An evaluation model for aboveground biomass based on Hyperspectral Data from field and TM8 in Khorchin grassland, China

Biomass is an important indicator for monitoring vegetation degradation and productivity. This study tests the applicability of Hyperspectral Remote-Sensing in situ measurements for high-precision estimation aboveground biomass (AGB) on regional scales of Khorchin grassland landscape in Inner Mongolia, China. Field experiments were carried out which collected hyperspectral data with a portable visible/NIR hyperspectral spectrometer (SOC 710), and collected aboveground net primary productivity (ANPP). Ground spectral models were then developed to estimate ANPP from the normalized difference vegetation index (NDVI), which was measured in the field following the same method as that of the Thematic Mapper(TM) from the Landsat 8 land imager (TM_NDVI). Regression analysis was used to assess the relationship between ANPP and NDVI based on coefficients of determination (R2) and error analysis. The estimation of ANPP had unique optimal regression models. By comparing the different spectral inversion models, we selected an exponential model associating ANPP with NDVI (ANPP = 12.523*e3.370*(0.462*TM_NDVI+0.413), standard error = 24.74 g m-2, R2 = 0.636, P < 0.001). This study suggests that the model can be used to monitor the condition and estimate the productivity of grassland at regional scales. The results still show a high potential to map grassland degradation proxies on the ground hyperspectral model. Thus, this study presents biomass hyperspectral inversion technology to remotely detect and monitor grassland degradation and productivity at high precision.


32
The tools for remotely sensing of vegetation have evolved significantly in recent decades [1], and 33 spectral imaging has become increasingly popular in remote-sensing research for correlating 34 spectral data with the biophysical properties of vegetation. Hyperspectral remote-sensing data 35 have subsequently been widely used to estimate vegetation biomass [2][3][4][5], vegetation cover

37
The accurate estimation of aboveground net primary productivity (ANPP) is an active area of 38 research and can provide valuable information about the productivity and ecosystem service value 39 of grassland [10]. ANPP is an important impact factor for desertification and are often used as 40 indicators for monitoring and evaluating grassland productivity and degradation [11]. The present 41 study aimed to develop models for estimating biomass and VC based on satellite data, which 42 allow an assessment over large areas at a low cost [12]. Desertification in the Khorchin grassland 43 is becoming worse with the rapid expansion of the population, overgrazing and the warmer and 44 drier trend associated with climate change [13]. The accurate estimation of the biomass of the 45 grassland over large areas using remotely sensed data is thus very important for monitoring 46 desertification and for improving the scientific management of grassland ecological resources.

98
This study was carried out in the field of Khorchin grassland which was State-owned Land and did not 99 involve endangered or protected species. Meanwhile, because this study supported by National

114
Field spectral data 115 The field data were collected using the SOC710 Hyperspectral Imaging System which 116 Manufactured by Surface Optics Corporation in America. The SOC710 is a precision instrument 117 with an integrated scanning system and analysis software that can quickly obtain high-quality 118 hyperspectral images at visible to near-infrared (NIR) wavelengths in the range 0.4-1.0 µm.

160
(3) n y y y where y is a measured biomass, y′ is an estimated biomass for the test data, and n is the number of 162 samples.

164
Optimal ground spectral models and tests of model accuracy 165 The optimal ground spectral models for biomass 166 From the analysis and evaluation of the relationships between ANPP and SOC_NDVI computed 167 from reflectance data obtained by the SOC710 in the field, we chose linear, logarithmic, power,

168
and exponential functions to fit and optimise the regression equations for selecting the best 169 regression model (Fig.2). 170 171

172
The relationships between ANPP and SOC_NDVI was significant (P < 0.001) for all functions 173 and met the assumptions of the statistical analyses. The exponential model was superior for ANPP,

174
with an R 2 of 0.636, indicated by bold type in Tables 1. 175

177
Tests of model accuracy 178 The accuracy of the models was tested to obtain the best regression models. We used test sets of 179 all field samples to analyse and evaluate the errors in the regression models (

Spectral inversion models 205
The spectral inversion models of TM8 for ANPP was calculated by Eqs. 4-5: To test the agreement between measured and predicted values, we applied Eqs. 6 to the TM8 208 NDVI greyscale image and obtained the patterns of ANPP distribution in the study area by grid 209 computing. The test data sets were then converted into vector diagrams defined by geographic 210 coordinates by geographic information system. The values at the test points were recorded in the 211 distribution patterns as the corresponding pixels predicting values of ANPP. The relationship 212 between actual and predicted values was used to evaluate the accuracy of model.

213
The correlation between the predicted and actual values was significant, as were the independent 214 validations for predicting biomass (SE = 24.74, MEC = 18.61%; Fig.4). This study suggested that 215 the spectral inversion models could be used to monitor grassland biomass at regional scales.

220
The main goal of this study was to establish more accurate models for estimating ANPP of the 221 grassland in Khorchin. We chose the NDVI vegetation index, which can be calculated from 222 spectral reflectance data acquired in the field and from data from Landsat TM7 Band 4 (TM4; Hyperspectral Imaging System, which is more accurate than the FieldSpec3 spectroradiometer.

230
The TM8 remotely sensed imaging data were only released in 2013, so they have not yet been 231 widely applied to monitor vegetational biomass. This study applied the field data for monitoring 232 the vegetation, thereby providing an informational baseline for this study area. The spectral 233 inversion model was ideal, indicating that TM8 remote imaging can be used for research on 234 vegetation biomass on a regional scale.

235
ANPP have their own optimal regression models based on the processing and statistical analysis 236 of experimental data in the study area. The optimal equations for the estimation of ANPP (Fig.5) 237 indicate that the relationship between SOC_NDVI and ANPP weakens at biomass >350 g m -2 for 238 grassland. Estimates of biomass above these levels are inaccurate or unreliable and may be The ground spectral models for ANPP can be applied to TM8 images, because measured spectral 246 characteristics of plants on the ground are intrinsically linked to those obtained by TM8 remote 247 sensing. Grassland yield over large areas can be estimated based on the ground spectral model.

248
The models, however, could be more accurate if field and satellite data are collected over several 249 years rather than only for one year. Also, the field and satellite data should be acquired at the same 250 time for maximal correspondence. In future field experiments, we will assess the collective 251 influence of these vegetational characteristics and the NDVI on biomass prediction and will seek 252 to obtain a modified NDVI for estimating the biomass of dense vegetation under natural 253 conditions.

255
This study developed a relatively accurate model for estimating AGB and tests the applicability of 256 hyperspectral data from field and TM8 to map AGB on regional scales by a regression analysis 257 method. The methodology we adopted in the study was a first attempt to Retrieval of vegetation 258 biomass from ground hyperspectral remote sensing in Khorchin grassland. 259 The accuracy of ground spectral inversion is affected by many factors, and the quality of the selected 8 260 remote sensing image data has the greatest impact on the fitting accuracy of the model. Landsat 8 261 satellite data is selected for remote sensing data, which has higher geometric accuracy and 262 signal-to-noise ratio than previous Landsat data, which effectively expands the application range of 263 image data. In the aspect of imaging mode, the sweep pendulum design of OLI imager has good 264 stability and improves the image quality, and in the aspect of geometric accuracy, L1T data product is a 265 data product after precise correction, and the product accuracy has been greatly improved. In this paper,

266
TM8 data is used to retrieve vegetation biomass, and the results show that calculated R 2 and SE and 267 MEC values for various regression models vary among ground spectral models. By comparison,

268
the exponential regression models we developed show a stronger relationship between spectral 269 reflectance and ANPP. An exponential equation was optimal for estimating ANPP in the Khorchin 270 grassland. Accuracy verification indicated that the relationship between the actual and predicted 271 biomass was significant. Estimating ANPP with high accuracy based on NDVI derived from TM8 272 satellite data is thus possible, which accumulates experience for the application of TM8 data in 273 vegetation monitoring field.

274
The accuracy of this technique depends on living, green biomass and not on senesced or dead