Estimating Canopy Characteristics of Inner Mongolia ’ s Grasslands from Field Spectrometry

This study was designed to estimate the canopy biophysical characteristics of semi-arid grassland ecosystems by using in situ field spectrometry measurements to identify important spectral information for predictions at broader spatial scales. Spectral vegetation indices (VIs), reflectance spectra, continuum removal spectra, and the amplitude of the red edge peak (drre) based on 61 well-replicated field measurements across a large area in Inner Mongolia were used to develop empirical models for estimating four key canopy biophysical features: percent green coverage (PGC), canopy height (H), green aboveground biomass (GBM), and total aboveground biomass (TBM). The results showed that NDVI, EVI, NDSVI, and LSWI were useful for estimating canopy biophysical features, with NDSVI being the most significant variable. The PGC was accurately estimated with spectral reflectance at 441 nm and 2220 nm (R 2 = 0.71), while the maximum depth of band (Dc), absorption area (Darea) in the red domain and drre were selected for estimating TBM and GBM (R 2 = 0.51 and 0.44). Among the four canopy features, PGC received the highest confidence from all of the models (R 2 = 0.81), while H was the most difficult to estimate (R 2 = 0.49). Finally, the degree of disturbances and ecosystem types appeared to be a significant variable for model development. OPEN ACCESS Remote Sens. 2014, 6 2240


Introduction
Accurate quantification of vegetation canopy characteristics has been a research focus for its crucial role in ecosystem studies on productivity, carbon cycles, nutrient allocation, and biological diversity [1][2][3].Despite previous efforts to quantify canopy biophysical properties using various methods at multiple spatial scales, the task remains challenging [3].Remote sensing technology, meanwhile, viewed as a time-and cost-efficient approach among the most promising methods, allows scientists to proceed with large-scale monitoring of ecological properties at broader scales [4,5] and has been widely used to estimate the biophysical characteristics of different vegetation types on landscape and regional scales.However, only a limited number of applications of hyperspectral field spectroscopy data have been promoted due to their capability of quantifying land surface properties, such as canopy cover, height, leaf area index (LAI), etc.
Hyperspectral scanners consisting of individual channels with high spectral resolution at 10 nm and 400-2500 nm spans have the potential to significantly improve our ability to discriminate between the characteristics of spectral features for analysis, classification, and monitoring of canopy properties [6,7].Different vegetation types have distinct pigment concentrations, leaf characteristics, and canopy structures, indicating the existence of unique spectral properties of absorption and reflectance for each species or vegetation type [7][8][9][10].Clearly, narrow bands of hyperspectral remote sensing are more useful than broad bands in the quantification of vegetative biophysical variables; in particular, the red portion of the electromagnetic spectrum might provide more stable information on canopy attributes [11].
Hyperspectral studies have been widely conducted in laboratories, aiming toward their ultimate use in actual fields.Several challenges remain for their effective use.First, canopy reflectance in the field is complex for its similarity of biochemical composition among species, overlapping of their absorption features for biochemical constituents in live vegetation [7], high atmospheric absorption interruption [12][13][14], and variation of vegetation canopy water content caused by incident precipitation that is particularly vital in semi-arid vegetation [12].Secondly, there is no consensus about the relationship between canopy properties and the reflectance that can be used to scale up to the entire canopy because each measure within the sensor view angle includes variable vegetation, mosaics of the soil background, and other non-photosynthetically active vegetation [5,12,15,16].Finally, spectral variations may also occur for the same species under different conditions of age, microclimate, soil characteristics, precipitation, topography, phenology, and stressors [17][18][19][20][21][22].Clearly, substantial and extensive in situ spectral data are needed to promote the uses of hyperspectral reflectance from a remote sensing perspective before it can be linked with satellite data for applications at broader temporal and spatial scales [14,23].
Here, an intensive field campaign was conducted to explore the usefulness of in situ hyperspectral data across a large area of semi-arid landscape in Inner Mongolia, China.Grasslands are the most dominant ecosystems on the Mongolian Plateau [24] and the most important natural resources for its large livestock population and economic development.Grassland degradation in recent years has occurred due to increasing climate extremes (e.g., cold winter) [25] and growing human pressures [26].Therefore, accurate monitoring and estimating canopy biophysical features of grasslands by using high-resolution spectral measurements becomes an urgent demand.
In this study, our objectives were set to: (1) investigate the potential uses of field spectrometry in estimating the biophysical characteristics of different grassland types with different human disturbances across a large spatial area in the arid and semi-arid areas of Inner Mongolia; (2) depict the sensitive spectral properties and any quantitative measures generated from these spectral properties that can be used to estimate biophysical characteristics by developing empirical models; and (3) develop empirical models by identifying significant remote sensing signals to estimate community functions such as biomass.

Study Area
The study area was located between 38.28°-46.60°Nand 107.65°-118.12°E in Inner Mongolia, China (Figure 1).The region is characterized by an arid and semi-arid continental climate with a strong climate gradient and various land use practices.The annual precipitation ranges from 150 mm to 400 mm and the dominant vegetation types include a typical steppe and a desert steppe [27], both under intensive human disturbances.Among the 61 sampling sites, 28 were measured in 2006 and 33 were measured in 2007.Considering the variations in grassland types and human activities, 11 sites were recorded as having -low human disturbance‖ for the typical steppes (i.e., Class I) and 24 sites as having -intermediate human disturbance‖ (i.e., Class II).The remaining 26 sites were in the desert steppe and assigned as Class III.The typical steppe has an annual precipitation of 250-400 mm, annual accumulated temperature of 2600-2800 °C, and is dominated by Stipa krylovii, Leymus chinensis, Cleistogenes squarrosa, Koeleria cristata, Artemisia spp., and Potentilla spp., while the desert steppe has an annual precipitation of 150-250 mm, annual accumulated temperature of 2200-2500 °C, and is dominated by Stipa klemenvii, S. breviflora, and S. glareosa [27].

Field Spectral Measurement
Field reflectance spectra measurements were conducted from 11 July to 25 August 2006 and 25 June to 24 July 2007.All spectral measurements were taken between 10.5 h and 14.5 h on clear and sunny days.A GER3700 (Geophysical and Environmental Research Corporation, Buffalo, NY, USA) spectroradiometer was used for taking in situ spectral data.The GER 3700 is a three-dispersion grating spectroradiometer using Si and PbS detectors with a single field of view of 25°.It has 647 wave bands and a wavelength range of 350-2500 nm with sampling intervals of 1.5 nm in the 350-1050 nm, 6.2 nm in the 1050-1900 nm, and 9.5 nm in the 1900-2500 nm.The spectral resolution is 3 nm over the range of 300-1050 nm, 11 nm over the range of 1050-1900 nm, and 16 nm over the range of 1900-2500 nm.The sensor was hand-held approximately one meter above the ground at the nadir position.Three replicates were conducted at each plot and five samplings (north, south, east, west, and

Vegetation Sampling
After taking a spectral measurement, percent green cover (PGC, %), canopy height (H, cm), green aboveground biomass (GBM, g•m −2 ), and total aboveground biomass (TBM, g•m −2 ) were measured or harvested using three 50 cm × 50 cm plots.PGC is the percentage of the ground area that is covered by green leaves and measured through visual estimations.The harvested green plants and non-green plants were oven-dried at 65 °C for 48 h and then weighed for green aboveground biomass (GBM) and total (green plus non-green) aboveground biomass (TBM).

Data Analysis
A moving Savitzky-Golay filter [28] with a frame size of 15 (second-degree polynomial) was applied to the averaged reflectance spectra to eliminate noise from the sensor using MATLAB 2007.Moreover, the reflectance of the following wavelengths was excluded due to high levels of noise or atmospheric water vapor absorption: <400 nm, >2400 nm, 1350-1590 nm, and 1780-1970 nm [7,12].Ultimately, 517 valid wavebands were obtained from each spectrum.
The amplitude of the red edge peak (dr re ) was determined as the maximum first derivative of the reflectance between 680 and 750 nm [34].The continuum removal analysis, a way of normalizing the reflectance spectra, was applied for the absorption features from 550 nm to 750 nm in the visible domain, the maximum band depth (D c ), the area under the curve (D area ), and the normalization of D c by the absorption feature (D c /D area ) were calculated as follows: the D c was the maximum band depth (BD) at the band center.The BD was calculated by subtracting the continuum-removed reflectance at a certain wavelength from 1 and, thus, the value of the BD is between zero and one.The continuum-removed reflectance is obtained by dividing the reflectance value for each waveband in the absorption feature by the reflectance level of the continuum line at the corresponding wavelength.The D area was calculated as the integration of the BD over the spectral interval of interest.The D c /D area was calculated by dividing the D c by D area , i.e., the normalization of the D c by the absorption feature [35][36][37].

Statistical Analysis
Radiative transfer models, being functions of canopy, leaf, and soil background characteristics, had theoretically physical foundations.However, they had the ill-posed nature of model inversion [38].Statistical approaches, being sensor-specific and dependent on site and sampling conditions, but easy to implement, were used in this study.Descriptive statistics and univariate and multivariate analyses were applied to initially explore the statistical properties of each dataset.Although, the -full spectrum‖ methods, such as partial least squares regression, had better predictive powers for biophysical variables [39].Here, a stepwise linear regression, which can establish a minimum set of wavelengths and permit an optimum modeling of the target variable, was used to select wavelengths from spectral property variables.To ensure the normality of residuals and to improve the statistical models, vegetation parameters were transformed by natural logarithmic and square root arithmetic and a general linear model was used to examine the effects of class variables on the grass canopy properties (SAS 9.1 TS level XP-PRO platform).In this model, spectral information and the class variable were set as the independent variables, while canopy characteristics were set as the dependent variables.

Variation in the Canopy Properties and Reflectance Spectrum
The canopy characteristics among the sampled grasslands were highly variable (Table 1).PGC ranged from 3.0% to 78.8% with an average (standard deviation or SD) of 28.0(21.77)%,H ranged Vegetation reflectance based on spectral measurements also showed a considerable difference between the two years and among the three vegetation classes (Figure 2).The reflectance spectra in 2007 were higher than those in 2006 in visible and near-infrared regions, excluding shortwave infrared regions.Class III, mostly in the desert steppe with short and sparse vegetation, had higher reflectance in the visible and near-infrared domains than Class I and Class II in typical steppes (Figure 2).Additionally, there were noticeable spectral differences between Class I and Class II for their different disturbance degrees, especially in the visible region.The reflectance in the second shortwave infrared domain (1900-2450 nm) of Class II was higher than that of Class I, which was higher than that of Class III.This pattern corresponded well to the canopy development conditions (Table 1).The visible region captures the changes of chlorophyll absorption, the near-infrared region reflects optical properties of leaves, the first shortwave infrared region (1500-1900 nm) is related to leaf biochemical characteristics, and the second shortwave infrared bands for the development of cellulose and lignin as a plant ages.

Estimations from VIs
Among the four vegetation indices, NDSVI and EVI were selected as the significant independent variables through stepwise regression for estimating transformed PGC (R 2 = 0.64; p < 0.001), while NDSVI alone was selected to estimate transformed GBM and TBM (R 2 = 0.24, 0.32 and p < 0.001, 0.001, respectively).The transformed H was linearly correlated with NDVI (R 2 = 0.40), while NDSVI correlated well with transformed GBM and TBM, with R 2 of 0.24 and 0.32, respectively (Figure 3).(c) (d)

Estimations from Reflectance Spectra
Using the Pearson product moment correlation function for all wavelengths of spectral measurements, we identified several spectra that had high correlations with canopy biophysical properties (Figure 4), including reflectance at 421 nm, 441 nm, 676 nm, 723 nm, 993 nm, 1210 nm, 1640 nm, and 2220 nm.These spectra were consequently selected to explore their roles in estimating PGC, H, GBM, and TBM (Figure 5) through stepwise regression.PGC could be estimated from spectra in visible (400-700 nm) and part-infrared (700-1200 nm) regions.Our confidence level for estimating transformed H, GBM, and TBM varied from 39% to 44% (Figure 5).Overall, it seemed that the canopy characteristics, except for H, could be more confidently estimated by reflectance spectra than by using VIs.Reflectance at 441 nm and 2220 nm explained 71% of the transformed PGC (Figure 5).

Estimations from Continuum-Removal Spectra and dr re
The predictive powers of canopy characteristics by continuum removal spectra and dr re were further improved for GBM and TBM (Table 2), with R 2 values for transformed PGC, H, GBM, and TBM that varied from 0.43 to 0.54.

Estimations Using All Independent Variables
Using all of the independent variables, including VIs, reflectance spectra, continuum removal spectra, and dr re variables, the predictive powers for canopy characteristics were further improved (Table 3).The R 2 value for transformed PGC from dr re , reflectance at 441 nm and 1210 nm was 0.72, while it was 0.55 for transformed GBM from dr re , NDVI, reflectance at 421 nm and 723 nm, and 0.57 for TBM from dr re , reflectance at 723 nm and NDVI.However, the R 2 value for transformed H remained around 0.40.

Estimations by Disturbances and Ecosystem Type
Disturbance intensity played a significant role in the development of our empirical models for estimating canopy features (Table 4).The R 2 value for estimating transformed PGC increased from 0.71 to 0.81 when disturbance intensity and grassland type were included as category variables in our linear models (Figure 5, Table 4).More importantly, inclusion of class variable in our model development reduced the number of drivers, while R 2 increased from 0.51 to 0.63 for TBM, from 0.44 to 0.59 for GBM, and from 0.40 to 0.49 for H (Tables 2 and 4, Figure 5).

Discussions
In this study, VIs, reflectance spectra, continuum removal spectra, and dr re based on 61 field measurements across a large area in Inner Mongolia were used to develop empirical models for estimating four key canopy characteristics: PGC, H, GBM, and TBM.Our analysis indicated that confidence models could be developed with careful selection of spectral drivers.These models, using different sets of independent drivers, provide us with alternative options (Sections 3.2-3.6).Evaluation of model performances is an essential premise; however, there is still a lack of a generally agreed upon set of -optimal‖ statistical measures for assessing model accuracy [40].Wide acceptance and the use of a coefficient of determination (R 2 ) were utilized in this study.
The conventional biometric approaches for quantifying canopy characteristics in the field are very accurate; however, they are labor-intensive and destructive, resulting in only a small number of communities that can be measured.Remote sensing technology (e.g., the hand-held spectrometer), consequently, becomes an appealing alternative due to its non-destructive, speedy sampling features and coverage of large areas.Mutanga et al. used the field spectrometer successfully to estimate canopy nitrogen, phosphorus, potassium, calcium, and magnesium concentrations in a savanna rangeland based on continuum-removed absorption features and succeeded with a R 2 of 0.43-0.80[14].Similarly, Haboudane et al. detected canopy pigment levels (i.e., chlorophyll and carotenes) with a confidence level of >80% [41].In this study, we found the highest confidence in estimating PGC and satisfactory confidence for GBM and TBM.The most difficult task remains the prediction of H. Nevertheless, it appeared that the hyperspectral spectra in this study contained rich information about canopy characteristics.
Numerous studies have shown that remotely sensed VIs have close correlations with canopy biophysical characteristics and, at the same time, they are less sensitive to other variables such as canopy geometry, soil backgrounds, solar zenith angles, and atmospheric conditions [11,42].One of our objectives in this study was to examine the feasibility of utilizing these widely used and available indices in the remote sensing community to estimate the canopy biophysical characteristics, which can be scaled up to broader spatial and temporal scales.Here, we confirmed that VIs could be effectively applied for such a purpose (Figure 3, Tables 3 and 4).Surprisingly, LSWI was not selected during the stepwise regression as an important variable.This is likely due to the fact that the region is very dry in the summer months and the surface moisture remains low and similar [43].Our success is also due to the fact that we did not have the saturation problem of receiving reflectance data from sparse or short vegetation.
Hypothetically, canopy characteristics could be estimated by using proper spectral information.We found that there was a great potential for the use of reflectance spectra in estimating PGC, but this use was questionable for H. Several authors concluded that the red-edge region was not sensitive to atmospheric or background effects [14,44], regardless of its high correlations with aboveground biomass and leaf area [45].Moreover, derivative spectra are commonly used to suppress background signals in estimating a range of canopy properties [12], including the chlorophyll absorption feature and the water absorption features [46].In addition, a study based on a laboratory experiment also showed that the red edge inflection point had a very poor relationship with LAI, particularly when the plant species were pooled together [47].In this study, the predictive performance of dr re was weaker than that of D c (Table 2) when constructing empirical models based on continuum-removed spectra and dr re to estimate canopy properties, likely due to dry climate conditions in arid and semi-arid regions [48].However, when considering all possible spectral information, dr re was an essential supplement for detecting PGC, TBM, and GBM (Tables 3 and 4).Clearly, the importance of the red-edge region for estimating vegetation biophysical characteristics in arid and semi-arid regions needs further investigations.
The continuum removal approach has been effectively applied to improve the estimation of biomass [7,14,35,36].The amount of green biomass, the cumulative product of photosynthetic abilities and ecosystem respiration, is a function of the absorption feature of electromagnetic energy at specific wavelengths in the visible region [14,49].In this study, we used continuum removal on red absorption features and dr re to build regression models for GBM and TBM and found that absorption features in the red region were more closely related to biomass.As suggested by Mutanga et al., the technique of continuum removal on absorption features may also be applied to data acquired by airborne and space-borne imaging spectrometers with high-quality radiometric and geometric calibration of hyperspectral imagery to predict and ultimately map the quality and the quantity of grasslands [14].Underwood et al. also showed that the processing techniques of continuum removal performed better for classifying one non-native plant of the target species using NASA's hyperspectral imagery AVIRIS (Airborne Visible/Infrared Imaging Spectrometer) [50].The best empirical models in this study, however, were developed using all of the above spectral information (Table 3), suggesting that a combination of individual spectrum, quantitative indices, and continuum removal approaches is needed.More importantly, selected independent variables for each combination seemed to vary by dependent variable, indicating different sources for our estimations.
A major finding of this study is the importance of including disturbances in predictive models, which were significantly improved from their inclusion (Table 4).Logically, the structural evenness of the sampling community will largely determine vegetation reflectance.Disturbances, depending on their intensity, type and frequency, will increase the heterogeneity of the vegetation.In a highly heterogeneous Mediterranean grassland, Darvishzadeh et al. found that canopy chlorophyll content could be estimated with the highest accuracy (R 2 = 0.70) and leaf chlorophyll content could not be estimated with acceptable accuracy, while LAI was estimated with intermediate accuracy (R 2 = 0.59) [51].For estimating PGC in this study, the confidence level was increased to 81% from 71% when disturbance was considered in the analysis.Similar degrees of model improvement were also found for TBM and GBM (Tables 3 and 4).Röder et al. pointed out that an estimation of biochemical and biophysical characteristics in heterogeneous grasslands with mixtures of different grass species can be challenging in remote sensing [52].However, Darvishzadeh et al. also demonstrated that by using hyperspectral measurements with a large number of narrow spectral bands and powerful multivariate regression techniques, biophysical, and, to a lesser extent, biochemical grass characteristics could be retrieved with acceptable accuracy [53].
Finally, limited spectral measurement in a large area with higher spatial heterogeneity may affect the correlation between spectral properties and canopy features as well as the predictive performances of biophysical features based on hyperspectral field spectroscopy.Therefore, when measuring field spectral reflectance, each sampled area should meet a certain scale to be representative of spectral properties of this community type, which can be done by increasing the number of repetitive or homogeneous measurement-a necessity for improving the predictive accuracies based on field spectrometry.

Conclusions
Previous studies on the utility of hyperspectral field spectroscopy for detecting canopy properties such as LAI, chlorophyll, and foliar nutrients were conducted mainly in laboratory conditions or in small field regions (e.g., plot and stand levels).Here, we conducted intensive field campaigns across the semi-arid region of Inner Mongolia with the goal that the lessons learned from this study can be applied later at broader spatial and temporal scales.We found that: (1) quantitative VIs of NDVI, EVI, NDSVI, and LSWI were useful for estimating canopy biophysical features, with NDSVI being the most significant variable; (2) PGC in our region was accurately estimated with spectral reflectance, while D c , D area , and dr re were selected for estimating TBM and GBM; (3) among the four canopy features, PGC received the highest confidence from all of the models, while H was the most difficult parameter to predict; and (4) the degree of disturbances is a very significant variable in developing empirical models using remote sensing measurements.
center positions) were conducted in each of the 61 sites within a 250 m × 250 m area.Altogether, we had 915 measurements in 2006 and 2007.

Figure 1 .
Figure 1.Spatial locations of 61 sampling sites in the study area overlaid on land cover types in Inner Mongolia, China.The land cover types were from the 1:4,000,000 Vegetation Atlas of China compiled by the Editorial Board of Vegetation Maps of China, 2001.

Figure 2 .
Figure 2. Changes in spectral reflectance of (a) 2006 and 2007; and (b) Class I, Class II, and Class III in Inner Mongolia.

Table 1 .
Summary statistics of the biophysical characteristics from 61 sampled communities in Inner Mongolia.

Table 2 .
Empirical models and performances from stepwise regressions using continuum-removed spectra and dr re to estimate canopy properties (n = 61).

Table 3 .
Empirical models and performances of stepwise regressions based on vegetation indices (VIs), reflectance spectra, continuum removal spectra, and dr re (n = 61).

Table 4 .
Empirical models and performances of the general linear model with disturbance as a class variable (n = 61).Class I: less human disturbances in the typical steppe, Class II: more disturbances in the typical steppe, and Class III: desert steppe.