Local spatial interaction modelling based on the geographically weighted regression approach
Local spatial interaction modelling based on the geographically weighted regression approach
复制标题
DOI：
10.1007/978-94-017-2296-4_4
复制
发表时间：
2001
期刊：
影响因子：
2.7
通讯作者：
T. Nakaya
中科院分区：
文献类型：
--
作者：
T. Nakaya
空间分析的一个最近的主要趋势是局部建模，空间分析人员通过这种建模来研究地理现象中的局部特性（Fotheringham，1997年）。事实上，由于地理环境的不同，空间过程往往在空间上有所不同，因此出现了空间非平稳性（Jones III和Hanham，1995年）。在这种情况下，假设普遍接受的属性的全球模型无法捕捉所研究的真实的现象。我们可以说，在疾病绘图中对局部发病率的推断是局部建模的最简单形式（Openshaw等人，1987，Nakaya，2000）。对于更复杂的关联分析，Casetti（1972）的展开方法是流行的，以明确地建模回归分析中的非平稳性（例如Casetti，1990）。根据该方法，我们可以通过位置变量的多项式或调和展开级数来确定回归参数的地理漂移。最近，纽卡斯尔学校（Brunsdon等人，1996年，Fotheringham等人，1998年）开发了一种更通用的局部回归方法，称为地理加权回归（GWR）。该方法利用移动加权核估计局部回归系数。本文的目的是发展的方式，GWR的方法是适用于空间相互作用建模，并提出了一个实证的例子，适用于1980年下半年在日本的移民流动。虽然空间相互作用模型被认为是许多空间数学模型的基本组成部分，但它们对观测流的描述能力较差（Openshaw，1976，1979）。有许多考虑表明，在空间相互作用过程中的非平稳性有助于这个问题。起源地或目的地特定模型是经典的方法，假设重力模型中的解释变量的效果在起源地或目的地之间可能不同。Lovett和Flowerdew（1989）展示了一个例子，在泊松空间相互作用模型中纳入了许多起源或目的地特定的解释变量。将模型分别拟合到流的相对子集是根据反映不同底层过程的不同起源和目的地对来推断参数漂移的有效方法。例如，较长距离和较短距离的移徙主要分别受到就业因素和住房因素的影响。Gordon（1991）提倡多流模型，其中重力模型针对根据原点之间的距离分组的每个迁移数据子集分别进行校准
One of the recent major trends in spatial analysis is local modelling by which spatial analysts examine local properties in geographical phenomena (Fotheringham, 1997). Indeed, spatial processes tend to vary over space due to different geographical contexts so that spatial non-stationarity emerges (Jones III and Hanham, 1995). In such cases, global models that postulate universally acceptable properties fail to capture the real phenomena under study. We could say that inferences of local incidence rates in disease mapping are the simplest form of local modelling (Openshaw et al., 1987, Nakaya, 2000). As for more complicated association analyses, Casetti's (1972) expansion method is popular to model explicitly the property of non-stationarity in regression analysis (eg Casetti, 1990). According to the method, we can specify geographical drifts of regression parameters by polynomial or harmonic expansion series of locational variables. Recently, the Newcastle school (Brunsdon et al., 1996, Fotheringham et al., 1998) has developed a more generalised local regression methodology, called geographically weighted regression (GWR). The approach estimates local regression coefficients with a moving weighting kernel. The aim of this paper is to develop the way in which the GWR approach is applied to spatial interaction modelling and to present an empirical example applied to the migration flows in Japan during the latter half of 1980. Although spatial interaction models are regarded as basic components of a lot of spatial mathematical models, their poor ability to describe the observed flows has been pointed out (Openshaw, 1976, 1979). There are many considerations indicating that the non-stationarities in the spatial interaction process contribute to this problem. Origin or destination specific models are classic approaches that assume the effect of the explanatory variables in gravity models may differ between origins or destinations. Lovett and Flowerdew (1989) show an example to incorporate many origin or destination specific explanatory variables in Poisson spatial interaction modelling. Fitting models separately to relative subsets of flows is an effective way to infer parameter drifts depending on different origin and destination pairs reflecting different underlying processes. For example, longer and shorter distance migrations would be influenced mainly by employment factors and by housing factors, respectively. Gordon (1991) advocates multi-stream modelling in which gravity models are separately calibrated for each subset of migration data grouped by distances between origins