Please login first
Application of geostatistical modelling to study the relationships between the surface urban heat island effect and land-cover using Landsat time series data
* ,
1  Warsaw University of Technology

Published: 22 March 2018 by MDPI in 2nd International Electronic Conference on Remote Sensing session Posters

Development of remote sensing techniques has made a significant contribution to assess climatology phenomena and determine which predictors have a noticeable influence on the intensity of the surface urban heat island (SUHI) effects. The aim of this study is to analyse the effectiveness of the geostatistical modelling of thermal properties of land surface in an expanding city, Poznan in west Poland. The applied models – Ordinary Least Squares (OLS) and Geographically Weighted Regression (GWR) were used to explore the strength of the correlation between explanatory variables (e.g. porosity index, ISA, road density) and dependent variables defined as mean SUHI intensity (MSUHIiintensity ) and difference mean SUHI intensity between 2001 and 2011 ( ΔMSUHIiintensity) for each district within the city . In the research we employed two Landsat images (2001 and 2011) on the basis of which SUHI intensity and land-cover maps were generated. Classification results ( overall accuracy of 97,5% ) obtained through Artificial Neural Network (ANN) algorithm were used as explanatory variables to identify the impact of land-cover and its derived products on the SUHI effect within the city. On the grounds of the combination of the chosen predictors it was possible to examine whether land-cover types and their derivatives determined mean SUHI values (MSUHIiintensity) and if changes in land cover effected ΔMSUHIiintensity. On the basis of statistical indicators ( I-Moran index, AICc, VIF, R2) it turned out that the most suitable predictors were ISA, ΔISA and road density. The results for the cross-sectional GWR model (R2 = 0,730) were better than for the OLS (R2 = 0,470). In contrast to the cross-sectional analyses, the goodness of fit for the longitundinal OLS model (R2 = 0,501) was similar to the GWR results (R2 = 0,500). However, the GWR revealed that local regression residuals were differentiated – values for some regions in the city centre were overestimated and for the outskirts of Poznan R2 underestimation values were noted . This situation indicated that unlike other cities for which the longitundinal GWR modelling gave better results, for Poznan the GWR did not improve the modelling effectiveness (ZHOU, WANG, 2011; DEILAMI, KAMRUZZAMAN, 2017). This means that associations between dependent and explanatory variables are stationary and as a result ΔMSUHIiintensity is not spatially variable.

This study has identified associations between SUHI effects and remotely-sensed land-cover parameters in Poznan. Results demonstrated that the GWR methods have proved effective in modelling using cross-sectional analysis (R2 = 0,730). In the case of estimating thermal conditions variability between 2001 and 2011 applying the GWR did not improve the modelling results (R2 = 0,500), what could be explained by the different spatial structure of the city and a moderate climate with both maritime and continental elements.

Keywords: Surface urban heat island, geographically weighted regression, Landsat, multivariate regression, cross-sectional analysis, climate change