Vegetation biomass is one of the key issues for accurate estimation of wetland ecosystems. The seasonal dynamics of vegetation biomass is vital for the study of ecosystem carbon fixation and the carbon cycle. Suaeda salsa is a common plant found in the Yellow River Delta (YRD) that plays an important role in coastal wetland ecosystems. Because S. salsa mainly grows in tidal flats, it is difficult to make long-term field observations, and so there are few systematic estimations of S. salsa biomass.
In this paper, different models, including parametric models and artificial intelligence models, were tested and analyzed for estimating fresh weight of S. salsa biomass based on remote sensing images from the Chinese environmental satellite HJ-1A CCD and measured data. According to the spatial distribution of S. salsa cover type, coverage and growing patterns, 20 plots of 90m×90m were randomly selected for sampling. In each plot, 5 quadrats of 1m×1m (in the four corners and center of the plot) were sampled and measured. The total biomass of each plot was calculated by the average value of those 5 quadrats. Vegetation indices were then extracted and the components of K-L transform and K-T transform were calculated from the preprocessed HJ-1A CCD images. The correlation between biomass fresh weight, dry weight and remote sensing information variables were analyzed to determine the variables that significantly related to the biomass. Finally, parameter and nonparametric models were built based on these significant variables.
The parameter models used in this study include univariate linear, nonlinear regression and stepwise regression models. The non-parameter models used in this research are artificial neural network (ANN) models, including BP (Back propagation) networks, RBF (Radial Basis Function) networks and GRNN (General Regression Neural Network) networks. The optimal model was determined by comparison of the mean relative error (MRE) of regression models and ANN models. The major conclusions are: (1) there are significant correlations between several vegetation indices (RDVI, DVI, RVI, SAVI, MSAVI), the second component of the K-L transformation and S. salsa fresh biomass. Although there are no significant correlations between dry biomass of S. salsa and the remote sensing information variables, the trend of correlation is generally consistent. (2) It is feasible to build S. salsa fresh biomass models of the initial growing season by remote sensing information variables with regression models such as a univariate linear model of RDVI, a curve regression model of SMAVI and a multivariate stepwise regression model. However, of all the regression models, the linear regression model of three variables, RDVI, MSAVI and PC2, represents the best model for S. salsa fresh biomass. (3) The BP network model is the optimal model for estimation of S. salsa fresh biomass. The MRE of BP network model is 12.73%, which is 8.11% lower than the traditional multivariate stepwise regression model.