基于GIS的森林调查因子地统计学分析

GIS-based geostatistical analysis of forest inventory factors

空间自相关性是地理学第一定律,基于GIS的地统计学分析为研究森林调查因子的空间相关性和依赖性现象提供了科学方法。紫金山国家森林公园2002年森林资源二类调查主要因子地统计学分析案例研究表明：在100~1 000 m的空间尺度上,10个调查因子存在着较为普遍的空间相关性,并且随着抽样间距的增大,空间自相关MoranI系数逐渐降低。由空间差异引起的结构方差在系统方差中占主导地位,其中海拔、坡度、坡位等立地因子结构方差的比例超过90%,其他因子超过60%。3个立地因子的空间自相关幅度较小（1 950 m）,而林龄、胸径、郁闭度等测树因子的空间自相关阈值较大（〉3 000 m）。不同半方差理论模型单位面积蓄积量内插精度分析表明：在线性、高斯、指数、圆形、球体5个半方差模型当中,指数模型决定系数最高（0.918）,相关系数最高（0.954）,剩余标准差（21.438）和平均相对误差（20.591%）最低。在5个半方差理论模型中,指数模型在拟合精度、预测精度两个方面综合性能最好,与研究地区的斑块状森林景观结构存在着密切的关系。

Spatial autocorrelation is the first law of geography. GIS-based geostatistical analysis provides a scientific method to study the phenomenon of spatial autocorrelation and dependence. Geostatistical analysis of the major forest inventory factors in case study area of Zijinshan National Forest Park in 2002 showed that, at the spatial scale from 100 m to 1 000 m, there existed a common spatial correlation for the 10 investigated factors. With the increase of sample spacing, spatial autocorrelation coefficients of Moran 1 tended to decrease. Structural variance caused by space difference played a primary role in system variance, in which the proportions of structural variance of site factors（ elevation, slope, aspect） to system variance were more than 90 % and other factors over 60 %. Spatial autocorrelation ranges of the three site factors were just 1 950 m, while the ranges of forest measuring factors of age, DBH, tree canopy density were bigger than 3 000 m. Interpolation accuracy analysis of stock volume per unit area of different theoretical models of semi-variance showed that, among the five semi-variance models of the linear, Gaussian, exponential, circular and sphere, the determination coefficient of exponential model was the highest （0. 918 ）, the correlation coefficient was the highest （0. 954）, the residual standard deviation（21. 438 ） and the average relative error（20. 591% ） were the lowest. Evaluated from two aspects of the fitting accuracy and prediction precision, the exponential model outperformed all others, which was closely related with the patched forest landscape structure in the study area.