부산대학교 도서관

Grassland ecosystem is one of the most important terrestrial ecosystems in China. Grassland aboveground biomass (AGB) is not only the material base for maintaining grassland ecosystem, but also the most direct indicator to reflect grassland status. AGB determines the herbivore carrying capacity in grassland ecosystem. Therefore, it is important significance to retrieve AGB to study the regional carbon cycle and the sustainable use of grassland resources. We selected the temperate grassland as the study area, which was one of the most representative grassland types in China. In this study, we developed artificial neural network (ANN) technique for the retrieval of grassland AGB based on multi-temporal remote sensing, topography and meteorological data in this region. Eight variables (normalized difference vegetation index (NDVI), difference vegetation index (DVI), 2-bands enhanced vegetation index (EVI2), soil adjusted vegetation index (SAVI), modified soil adjusted vegetation index (MSAVI), elevation, growing season precipitation (GSP), and growing season temperature (GST)) were combined as the candidate input variables for establishing ANN model. Five vegetation indices (VIs) were calculated by the Moderate Resolution Imaging Spectroradiometer (MODIS) reflectance data, which represented a 8 day composite with a 250m spatial resolution for the period 2006–2010. GSP and GST was interpolated according to DEM by Anusplin software, included daily precipitation and temperature records during 2006 to 2010 from 35 climate stations distributed around the region. Field samples were obtained from multi-year field survey data, primarily in August from 2006 to 2010. The ANN model included three layers, one input layer, one hidden layer and one output layer. The hidden layer had 3 to 14 neurons and used a basic tansig sigmoid transfer function for each neuron. The networks were trained by the Levenberg-Marquardt algorithm and Bayesian algorithm. The results were showed as follow: the use of multi-temporal VIs and geological factors had advantages for AGB retrieval with ANN method. The precision of the two algorithms for estimating AGB based on ANN model was more than 63%. Although the retrieval errors remained larger, the predictions were relatively accurate in comparison to previous studies. Furthermore, ANN model based on the Bayesian algorithm was better than on Levenberg-Marquardt algorithm. It had more stable and reliable training process. Since AGB was influenced by elevation, precipitation and temperature factors, ANN could be used to solve complex nonlinear problems. The ANN is well potential application to the retrieval of temporally grassland production from high-dimensional data. This research provides a reference and technical method for grassland AGB estimation by remote sensing.