An assessment of coefficient accuracy in linear regression models with spatially varying coefficients

The realization in the statistical and geographical sciences that a relationship between an explanatory variable and a response variable in a linear regression model is not always constant across a study area has led to the development of regression models that allow for spatially varying coefficients. Two competing models of this type are geographically weighted regression (GWR) and Bayesian regression models with spatially varying coefficient processes (SVCP). In the application of these spatially varying coefficient models, marginal inference on the regression coefficient spatial processes is typically of primary interest. In light of this fact, there is a need to assess the validity of such marginal inferences, since these inferences may be misleading in the presence of explanatory variable collinearity. In this paper, we present the results of a simulation study designed to evaluate the sensitivity of the spatially varying coefficients in the competing models to various levels of collinearity. The simulation study results show that the Bayesian regression model produces more accurate inferences on the regression coefficients than does GWR. In addition, the Bayesian regression model is overall fairly robust in terms of marginal coefficient inference to moderate levels of collinearity, and degrades less substantially than GWR with strong collinearity.     

Dynamic land use change simulation using cellular automata with spatially nonstationary transition rules

The dynamic relationships between land use change and its driving forces vary spatially and can be identified by geographically weighted regression (GWR). We present a novel cellular automata (GWR-CA) model that incorporates GWR-derived spatially varying relationships to simulate land use change. Our GWR-CA model is characterized by spatially nonstationary transition rules that fully address local interactions in land use change. More importantly, each driving factor in our GWR model contains effects that both promote and resist land use change. We applied GWR-CA to simulate rapid land use change in Suzhou City on the Yangtze River Delta from 2000 to 2015. The GWR coefficients were visualized to highlight their spatial patterns and local variation, which are closely associated with their effects on land use change. The transition rules indicate low land conversion potential in the city’s center and outer suburbs, but higher land conversion potential in the inner near suburbs along the belt expressway. Residual statistics show that GWR fits the input data better than logistic regression (LR). Compared with an LR-based CA model, GWR-CA improves overall accuracy by 4.1% and captures 5.5% more urban growth, suggesting that GWR-CA may be superior in modeling land use change. Our results demonstrate that the GWR-CA model is effective in capturing spatially varying land transition rules to produce more realistic results, and is suitable for simulating land use change and urban expansion in rapidly urbanizing regions.     

Extreme Coefficients in Geographically Weighted Regression and Their Effects on Mapping

This study deals with the issue of extreme coefficients in geographically weighted regression (GWR) and their effects on mapping coefficients using three datasets with different spatial resolutions. We found that although GWR yields extreme coefficients regardless of the resolution of the dataset or types of kernel function: (1) GWR tends to generate extreme coefficients for less spatially dense datasets; (2) coefficient maps based on polygon data representing aggregated areal units are more sensitive to extreme coefficients; and (3) coefficient maps using bandwidths generated by a fixed calibration procedure are more vulnerable to the extreme coefficients than adaptive calibration.     

Multicollinearity and correlation among local regression coefficients in geographically weighted regression

Present methodological research on geographically weighted regression (GWR) focuses primarily on extensions of the basic GWR model, while ignoring well-established diagnostics tests commonly used in standard global regression analysis. This paper investigates multicollinearity issues surrounding the local GWR coefficients at a single location and the overall correlation between GWR coefficients associated with two different exogenous variables. Results indicate that the local regression coefficients are potentially collinear even if the underlying exogenous variables in the data generating process are uncorrelated. Based on these findings, applied GWR research should practice caution in substantively interpreting the spatial patterns of local GWR coefficients. An empirical disease-mapping example is used to motivate the GWR multicollinearity problem. Controlled experiments are performed to systematically explore coefficient dependency issues in GWR. These experiments specify global models that use eigenvectors from a spatial link matrix as exogenous variables.This study was supported by grant number 1 R1 CA95982-01, Geographic-Based Research in Cancer Control and Epidermiology, from the National Cancer Institute. The author thank the anonymous reviewers and the editor for their helpful comments.     

Alleviating the effect of collinearity in geographically weighted regression

Geographically weighted regression (GWR) is a popular technique to deal with spatially varying relationships between a response variable and predictors. Problems, however, have been pointed out (see Wheeler and Tiefelsdorf in J Geogr Syst 7(2):161–187, 2005), which appear to be related to locally poor designs, with severe impact on the estimation of coefficients. Different remedies have been proposed. We propose two regularization methods. The first one is generalized ridge regression, which can also be seen as an empirical Bayes method. We show that it can be implemented using ordinary GWR software with an appropriate choice of the weights. The second one augments the local sample as needed while running GWR. We illustrate both methods along with ordinary GWR on an example of housing prices in the city of Bilbao (Spain) and using simulations.     

Spatially non-stationary effect of underlying driving factors on surface urban heat islands in global major cities

Urban heat island (UHI) effect is among the most typical characteristics of urban climate. The analysis of surface UHI (SUHI) mechanisms has received the most extensive attention in the world. Here, we quantify the diurnal and seasonal SUHI intensity (SUHII) in global 419 major cities during the period 2003-2013. A geographically weighted regression (GWR) was established to assess the relationships between SUHII and several driving factors, and it further was compared to the ordinary least square (OLS) and stepwise multiple linear regression (SMLR) models. We show that GWR model has higher determination coefficient (R²) than OLS and SMLR models (Time: summer daytime, summer night, winter daytime and winter nighttime; GWR: 0.805, 0.458, 0.699 and 0.582; OLS: 0.732, 0.347, 0.473 and 0.320; SMLR: 0.732, 0.341, 0.468 and 0.316), indicating the spatially non-stationarity in the relationships. During the day, both vegetation activity and tree cover fraction have stronger cooling effect on SUHI in the summer of Asia. At night, there are stronger albedo effects on SUHI in the summer of Eastern Asia and Western North America and in the winter of Eastern Asia. Furthermore, temperature has stronger effect on daytime SUHI in Africa, Europe and South America in summer, and precipitation has stronger effect on nighttime SUHI in Africa and Europe in summer. Our results emphasize the spatial variation of the relationships between SUHII and relevant driving factors across global major cities, further indicating that the spatially non-stationary effect of driving factors on SUHII need to be considered in the future.     

Exploring the Impact of Non-normality on Spatial Non-stationarity in Geographically Weighted Regression Analyses: Tobacco Outlet Density in New Jersey

The principal rationale for applying geographically weighted regression (GWR) techniques is to investigate the potential spatial non-stationarity of the relationship between the dependent and independent variables—i.e., that the same stimulus would provoke different responses in different locations. The calibration of GWR employs a geographically weighted local least squares regression approach. To obtain meaningful inference, it assumes that the regression residual follows a normal or asymptotically normal distribution. In many classical econometric analyses, the assumption of normality is often readily relaxed, although it has been observed that such relaxation might lead to unreliable inference of the estimated coefficients' statistical significance. No studies, however, have examined the behavior of residual non-normality and its consequences for the modeled relationships in GWR. This study attempts to address this issue for the first time by examining a set of tobacco-outlet-density and demographic variables (i.e., percent African American residents, percent Hispanic residents, and median household income) at the census tract level in New Jersey in a GWR analysis. The regression residual using the raw data is apparently non-normal. When GWR is estimated using the raw data, we find that there is no significant spatial variation of the coefficients between tobacco outlet density and percentage of African American and Hispanics. After transforming the dependent variable and making the residual asymptotically normal, all coefficients exhibit significant variation across space. This finding suggests that relaxation of the normality assumption could potentially conceal the spatial non-stationarity of the modeled relationships in GWR. The empirical evidence of the current study implies that researchers should verify the normality assumption prior to applying GWR techniques in analyses of spatial non-stationarity.     

Examining the influences of air quality in China's cities using multi‐scale geographically weighted regression

This study evaluates the influences of air pollution in China using a recently proposed model—multi‐scale geographically weighted regression (MGWR). First, we review previous research on the determinants of air quality. Then, we explain the MGWR model, together with two global models: ordinary least squares (OLS) and OLS containing a spatial lag variable (OLSL) and a commonly used local model: geographically weighted regression (GWR). To detect and account for any variation of the spatial autocorrelation of air pollution over space, we construct two extra local models which we call GWR with lagged dependent variable (GWRL) and MGWR with lagged dependent variable (MGWRL) by including the lagged form of the dependent variable in the GWR model and the MGWR model, respectively. The performances of these six models are comprehensively examined and the MGWR and MGWRL models outperform the two global models as well as the GWR and GWRL models. MGWRL is the most accurate model in terms of replicating the observed air quality index (AQI) values and removing residual dependency. The superiority of the MGWR framework over the GWR framework is demonstrated—GWR can only produce a single optimized bandwidth, while MGWR provides covariate‐specific optimized bandwidths which indicate the different spatial scales that different processes operate.     

Geographically weighted regression‐based determinants of malaria incidences in northern China

Geographically weighted regression (GWR) is an important local method to explore spatial non‐stationarity in data relationships. It has been repeatedly used to examine spatially varying relationships between epidemic diseases and predictors. Malaria, a serious parasitic disease around the world, shows spatial clustering in areas at risk. In this article, we used GWR to explore the local determinants of malaria incidences over a 7‐year period in northern China, a typical mid‐latitude, high‐risk malaria area. Normalized difference vegetation index (NDVI), land surface temperature (LST), temperature difference, elevation, water density index (WDI) and gross domestic product (GDP) were selected as predictors. Results showed that both positively and negatively local effects on malaria incidences appeared for all predictors except for WDI and GDP. The GWR model calibrations successfully depicted spatial variations in the effect sizes and levels of parameters, and also showed substantially improvements in terms of goodness of fits in contrast to the corresponding non‐spatial ordinary least squares (OLS) model fits. For example, the diagnostic information of the OLS fit for the 7‐year average case is R² = 0.243 and AICc = 837.99, while significant improvement has been made by the GWR calibration with R² = 0.800 and AICc = 618.54.     

GBGWR模型在降水量空间分布模拟中的应用

近年来,我国大部分地区屡遭洪涝与干旱两种自然灾害侵袭,对重洪涝干旱区域进行空间插值具有重要的意义。针对传统地理加权回归（GWR）模型建模过程中模型识别和参数估计易受观测值异常点影响的问题,本文提出了一种基于吉布斯采样的贝叶斯地理加权回归（GBGWR）方法。运用基于吉布斯采样的马尔可夫链蒙特卡罗贝叶斯方法,估计地理加权回归模型参数,通过平滑函数降低观测值中异常点位数据,最后对湖南省1985-2015年35个观测站点的降水观测数据进行了空间分布模拟。试验结果表明,本文提出的方法相较于GWR模型性能提高了19.8%,相较于BGWR模型性能提高了8.2%,该方法可以有效降低异常值和"弱数据"对回归结果的影响,能够更加真实地模拟湖南省降水量的空间分布。     

一种基于半监督学习的地理加权回归方法

地理加权回归方法在小样本数据下回归分析精度往往不高。半监督学习是一种利用未标记样本参与训练的机器学习方法,可以有效地提升少量有标记样本的学习性能。基于此本文提出了一种基于半监督学习的地理加权回归方法,其核心思想是利用有标记样本建立回归模型来训练未标记样本,再选择置信度高的结果扩充有标记样本,不断训练,以提高回归性能。本文采用模拟数据和真实数据进行试验,以均方误差提升百分比作为性能评价指标,将SSLGWR与GWR、COREG对比分析。模拟数据试验中,SSLGWR在3种不同配置下性能分别提升了39.66%、11.92%和0.94%。真实数据试验中,SSLGWR在3种不同配置下性能分别提升了8.94%、3.36%和5.87%。SSLGWR结果均显著优于GWR和COGWR。试验证明,半监督学习方法能利用未标记数据提升地理加权回归模型的性能,特别是在有标记样本数量较少时作用显著。     

The pyrogeography of sub-Saharan Africa: a study of the spatial non-stationarity of fire–environment relationships using GWR

This study analyses the relationship between fire incidence and some environmental factors, exploring the spatial non-stationarity of the phenomenon in sub-Saharan Africa. Geographically weighted regression (GWR) was used to study the above relationship. Environment covariates comprise land cover, anthropogenic and climatic variables. GWR was compared to ordinary least squares, and the hypothesis that GWR represents no improvement over the global model was tested. Local regression coefficients were mapped, interpreted and related with fire incidence. GWR revealed local patterns in parameter estimates and also reduced the spatial autocorrelation of model residuals. All the covariates were non-stationary and in terms of goodness of fit, the model replicates the data very well (R ² = 87%). Vegetation has the most significant relationship with fire incidence, with climate variables being more important than anthropogenic variables in explaining variability of the response. Some coefficient estimates exhibit locally different signs, which would have gone undetected by a global approach. This study provides an improved understanding of spatial fire–environment relationships and shows that GWR is a valuable complement to global spatial analysis methods. When studying fire regimes, effects of spatial non-stationarity need to be incorporated in vegetation-fire modules to have better estimates of burned areas and to improve continental estimates of biomass burning and atmospheric emissions derived from vegetation fires.     

Do Factors Causing Wildfires Vary in Space? Evidence from Geographically Weighted Regression

This paper describes the results of a geo-statistical analysis carried out at the provincial level in Southern Europe to model wildfire occurrence from socio-economic and demographic indicators together with land cover and agricultural statistics. We applied a classical ordinary least squares (OLS) linear regression together with a geographically weighted regression (GWR) to explain long-term wild-fire occurrence patterns (mean annual density of >1 ha fires). The explanatory power of the OLS model increased from 52% to 78% as a result of the non-constant relationships between fire occurrence and the underlying explanatory variables throughout the Mediterranean Basin. The global model we developed (i.e., OLS regression) was not sufficient to fully describe the underlying causal factors in wildfire occurrence modeling. Indeed, local approaches (i.e., GWR) can complement the global model in overcoming the problem of non-stationarity or missing variables. Our results confirm the importance of agrarian activities, land abandonment, and development processes as underlying factors of fire occurrence. The identification of regions with spatially varying relationships can contribute to the better understanding of the fire problem, especially over large geographic areas, while at the same time recognizing its local character. This can be very important for fire management and policy.     

顾及空间非平稳特征的遥感干旱监测

遥感技术具备实时快速、时空连续、广覆盖尺度等独特优势,在全球气候恶化大背景下,利用遥感干旱监测方法相比于传统地面监测手段,能够提供实时、准确、稳定的旱情信息,辅助科学决策。目前常用遥感旱情监测方法大多依赖全域性数学模型建模,假定了旱情模式的空间平稳特性,因而难以准确反映旱情模式的局部差异特征。本文提出利用地理加权回归模型GWR (Geographically Weighted Regression),考虑旱情模式的空间非平稳特性,综合多种遥感地面旱情监测指数,以实现传统全域旱情监测模型的局部优化。以美国大陆为研究区,监测2002年—2011年共10年的旱情状态。研究表明,GWR模型能够提供空间变化的局部最佳估计模型参数,监测结果更加吻合标准美国旱情监测USDM (U.S Drought Monitor)验证数据,且与地面实测值的最高相关系数R达到0.8552,均方根误差RMSE达到0.972,显著优于其他遥感旱情监测模型。GWR模型具备空间非平稳探测优势,实现了旱情模式的局部精细探测,能够显著提升遥感旱情监测精度,具备较好的应用前景。     

大坝变形变系数回归建模

在变形分析建模中常采用回归分析建模,但普通的回归模型是一种静态的模型,当变形体结构或物理性质发生变化时,普通线性回归所建立的静态模型将不再适用。变系数回归模型是一种动态模型,有着更强的灵活性和适应性,因此,将变系数回归引入大坝变形分析建模中,采用局部线性估计的方法进行系数拟合。仿真和大坝变形建模实验表明,变系数回归得到的大坝变形模型优于普通的线性回归模型,预测精度更高。     

时序变差函数特征驱动下的月平均降水空间分布模拟

高精度降水场是水文、气象以及环境分析的重要数据支撑,直接影响相关服务的准确性。传统降水分布模拟大多依赖站点空间维的驱动因素,而忽略了降水时序变化特征对其空间分布的影响。使用2015—2017年中国湖北省83个国家气象观测站点和28个省级观测站点近3 a月平均累积降水资料,通过相关性分析,引入站点降水时序理论变差函数模型的拱高值(C)和块金值(C₀)作为影响因素,使用地理加权回归(geographically weighted regression, GWR)建立湖北省月平均降水分布模型。结果表明:（1）各站点降水的时序变差函数曲线与降水的季节性基本吻合。站点时序理论变差函数模型中,有25.3%能够在4个月内达到平稳,36.14%在6个月内达到平稳。（2）站点降水时序理论变差函数模型的C和C₀与逐年12月平均累积降水在0.01水平（双侧）上显著相关,平均相关系数分别为0.745和0.526,大于地理位置和高程对降水的影响。（3）引入C和C₀ 有助于提升GWR模型的整体拟合优度和插值精度。对比仅使用经纬度的GWR模型和引入时序理论变差函数特征的GWR模型,3 a平均整体拟合优度从0.852提升至0.912。验证集站点插值精度评价显示,3 a绝对误差、均方根误差和平均绝对百分误差下降幅度均大于60%。因此,引入时序理论变差函数特征的时空GWR模型能够获得较高精度的降水模拟结果,更适合具有丰富历史降水资料地区的降水空间分布估算。     

Modeling Urban Growth Using GIS and Remote Sensing

Based on remote sensing and GIS, this study models the spatial variations of urban growth patterns with a logistic geographically weighted regression (GWR) technique. Through a case study of Springfield, Missouri, the research employs both global and local logistic regression to model the probability of urban land expansion against a set of spatial and socioeconomic variables. The logistic GWR model significantly improves the global logistic regression model in three ways: (1) the local model has higher PCP (percentage correctly predicted) than the global model; (2) the local model has a smaller residual than the global model; and (3) residuals of the local model have less spatial dependence. More importantly, the local estimates of parameters enable us to investigate spatial variations in the influences of driving factors on urban growth. Based on parameter estimates of logistic GWR and using the inverse distance weighted (IDW) interpolation method, we generate a set of parameter surfaces to reveal the spatial variations of urban land expansion. The geographically weighted local analysis correctly reveals that urban growth in Springfield, Missouri is more a result of infrastructure construction, and an urban sprawl trend is observed from 1992 to 2005.     

一种基于CUDA的大数据量地理加权回归并行加速算法

针对传统地理加权回归（GWR）在大数据量计算中存在的计算效率低、内存占用大、数据规模受限等问题，本文提出了快速并行地理加权回归（FPGWR）算法，基于英伟达CUDA架构实现了GWR的并行加速，将串行过程分解为并行的独立回归计算模块，同时优化了内存使用模型，提高了算法的运行速度。对比FPGWR和传统GWR在不同数量级模拟数据上和真实数据上的运行速度，结果显示，FPGWR能够支持更大规模的样本量计算并有效提升运行效率，数据量越大加速效果越显著。     

Partitioning Spatial Model Uncertainty Based on Joint Spatial Simulation

In this study, an uncertainty analysis procedure for joint sequential simulation of multiple attributes of spatially explicit models used in geographical informational systems was developed based on regression analysis. This procedure utilizes information obtained from joint sequential simulation to establish the relationship between model uncertainty and variation of model inputs. Using this procedure, model variance can be partitioned by model input parameters on a cell by cell basis. In the partitioning, the correlation of neighboring cells is accounted for. With traditional uncertainty analysis methods, this is not possible. In a case study, spatial variation of soil erodibility from a joint sequential simulation of soil properties was analyzed. The results showed that the regression approach is a very effective method in the analysis of the relationship between variation of the model output and model input parameters. It was also shown for the case study that: (1) the uncertainty of soil erodibility of a cell is mainly propagated from its own soil properties; (2) the interactions of soil properties of neighboring cells could reduce uncertainty of soil erodibility; (3) it is sufficient for uncertainty analysis to include the nearest three neighboring cell groups; and (4) the largest uncertainty contributors vary by soil properties and location.     

地理加权回归模型的多重共线性诊断方法

地理加权回归是常用的空间分析方法,已广泛应用于各个领域,但利用此方法进行回归分析前,往往忽略了对设计矩阵进行局部多重共线性的诊断,从而导致对模型的估计不准确。因此,本文在引入了全局模型的多重共线性诊断方法的基础上,对这些方法进行了改进,改进后构建了加权方差膨胀因子法和加权条件指标方法——分解比法,用于诊断地理加权回归模型设计矩阵的多重共线性问题。实验结果表明,多重共线性不存在于全局模型,而可能存在于局部模型中,构建的两种方法能够有效地诊断地理加权回归模型的多重共线性问题,且加权条件指标方法——分解比法比加权方差膨胀因子法在诊断多重共线性问题上更有优势。