1. Introduction
As the major component of terrestrial ecosystems, forests hold 70%–80% of terrestrial aboveground and belowground biomass and play an essential role in the global carbon cycle and climate change [
1]. Forests are vulnerable to fire, logging, pests, and land conversion, which release carbon easily to the atmosphere [
2]. The annual carbon flux between forests and the atmosphere accounts for 90% of the flux between the atmosphere and Earth’s land surface [
3]. Therefore, mapping the magnitude, spatial distribution, and change of forest aboveground biomass (AGB) over time is important for improving estimates of terrestrial carbon sources and sinks. Furthermore, some studies of global change and the carbon cycle showed that there was a large uncertainty in the forest carbon stocks in tropical and boreal regions because of the lack of accurate forest biomass maps [
4,
5]. Regional to global information on the human impact on carbon stocks and ecological balance also requires accurate determination of forest biomass and monitoring of its changes, especially in areas of fragmented forest cover and in developing countries [
6,
7,
8]. These scientific objectives require accurate mapping of forest biomass at the national level for the assessment of forest carbon stocks and their dynamics.
Traditionally, field measurements for estimation AGB are collected at sampling sites by measuring tree heights and DBH (diameter at breast height). The biomass of a tree can be estimated from its specific allometric equation using these two variables or DBH alone. The biomass of a plot is typically expressed in the form of biomass density,
i.e., in weight per unit area. Previous studies on biomass estimation were developed based on the extrapolation of biomass from specific sites to large areas (
i.e., multiplying the mean biomass density estimated from direct field measurements by the forested area or using biomass expansion factor to convert timber volume to biomass) [
2,
9,
10,
11]. These methods can obtain good biomass estimation for a small area, such as a forest stand. However, the lack of field data in remote areas and the inconsistency of data collection methods among different management units over large regions were the major constraints to obtaining reliable large-scale biomass estimation using field-based methods [
12,
13]. Moreover, obtaining comprehensive, spatially complete, temporally uniform, and accurate forest inventory data was usually time-consuming and labor-intensive over huge areas and field campaigns were not well suited for detecting changes because a single measurement campaign can extend over several years in most countries, especially in developing countries with large land area [
12,
14,
15].
Satellite remote sensing is fairly advantageous for regional or global AGB estimations because of its large area coverage [
16]. However, as no remote sensing instrument has been developed that is capable of providing direct measurement of biomass, additional field-sampled biomass is required for establishing relationships between remote sensing signals and biomass [
17]. Optical satellite data can provide fine results in the discrimination of largely different biomass classes across a variety of spatial and temporal scales due to its spectral sensitivity to different species. Furthermore, satellite-based optical imaging spectrometers are perceptive to vegetation structural parameters such as tree abundance, basal area, stem density and crown size, which are correlated to some degree with biomass because vegetation spectral reflectance contains information on the vegetation chlorophyll absorption bands in the visible region (especially in blue band and red band) and sustained high reflectance in the near-infrared region [
18]. Particularly, recent research has focused on the utility of short wave infrared (SWIR) data for estimating AGB [
19,
20]. Besides application of single-band spectral signatures, various vegetation indices (VIs) derived from TM (Thematic Mapper), ETM+ (Enhanced Thematic Mapper Plus), AVHRR (Advanced Very High Resolution Radiometer), and MODIS data have been used to estimate forest biomass in various areas, including southern Asia, boreal and temperate forest zones in America and Asia, subtropical areas in South America, southwestern United States and tropical Africa, America and Asia [
4,
19,
20,
21,
22,
23,
24,
25]. Although passive optical remote sensing is a less expensive technology for biomass estimation over vast areas, frequent cloud coverage in mountain areas and the low saturation level of spectral bands and VIs on biomass estimation in medium to high biomass forests are the major disadvantages of this technology [
26,
27,
28].
LiDAR (light detection and ranging) is an active sensor type based on laser ranging, which emits a laser pulse and measures the distance based on half the elapsed time between the emission of a pulse and the detection of a reflected return [
29]. When the surface is covered by vegetation, echoes are a function of the vertical distribution of vegetation and the ground surface within a footprint. LiDAR data capture the horizontal and vertical structure of vegetation in great detail, resulting in accurate estimates of forest biomass across a broad range of forest types and biomes [
30,
31,
32,
33]. LiDAR systems acquire data either over small footprints (<1 m, point-cloud data, or waveform data, airborne) or large footprints (>10 m, waveform, either airborne or spaceborne) [
34,
35]. LiDAR is the only sensor type available at present whose signal does not saturate in high biomass forests (e.g., 1200 Mg·ha
−1 and 1300 Mg·ha
−1 [
36,
37]). Estimating biomass with airborne LiDAR data is often more accurate [
38,
39], but the associated large data volume, as well as sophisticated technical equipment and high acquisition costs to observe remote areas limit its application at regional and global scales [
16,
19,
27]. The Geoscience Laser Altimeter System (GLAS), onboard the Ice, Cloud, and land Elevation Satellite (ICESat), was a spaceborne LiDAR system recording full waveforms over large footprints [
40]. GLAS used 1064-nm laser pulses operating at 40 Hz and recorded the echo of those pulses from footprints ~65 m in diameter, spaced 172 m apart [
41]. GLAS footprint data as a manner of sampling are crucial for associating field plot data with optical imaging systems because LiDAR samples could compensate for the lack of systematic spatial sampling of AGB from ground measurements. One major limitation of GLAS was the lack of imaging capability and the fact that it provided relative sparse sampling information on forest structure [
42]. Therefore, multi-sensor data synergy is required to estimate forest structural parameters at the regional scale.
Over the past few years, the integration of optical remote sensing and GLAS data has been successfully used to estimate forest volume and biomass because the signature from these two types of data provide relevant information content, e.g., species and forest coverage from the optical data, and three-dimensional vertical forest structure information from LiDAR data [
42,
43]. For example, Boudreau
et al. [
12] used GLAS waveform metrics, airborne profiling LiDAR, and field plots to estimate forest volume, biomass, and carbon in Quebec, Canada, covering an area of 1.3 million km
2. A generic airborne LiDAR-based biomass equation (R
2 = 0.65) was developed from field inventory data as a function of airborne profiling metrics. A second model was explored that predicted biomass as a function of GLAS waveform variables. Thus, GLAS footprints became the regional sampling tool used to estimate forest biomass on forest cover strata that were derived from Landsat-7 ETM+ [
44]. Nelson
et al. [
30] developed a GLAS-based equation to estimate timber volume in south-central Siberia using GLAS and MODIS data. Although only 51 ground plots co-located with GLAS footprints were sampled, regional total volume estimates and per-hectare estimates were compared with ground-based study results for the entire 811,414 km
2 area, with differences of 1.1% and 11.9%, respectively. Baccini
et al. [
45] mapped AGB across tropical Africa (1-km resolution) using MODIS data with a large dataset of field measurements. They found a strong relationship between the estimated AGB and GLAS height metrics (average height and height of median energy). That study noted, “…there are currently limited high quality field biomass estimates available at sufficient spatial extent to develop and independently validate maps of AGB across tropical regions…” [
45]. Baccini
et al. [
20] estimated the carbon density of aboveground live wood vegetation for the pan-tropics at a spatial resolution of 500 m from MODIS data using GLAS waveform metrics. Saatchi
et al. [
4] mapped biomass (above- and belowground) in tropical regions across three continents using a combination of data from 4079
in situ inventory plots and GLAS samples of forest structure, plus optical and microwave imagery (1-km resolution). They developed continent-based allometric equations to provide the best models to convert Lorey’s heights derived from GLAS data to forest biomass. A data fusion model based on the maximum entropy approach was used to extrapolate AGB derived from GLAS footprints to the entire landscape. This benchmark map provided comparable estimates of AGB for 75 developing countries.
China, as one of the world’s fastest developing countries, needs to produce robust estimates of forest biomass and carbon stocks for successful implementation of climate change mitigation policies. As one of the five most forest-rich countries [
46], China is rich in temperate forests and subtropical forests. Timely and accurate measurements of forest biomass and its distribution are increasingly needed to support a wide range of activities related to sustainable forest management and carbon accounting. Previous studies on estimates of forest biomass in China were based on statistical analysis of the biomass-volume relationship based on nationwide forest inventory data [
47,
48]. Despite the high precision of such inventories, they do not provide maps of biomass at a resolution useful for assessing land-use change. An AGB map of China with clear and detailed spatial distribution is urgently needed. However, the benchmark map of Saatchi
et al. [
4] did not cover the entire land of China. Similarly, the pan-tropical map generated by Baccini
et al. [
20] only covered the area in southern China (below 30°N). More importantly, no field survey samples from the Chinese territory were included in the two studies.
In this study, we explore the capabilities and limitations of satellite remote sensing data (GLAS and MODIS) and ground measurements for mapping AGB in China. The slope effect on biomass estimation in a footprint was considered. The potential information on biomass from GLAS waveforms and GLAS samples for reliable biomass estimation were studied using field data. Proper prediction models were used to extend the biomass estimates from GLAS samples to all forested areas throughout China. Biomass estimation results were compared with field biomass data within a footprint, with the aim of analyzing the effect of slope on the accuracy of the estimates. The biomass map was validated with other biomass estimates derived from forest inventory data.
4. Discussion
A method to assess the AGB of Chinese forests was developed by combining large footprint LiDAR data, field measurements, and optical remote sensing data. The results demonstrated that using a small part of forest inventory data to estimate national AGB was feasible, and the estimates were reasonable. GLAS footprint data as a method of sampling were important for linking field plot data to MODIS data because the acquisition of field plot data is time-consuming, and LiDAR samples could extrapolate biomass of limited field plots to more footprints in a greater range. The LiDAR waveform records the interaction of a laser beam with trees in a footprint, starting from the crown. The crown shape reflects the basic geometric structure of a tree. The crown shape of a needleleaf tree is similar to a cone. Therefore, to estimate biomass within a GLAS footprint located in a needleleaf forest, fewer waveform parameters reflecting the characteristics of canopy vertical structure were used in the regression model, such as the mean canopy height or h50. The structure of a broadleaf tree is more complex, and cannot be described by one type of tree crown. More parameters reflecting the overall structure of canopies were used in the models. Moreover, we found waveform energy heights of 75% (h75) were positively correlated with biomass, and used them in the estimation models. Other researchers have also found that h75 was a good indicator of forest vertical structure [
46]. In addition, a new waveform parameter h
slope proved to be helpful for predicting footprint-level biomass. In study of boreal forest biomass with large footprint LiDAR at similar latitudes, Boudreau
et al. [
12] used variables of GLAS waveform extent, waveform front rising slope angle, and a terrain index to estimate biomass in Quebec, Canada. We determined the mean height, median height, and h100 (similar to waveform extent) as the main parameters for predicting AGB in zone I at the same latitude as Quebec. Lefsky
et al. [
36,
53] found that the mean canopy height and the square of the median canopy height had high correlations with AGB in Oregon, USA. We discovered that h50 (similar to median height) and h75 were useful for estimating AGB in forest zone II, at a latitude similar that of Oregon.
The GLAS models tend to overestimate field aboveground biomass with increasing slope [
56,
58]. The GLAS-sampled biomass was divided into three classes according to the slope in the footprints (all slopes, slopes <20° and slopes ≥20°) to analyze the impact of slopes on AGB estimates.
Figure 4 shows that AGB estimates using footprints across all slopes compared well with biomass derived from field data on moderate slopes (0°–20°). This fact may partly be due to logging activity in areas with low to moderate slopes, leading to younger, secondary forests with coverage condition simplifying estimation of biomass. AGB estimated from GLAS footprints on slopes <20° was closer to the biomass derived from field data than all other slopes (
Figure 4). In steeply sloped areas, better results were achieved using GLAS footprints on slopes ≥20° rather than on all slopes. This could mean that slopes on <20° and ≥20° combined are more appropriate for predicting biomass than using footprints across all slopes. Another issue that should be noted is that, the biomass range of field data may not to include high values of AGB under moderate terrain conditions of slope <20°, and could be a potential source of error in estimating AGB of GLAS footprints. This uncertainty should be checked before training the AGB estimation models to ensure the representativeness of the training samples.
After calculating biomass from the GLAS footprints, the GLAS-based estimates of biomass can serve as a calibrated inventory dataset for all of China, overcoming the dependency on national forest inventory data. Extrapolating a large number of GLAS-based AGB estimates to a continuous map spanning China requires widespread surface coverage. MODIS image products with coarser spatial resolution and high spectral resolution were useful for scaling forest AGB from plots to sub-continental scales. MODIS images used in this study included optical bands centered on visible, near infrared, and middle infrared, that many studies have indicated are sensitive to forest cover and forest stand structure. MODIS images in high forest cover areas saturate at relatively low levels of forest biomass. Fortunately, good results in AGB estimation of GLAS footprints could offer reliable biomass samples for training across the full range of AGB within a MODIS image. This could improve the saturation of reflectance values in densely forested areas more than using only MODIS images and forest inventory data for biomass estimation.
As expected, the highest forest AGB was found in the southwest of Xizang province (>300 Mg·ha
−1), while the lowest forest AGB was found in Hebei province in the north, and Jiangsu province in the south (<50 Mg·ha
−1). Good agreement (relative error <20%) between estimates of AGB and forest inventory data was seen in most southern provinces of China. Large discrepancies mainly occurred in provinces located in north-central and east-central China. The potential reason for this is low forest coverage in these areas. The low cover could lead to mistakes in forest/non-forest classification, producing significant differences in AGB estimation. As shown in
Figure 6, the biomass ranges were different in seven forest zones. In zone I, the Northern Da Hinggan Mountains suffered from serious fires in 1987. Nearly 20 years later, a field survey was conducted in 2006, and showed that regenerated forest coverage had improved significantly under intensive management, but the stand density and canopy height had not reached levels that existed before the big fire. Before the big fire, the average biomass of larch (
Larix gmelinii (Rupr.) Kuzen.) in this area was about 105.05 Mg·ha
−1 [
55]. After the big fire, it is expected that forest biomass density would decreased. In our research, a biomass range between 60.1 Mg·ha
−1 and 90 Mg·ha
−1 accounted for the major part of total AGB, consistent with the result of a previous study that the average biomass of a natural young larch forest was about 66.93 Mg·ha
−1 [
55]. The Xiao Hinggan Mountains has not undergone any large natural disasters in the last 20 years. Thus, forests in this region relatively undestroyed compared with the Da Hinggan Mountains. Therefore, the largest biomass range, from 90.1 Mg·ha
−1 to 120 Mg·ha
−1, occurs in zone II. The second largest biomass range was from 120.1 to 150 Mg·ha
−1. The composition of biomass ranges in zone III was similar to that in zone I. Because the area of forested land in zone III was smaller than in zone I, the amount of total AGB in zone III was less than in zone I. Affected by the long-term human activity, forest zone III was dominated by secondary and man-made forests. Chinese pine (
Pinus tabulaeformis Carr.), armand pine (
Pinus armandi Franch.), and Japanese red pine (
Pinus densiflora Sieb. et Zucc.) were the major species of needleleaf trees, while typical species of broadleaf trees were oak (
Quercus L.) and birch (
Betula L.). The average biomass of these species at the age of about 30 years were 96.37 Mg·ha
−1, 88.9 Mg·ha
−1, 89.27 Mg·ha
−1 (at the age of 60 years), 49.13 Mg·ha
−1, and 84.47 Mg·ha
−1, respectively [
55]. This is consistent with our conclusion that most of the forest biomass was in the range between 60 Mg·ha
−1 and 90 Mg·ha
−1. Forest zone IV had the largest forested area of all forest zones, and the largest total amount of biomass, ranging from 60.1 Mg·ha
−1 to 90 Mg·ha
−1. Of the seven forest zones, zone IV had the highest total biomass because it was dominated by middle-aged man-made forests and young secondary forests. Biomass ranges between 150.1 Mg·ha
−1 to 200 Mg·ha
−1 accounted for the greatest amount of biomass in zones V and VII. Zone V was mainly covered by tropical monsoon forests with high stand density and canopy height. The forest region in zone VII was located in the southeastern Xizang province, and retains large areas of virgin forest. The altitude of the mountain valley is relative low, and the region has high levels of precipitation and accumulated temperature. All of these conditions result in a high biomass range (>150 Mg·ha
−1) in zone VII.
It was a difficult to evaluate the national forest AGB. In order to comprehensively evaluate the AGB estimates, validation was carried out at two scales, the forest stand and nationwide scales. As much as possible, forest stands with the greatest ranges in spatial distribution of biomass were selected. All values of R
2 were lower than 0.7, mainly because more than one pixel were dropped in one forest stand, so the value of biomass in the stand was compared with the average AGB value of two or more pixels. The R
2 values were below 0.55 in the Tahe and Qinling regions. The low R
2 value in Tahe was mainly due to mature forest suffered great loss during the big fire in 1987. Relatively low stand density and canopy heights would affect biomass estimates in GLAS footprints. The field plots in the Qinling area were located on the north slope of Qinling Mountain. The complex topography makes it difficult to estimate AGB with high accuracy. However, the variances of the forest stands in Tahe and Qinling were lower than for that in Lushuihe and western Sichuan. It may be inferred that forest stands in Tahe and Qinling were more homogeneous than that of the other two places. There were small errors between the national AGB estimation and forest inventory data (12,622 Mt
vs. 12,617 Mt). A total of 415,000 million fixed forest inventory sample plots (
http://www.forestry.gov.cn/) were measured in the seventh forest resource investigation, while fewer than 30,000 plots were used in this research, with the intent of showing that a reasonable and reliable AGB map can be predicted by a limited number of field plots and satellite data. Indeed, large differences existed in some provinces, and more studies are needed at the next step, with more detailed forest inventory data for each province. The forest AGB map would be useful for depicting and quantifying the distribution of forest aboveground biomass over entire landscapes in China.
To reduce errors, LiDAR data were selected during the growing seasons, according to geographical forest distribution. For example, in high-latitude or cold areas, a summer GLAS dataset was selected. In contrast, at mid-latitudes, all datasets except for winter datasets were used for modeling. In the process of extrapolating from GLAS biomass samples to the 500-m spatial scale, MODIS images were processed in a similar manner to maintain the consistency of the acquisition phase. There are some other limitations in the forest AGB estimation. First, Chinese forested lands were roughly divided into seven zones. There are more than 15 eco-regions associated with forests in the Chinese Vegetation Atlas [
49]. Field measurements in these regions were time-consuming and labor-intensive. Regions aggregated in several larger zones were required for macroscopic study using remote sensing techniques. We aggregated the forests in China into seven zones based on the geographic characteristics of the forested areas [
49,
51] to ensure the rationality of partitioning. Second, the forest inventory data represented an uneven sampling of spatial distribution because of human resource constraints and data access restrictions. Third, the acquisition time of GLAS data, MODIS images, and field plots were different in 2005 and 2006, and from 2003 to 2010 (most were between 2003 and 2007, except the field survey in Yunnan in 2010). However, most field data were collected at the end of the growing season between 2003 and 2007 to obtain the biomass at the peak of accumulation, so this could not be a major source of the error. Therefore, seasonal differences in acquisition time between GLAS data and MODIS imagery was likely a potential bias. As GLAS data were acquired over a broad range of latitudes, longitudes, and elevations, the effect of the presence or absence of leaves in GLAS data had to be considered for generating AGB map. Furthermore, snow might be present in high-latitude or high-elevation areas, which could introduce errors in the regression between GLAS-derived biomass and MODIS data.
5. Conclusions
We used GLAS waveform data and MODIS imagery to generate an accurate forest AGB map of China. A statistical relationship between field biomass estimates and GLAS metrics was examined. When GLAS waveform parameters associated with terrain factors (e.g., lead, trail, hslope) were considered in a regression analysis, the R2-a of the estimation models were significantly increased (e.g., R2-a changed from 0.698 to 0.805 in biomass estimation model of broadleaf forests in zone III). In biomass estimation from GLAS waveform data on slopes <20° and ≥20°, the bias between biomass estimates of the models and biomass estimated from field data were smaller than the bias between biomass estimates of models using GLAS footprints on all slopes and biomass estimated from field data. The assessment of the national biomass map using forest inventory data was performed at two different aggregated scales: (1) At the forest stand level, high correlation appeared in Lushuihe, Jilin province with an R2 of 0.672, and the smallest RMSE was found in Tahe, Heilongjiang province, with a value of 8.05 Mg·ha−1. (2) Comparison of national biomass estimates summed in each province with statistical biomass derived from forest inventory data was performed for 31 provinces. Fifteen provinces had relative errors less than 20%, and relative errors of ten provinces were between 20% and 50%. The other six provinces had relative errors greater than 50%, but less than 100%. These result showed overall high consistency between estimated AGB and statistical results at the province scale.
In summary, this study demonstrated that GLAS and MODIS data can be used to estimate forest aboveground biomass at a national scale. GLAS footprints as a way of sampling can be extrapolated to estimate biomass for all of China at a spatial resolution of 500 m. The plot-based methodology using limited ground measurements would reduce the dependence on forest inventory data. Application of this method for different years will provide a chance to understand the impact of forest disturbance on biomass change, and to fill in gaps in some years when the inventory data is incomplete. The mapped area covered all forested land in China, and the total forest AGB was about 12,622 Mt in 2006. The relatively fine-scaled, spatially explicit forest aboveground biomass map provides critical AGB information for forest carbon cycle studies and forest resource management.