predominantly less than 150 Mg ha−1.

Fig. 15
figure 15

Histogram of AGB in 10 Mg ha−1 bins, 10 indicts 0–10 bin.

Full size image

Based on the assessment of the field data above and comparisons with existing AGB maps, we would recommend that this new dataset represents the best estimate of the total carbon storage, and its distribution, for China in this time period.