Grassland ecosystems are one of the most important types of terrestrial ecosystems on the planet. They provide the ecosystem functions of soil and water conservation, wind erosion prevention, sand fixation and air purification. They also play an important role in the global terrestrial circulation among ecosystems. In addition, grassland ecosystems are important for “livestock production”. China’s grasslands encompass an area of approximately 400 million hectares, accounting for 41.7% of the country’s total land area [1
], and constituting the single largest type of terrestrial ecosystem in China. Grassland biomass is the most direct indicator of grassland’s ecological status [2
]. Thus, a precise and rapid method for the estimation of grassland biomass is of great importance for both basic science and the management and protection of grasslands.
Grassland biomass directly reflects the level of grassland productivity. Productivity refers to the amount of organic matter produced by autotrophic organisms in a given unit of area over a given unit of time. Grassland net primary productivity (NPP) is obtained by subtracting autotrophic respiration [3
] from the total amount of organic matter fixed by grassland vegetation. Grassland biomass is defined as the yield of fresh grass or hay that can be harvested at a certain time, and is close to the net primary productivity without regard to disturbances. Productivity is the basis for the formation of biomass, and biomass is the manifestation of productivity. The currently available methods for the estimation of grassland productivity and biomass primarily include field surveys, statistical modeling [4
], process modeling [8
] and parameter modeling [12
With the development of remote sensing technology, more and more fields have begun to involve remote sensing technology [17
], and the development of remote sensing technology has stimulated studies on vegetation productivity and biomass. Due to the simple calculations involved and the high accuracy of the approach, statistical regression models using remote sensing data have been widely applied for the estimation of grassland biomass. The essence of the method is the establishment of a regression model between biomass measurements and either single-band remote sensing or vegetation index data, using satellite remote sensing data as the input parameters to estimate biomass. Initially, as the use of single-band information was simple, data of this type were used for estimating grassland biomass. However, due to strong interference from a number of factors (e.g., air, soil, sensor performance and the angle of the sun), the resulting estimation accuracy was found to be poor [19
]. As such, the vegetation index is now commonly used as the form of input data for building empirical regression models of vegetation biomass. Piao, et al.
] used the normalized difference vegetation index (NDVI) calculated by NOAA/AVHRR to establish a grassland vegetation biomass estimation model for China, and used the model to study the spatial distribution characteristics of China’s grassland vegetation biomass. Advances in remote sensing technology have generated higher-resolution images that are now used to estimate grassland productivity and biomass through methods, such as the application of MODIS data. Xu, et al.
] performed a systematic estimation of China’s grassland productivity by region using a combination of MODIS data and ground survey data for the same time period. For different types of grassland areas, they established relational models between NDVI and field survey biomass data that allowed them to estimate the distribution of grass production in China. Yang, et al.
] employed enhanced vegetation index (EVI) data from MODIS to estimate aboveground biomass in Tibet and analyze the relationship between the grassland’s aboveground biomass and climatic factors. Gao, et al.
] used the MODIS vegetation index to conduct in-depth research on the spatial distribution of the aboveground and underground biomass of the Xilingol grassland.
Using a remote sensing vegetation index to estimate aboveground biomass provides accurate and rapid results, but the limitations of the vegetation index itself may affect the obtained grassland biomass estimates. Specifically, in low-coverage grasslands, due to the significant influence of the soil background and grassland vegetation types, the estimation results exhibit high error rates. In addition, in high-coverage grasslands, the NDVI shows decreased sensitivity. A “saturation” phenomenon appears when the NDVI is higher, resulting in a decline in the accuracy of grassland biomass estimates. The MODIS productivity products regularly published by NASA are surface photosynthetic products obtained from model estimates. After processing these products, we established a relational model between the MODIS productivity products and field-measured aboveground biomass. We then estimated aboveground biomass by accepting the MODIS productivity data as input parameters to avoid or reduce saturation problems if the input parameters were too high. In the present study, based on this strategy, we used ground survey data from the Xilingol grassland for the years 2005–2012 and MODIS productivity data for the same time period to establish statistics-based models for estimating biomass. We further tested the accuracy of the models and selected the optimal model for estimating the aboveground biomass of the Xilingol grassland during the growing period.
4.1. Potentials Analysis of Model
A significant empirical relationship was found between the accumulated PSNnet data and the aboveground biomass data, showing good prospects for the application of MODIS productivity data in combination with ground sampling data to establish models for biomass inversion. There are three reasons to support this conclusion, as follows:
Through the accumulation of PSNnet data every eight days from the beginning of the growing season to the peak of the growing season, a good correlation was achieved between the obtained accumulated PSNnet data and the peak growing season NDVI data for the corresponding spatial point, showing coefficients of determination (R2
) up to 0.75 (Figure 7
). Because there have been many previous studies on statistical models of the NDVI and aboveground biomass [35
], we can assume a good correlation between the PSNnet data and the aboveground biomass data. Furthermore, we can assume that using the PSNnet data and the aboveground biomass data to build the model and then retrieve the aboveground biomass data is a practical method. In addition, the MODIS productivity data used in this study fully considered the effects of temperature, precipitation and other climatic factors during the process of estimating vegetation productivity. Compared with the NDVI data employed in traditional methods of biomass estimation, MODIS productivity data can better reflect the effects of environmental stresses.
The temporal matching between remotely sensed images and ground survey greatly affects the accuracy of remote sensing based models for grassland biomass estimation. The database of the extensive filed samples and their matching remotely sensed data is the basis of improving the model precision and stability. In this study, a sound database combining multi-year accumulated PSNnet data and ground survey biomass data with strict temporal matching was developed, which was further applied to biomass estimation models.
Two methods are often used to evaluate model performances. One is based on the coefficient of determination (R2
), another way is to assess model error. In general, a high R2
or a low error value often indicates a good fit between the model developed and the ground survey data [39
]. In this study, we compared the correlation between NDVI data and biomass, as well as accumulated PSNnet data and biomass. The result showed that R2
between calculated PSNnet data and biomass was a little higher than R2
between NDVI data and biomass, as shown in Figure 8
. We used ground survey data from the Xilingol grassland for the years 2005–2012 and MODIS productivity data for the same time period to establish statistics-based models for biomass estimation, with an overall accuracy of 69%, which is close to highest accuracy (74%) by Jin, et al
]. In addition, NDVI data are prone to an “oversaturation” phenomenon if the vegetation coverage is higher, which decreases the sensitivity of biomass estimation, whereas MODIS productivity data can overcome this oversaturation problem. Therefore, in high vegetation cover conditions, the biomass estimation accuracy by MODIS productivity data would be higher than the biomass estimation accuracy by NDVI data.
4.2. Uncertainties of Model
In this study, based on ground survey data from the Xilingol grassland for the years 2005–2012 and MODIS productivity data for the same time period, we developed statistical models based on the relationship between the PSNnet data and aboveground biomass data and then selected the optimal model to estimate the grassland aboveground biomass of the Xilingol grassland. However, our estimate still retains some uncertainties for the following reasons. First, there exists the scale effect between remotely sensed data and ground survey data. The spatial resolution of MODIS productivity data is 1 km, differing from the size of quadrat. In the course of the sampling, we have taken the average value of multiple samples to reduce the estimation error caused by the scale effect. Second, the remotely sensed data used in this study came from the MOD17A2 eight-day PSNnet product, which is a global land vegetation net productivity product calculated by model and has been validated widely, across most of the world. In the process of performing these calculations, nevertheless, since some of the maintenance respiration costs and all of the growth respiration costs have not been accounted for in the daily timestep, the daily output from this algorithm is termed PSNnet (net photosynthesis), to differentiate it from the true daily NPP. This reason has magnified the contrast between net photosynthesis and actual biomass, and becomes the important uncertainty source for biomass estimation. In addition, although the MODIS-PSNnet dataset has been processed with a series of corrections, there is still a certain level of deviation, which causes uncertainty in the estimation. Understanding and identifying the sources of uncertainties and then devoting efforts to improving them are critical in improving grassland aboveground biomass estimation performance; therefore, more research is needed in the future for reducing the uncertainties from different sources in the aboveground biomass estimation procedure.
4.3. Comparison with Previous Estimates
As shown in Table 4
, the average aboveground biomass densities were estimated to be 27.23 g·m−2
for the temperate desert steppe, 69.10 g·m−2
for the temperate steppe and 115.23 g·m−2
for the temperate meadow steppe. A comparison of our estimates with these previous values (Table 5
]) showed that our estimate was significantly lower than the previous estimates, and the reasons for this difference could be related to the following three aspects. First, the influence on ground survey data from different times and regions is the major reason for this difference. For example, we used field sampling data from multi-year field survey data collected by our research group and a large-scale field campaign organized by the GMSC between 2006 and 2012, whereas Ma, et al.
] gathered 113 field samples from 2002–2005, Piao, et al.
] used national grassland resource inventory data between 1982 and 1999. Previous studies have mainly focused on large scales, including all types of grasslands in China or Inner Mongolia, whereas this study only concerns grasslands in Xilingol. Second, the sampling method employed in these studies may have contributed to this difference. Although biomass harvest is a commonly used method of grassland quadrat survey, there are differences in actual practical operations. Different standards of sampling, including sample locations (enclosed or not) and collected objects (including standing dead biomass and litter biomass or not), dramatically influence sampling. The third aspect is the approach taken in the estimation process. For example, Ma, et al.
] calculated the biomass densities for different grassland types based on field samples and further estimated the biomass according to different grassland types. Compared with this method, the method based on remote sensing can represent the spatial details of the aboveground biomass across the entire study area, thereby expanding the study area and reducing the uncertainty of the estimates.
In addition to the studies cited above, several studies have obtained results similar to our results. Jin, et al.
] estimated the average aboveground biomass density of the temperate desert steppe, the temperate steppe and the temperate meadow steppe in in the Xilingol grassland to be 23.1, 55.7 and 98.6 g·m−2
, respectively. Gao, et al.
] used field-based biomass samples and MODIS time series data sets to establish two empirical models based on the relationship of the normalized difference vegetation index (NDVI) to aboveground biomass in the Xilingol grasslands of northern China. The results showed that the average aboveground biomass densities for temperate desert-steppe, temperate steppe and temperate meadow-steppe were 21.2, 59.6 and 111.3 g·m−2
, respectively. The field biomass measurements obtained at the same time and from the same regions contributed to the consistency of these results. However, several differences remained among studies as a result of the use of different remotely sensed data as input parameter and different regression models. The MODIS productivity data is a global land vegetation net productivity product calculated by model, differing from the vegetation index data. Using a remote sensing vegetation index to estimate aboveground biomass provides accurate and rapid results, but the limitations of the vegetation index itself may affect the obtained grassland biomass estimates. Compared with NDVI data, MODIS productivity data can overcome the problem of high error rates in low-coverage grasslands and the oversaturation problem in high-coverage grasslands. In addition, MODIS productivity data can better reflect the effects of environmental stresses. However, the method using MODIS productivity data for biomass estimation should be improved to achieve more accurate estimates of grassland biomass in the future.
This study took the Xilingol grassland as a case study and used ground survey data and MODIS productivity data for the growing seasons of the years 2005–2012 to build a unitary linear regression model to retrieve the aboveground biomass of the Xilingol grassland, and then analyze the spatial and temporal distribution of aboveground biomass. The grassland aboveground biomass averaged 14.35 Tg in the Xilingol grassland during the years 2005–2012, and showed a gradually decreasing trend from east to the west. There were large interannual variations in the aboveground biomass, ranging from 10.56–17.54 Tg, and the aboveground biomass showed differences among different grassland types.
The study made up for the inadequacy of vegetation index to estimate the grassland biomass, and provided an improvement for grassland biomass estimation. Although the study showed promising results for remote sensing based grassland biomass estimation, there are limitations to the accuracy of biomass inversion using MODIS productivity data, therefore, further work is needed to improve the estimation accuracy. In addition, spatio-temporal patterns of aboveground biomass and their relationships with climate factors still need further research in the Xilingol grassland. Gao, et al.
] have already studied the relationship between GSP (growing season total precipitation) with grassland biomass and the relationship between GST (growing season average temperature) with grassland biomass. However, the climate factors are much more than these. For example, how the growing season maximum and minimum temperatures affect grassland biomass and whether the temperature in winter affects grassland biomass need in-depth research. Therefore, the grassland response to climate change is complex, and deserves more detailed and deeper inquiry in future research.