欢迎访问林业科学,今天是

林业科学 ›› 2017, Vol. 53 ›› Issue (6): 56-66.doi: 10.11707/j.1001-7488.20170607

• 论文与研究报告 • 上一篇    下一篇

基于GWR的大兴安岭森林立地质量遥感分析

李明泽, 郭鸿郡, 范文义, 甄贞   

  1. 东北林业大学林学院 哈尔滨 150040
  • 收稿日期:2016-09-02 修回日期:2017-05-16 出版日期:2017-06-25 发布日期:2017-07-14
  • 通讯作者: 范文义
  • 基金资助:
    国家自然科学基金项目(31470640)

Remote Sensing Analysis of Forest Site Quality in Daxing'an Mountain Based on GWR

Li Mingze, Guo Hongjun, Fan Wenyi, Zhen Zhen   

  1. College of Forestry, Northeast Forestry University Harbin 150040
  • Received:2016-09-02 Revised:2017-05-16 Online:2017-06-25 Published:2017-07-14

摘要: [目的] 建立遥感信息模型,估算森林地位级指数,对森林立地质量的空间分布进行系统研究,为森林经营管理和营林造林提供数据支持与理论依据。[方法] 以黑龙江省大兴安岭地区为研究区域,基于研究区域的多光谱数据结合地面森林资源清查数据,采用以最小二乘为基础的多元线性模型(全局模型)和以地理加权回归(GWR)为基础局域回归模型提取包括遥感因子土壤修正植被指数(MSVI)、差值植被指数(DVI)和林分因子林分平均胸径(ADBH)、林分郁闭度(FCC)在内的4个因子作为自变量,建立地位级指数全局估算模型和局域估算模型。对比2种方法,最终采用地理加权回归模型来绘制研究区域地位级指数空间分布图,对研究区域的立地质量进行评价与分析,并探索研究区域内森林地位级指数的空间分布状态随地形的变化趋势。采用全局Moran I描述不同空间尺度下模型残差的空间自相关性(以8 km为间隔计算从8 km到80 km)。[结果] 大兴安岭地区地位级指数呈现明显的聚集分布,具体表现为东高西低、北高南低,并且最大值出现在北部区域。遥感因子和林分因子影响森林地位级指数的空间分布。地理加权回归模型的拟合和精度明显好于全局模型,其中全局模型的Radj2为0.48、AIC为1 816、RMSE为1.74,地理加权回归模型的Radj2为0.53、AIC为1 784、RMSE为1.29。通过模型模拟结果和实测值的对比分析发现,地理加权回归模型具有更高的验证精度和更好的拟合效果。基于地理加权回归模型残差分析可知,地理加权回归模型能够很好地解决模型残差的空间自相关性,产生更为理想的模型残差。[结论] 全局模型和局域地理加权回归模型能够较为有效地估算黑龙江省大兴安岭地区森林地位级指数,加入样地位置信息进行回归分析的地理加权回归模型能够更有效地降低数据的空间自相关性,结果更符合传统统计模型中关于残差间相互独立的基本假设,使得建模过程更加科学合理。

关键词: 立地质量, 多光谱遥感, 地理加权回归模型, 多元线性回归模型

Abstract: [Objective] This paper was to establish a model of remote sensing information, and the forest site class index was successfully estimated. The spatial distribution of forest site quality was analyzed systematically and scientifically, which provides certain data support and theoretical basis for forest ecosystem management and afforestation.[Method] In this study, Daxing'an Mountain in Heilongjiang Province was taken as the research area, two types of response variables, including the remote sensing factors (modified soil vegetation index,MSVI;difference vegetation index,DVI) and the stand factors (average diameter at breast height,ADBH; forest canopy closure,FCC) were considered in the modeling processes. Both global and GWR (geographically weighted regression) modeling techniques were utilized to fit the models to evaluate and analyze the site quality of the study area and to explore the spatial distribution of forest site class index along with the changing topography. By comparing the two method, we finally chose the GWR model to map the site class index space distribution. The global Moran I index was used to characterize the spatial autocorrelation of the model residuals at different spatial scales (8 km to 80 km).[Result] The result showed that the spatial distribution of the site class index in Daxing'an Mountain region tended to be a clustered distribution, and a high site quality index appeared in the northeastern part of the study area while the southwestern portion with a low site quality index, also the maximum value was observed in the northern region. Both remote sensing factors and stand factors affect the distribution of forest site class index.The GWR model outperformed the global model in both model fitting and validation performances. The Radj2 of the globe model was 0.48, the AIC was 1 816 with a RMSE of 1.74, while the Radj2 of the GWR model was 0.53, the AIC was 1 784 and the RMSE was 1.29.[Conclusion] Global model and GWR model can effectively estimate forest site class index, the GWR model can solve the spatial autocorrelation of the model residuals, and generate more ideal prediction result, which is feasible to estimate the site class index.

Key words: site quality, multi-spectral remote sensing, geographically weighted regression (GWR) model, multiple linear regression model

中图分类号: