Quantifying the effect of surrounding spatial heterogeneity on land surface temperature based on local climate zones using mutual information

Local climate zone (LCZ), a landscape classification scheme that segments urban region into 17 distinct zones, has become the standard for analyzing urban thermal environments since 2012. The characteristic features of each LCZ decide the standard regime for their Land Surface Temperature (LST). In heterogeneous cities, the variation in LST is observed for similar but spatially dispersed LCZs and has not yet been analysed by the researchers. The surrounding spatial heterogeneity could be partially responsible for such variation. Hence, we have presented a framework to analyze the LST variation of similar but spatially dispersed LCZs by considering the surrounding LCZ pattern1details in the study area of Nagpur (India). The framework uses machine learning techniques like random forest (RF), Xgboost, deep learning to confirm that the LST variation is due to surrounding LCZ pattern. Later it indentifies the factors responsible for LST variation in each LCZ type using our proposed mutual information based Adjacent LCZ preference Estimator (ALPE). Deep learning model captures 90-92% of the neighbourhood heterogeneity responsible for LST variation as compared to RF (0.83- 0.79) and Xgboost (0.86-0.81). The reduction in bias (by 1.05°C -1.39°C) is observed while estimating LST incorporating surrounding LCZ pattern. This confirms that the external heterogeneity affects the LST of the corresponding LCZ. Further, the result analysis of ALPE suggests that characteristics of open LCZ types (when they are present in surrounding) highly influence the LST of compact and open LCZs.