Geographically Weighted Regression Based Estimation Of Regional Forest Carbon Storage

Global climate issues have confirmed the irreplaceable role of forest carbon stocks in the global carbon cycle. To verify whether the geographically weighted regression(GWR) model method which considers the effects of spatial heterogeneity and establishes the local regression model, can improve the estimation accuracy of forest carbon stocks, instead of the more commonly used methods of global regression model such as ordinary least squares analysis(OLS). Xianju County in Zhejiang Province was selected as a study area. by using GWR method,forest management inventory data, combined with Landsat TM image data and DEM data we developed a serials of forest carbon mpdels and apllied to estimating forest carbon stock and its distribution in the study area. Paper first describes the basic principles of GWR method introduced and compared the commonly used Gaussian weight function,Tricube weight function; and optimization selection Bandwidth of these two types weight function.Analysis is included comparison to traditional regression, co-kriging interpolation and sequential Gaussian co-simulation. Taking into account the elevation can be defined as a third dimension of space coordinates and spatial relationships observation point, the geographically and altitudinal weighted regression(GAWR) model is discussed, and with the altitude as the dependent variable contrast GWR model to evaluate its effectiveness; available of geographically and altitudinal weighted regression(GAWR) model was then tested in smooth terrain. Results showed that:1. The forest aboveground carbon density ranged from 0 to 89.964 Mg/hm2 with a mean value of15.555 Mg/hm2 estimated by the GWR(T) model for Xianju County. Meanwhile, the mean forest aboveground carbon density calculated from diameter measurements were 15.854 Mg/hm2, the result from GWR(T) model was lower than diameter measured by 1.716 %, R2 = 0.654(P < 0.01), retain at least 70% of the space heterogeneity; and carbon density distribution was consistent with the actual situation. Geographically weighted regression based estimation of regional forest carbon storage is effective, GWR method to estimate a reasonable result and high precision in estimating carbon stocks in terms of the area.2. The estimated results also had a higher accuracy with the RMSE = 9.802(P < 0.01) than traditional regression method with the RMSE = 15.033(P < 0.01) and co-kriging interpolation method with the RMSE = 16.427(P < 0.01). GWR method can effectively estimate the regional forest aboveground carbon stocks reasonably and accurately. Global regression reflects the average state of regionalized variables, while the GWR model better reflect the true variation of spatial data features of the same time; Cokriging interpolation although retain nearly 70% spatial heterogeneity, but had a lower prediction accuracy; sequence Gaussian co-simulation and geographically Weighted Regression prediction error is small, ranging span, more suitable in forest carbon stocks spatial variability analysis and forecasting.3. Intercept of GWR(T) model, which modeled by James P. LeSage’s program with Tricube weightfunction, ranges in-50.595~147.780, coefficient of Band3 ranges in-0.912 ~ 1.224, not the constant value 26.718、-0.124 in global regression. However the inter-quartile range of local estimates is smaller than 1 standard deviation of the equivalent global parameter, the parameter under study has a certain non-stationary, not significantly.