Comparison of Geographically Weighted Regression and Regression Kriging to Estimate the Spatial Distribution of Aboveground Biomass of Zagros Forests

Aboveground biomass (AGB) of forests is an essential component of the global carbon cycle. Mapping above-ground biomass is important for estimating CO2 emissions, and planning and monitoring of forests and ecosystem productivity. Remote sensing provides wide observations to monitor forest coverage, the Landsat 8 mission provides valuable opportunities for quantifying the distribution of above-ground biomass at moderate spatial resolution across the globe. The combination of the sample plots and image data has been widely used to map forest above-ground biomass at local, regional, national, and global scales. Many predictive methods have been suggested to estimate forest aboveground biomass from sparse sampling points into continuous surfaces, ranging from regression methods such as Geographically Weighted Regression (GWR) and geostatistical methods such as Regression Kriging (RK). Researchers have been particularly interested in understanding the causes and effects in ecosystem functions of spatial autocorrelation and heterogeneity, over the past decade. Where in forestry data include spatial autocorrelation and heterogeneity, the independence and homogeneity assumptions of standard statistical approaches, such as ordinary least squares (OLS), may be violated. Many spatial models (such as Geographically Weighted Regression and Regression Kriging) have been developed in recent years to discuss spatial effects in the relationships between variables. Spatial models can be divided into global and local models depending on the spatial scales used in the modeling process. A global model usually involves, a tool to model spatial autocorrelation between observations in neighboring locations, through either a covariance matrix that can be calculated using a variogram or spatial weight matrix based on neighborhood proximity. Global models, of course, do not well represent spatial differences at any given location and may not be successful in dealing with spatial heterogeneity. By comparison, local models, such as geographically weighted regression, adequate a regression relationship within a given bandwidth for each spatial location using the neighbors. From the relationships between variables, the local models are more useful in exploring locational spatial variation (heterogeneity). In the present study, using a Landsat 8-OLI image, and Geographically Weighted Regression and Regression Kriging modeling were compared for the estimation of aboveground forest biomass. In this study, we gathered aboveground biomass data from a total of 184 (30 × 30 m) sample plots in Zagros forests in the Kohgiluyeh and Boyer-Ahmad Province. The datasets corresponded to the Landsat 8 image pixel values. We applied the species-specific allometric equations for individual trees to estimate forest aboveground biomass. The aboveground biomass at plot-level is simply the summation for all trees within the same plot. The estimates were evaluated by ten-fold cross-validation and performances of the model was evaluated using the coefficient of determination (R2) and relative root mean squared error (RMSE%). The efficiency of the predictions can be described with the scatterplots showing the relationships between the forest above-ground biomass estimates and reference data. Results showed 1) that Geographically Weighted Regression (R2 = 0.61, RMSE%= 22) was a fairly better approach and could provide promising results for the prediction of forest above-ground biomass compared to Regression Kriging (R2 = 0.47, RMSE%= 28) and 2) scatterplots depicted that the problems of overestimation and underestimation for all the prediction were apparent.