Enhanced Urban Heat Island Modeling with Machine Learning and Regression Kriging in a Topographically Diverse Medium-Sized City

doi:10.21203/rs.3.rs-5009682/v1

Download PDF

Research Article

Enhanced Urban Heat Island Modeling with Machine Learning and Regression Kriging in a Topographically Diverse Medium-Sized City

https://doi.org/10.21203/rs.3.rs-5009682/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Urban Heat Islands (UHIs) pose significant environmental challenges, particularly in medium-sized cities with diverse topographies. This study enhances UHI modeling by integrating the Random Forest machine learning algorithm with regression kriging techniques. Focusing on Cluj-Napoca, we address the complexities of spatial temperature variability and improve UHI mapping accuracy. The Random Forest algorithm models air temperature variations, and regression kriging enhances spatial interpolation by integrating explanatory variables and residuals. The findings reveal that the Random Forest algorithm significantly improves temperature variability explanation and reduces prediction errors. Results highlight notable spatial trends, with high-temperature sensors concentrated in the central eastern part of Cluj-Napoca. Improved UHI modeling has substantial implications for urban planning and smart city initiatives. Accurate temperature mapping enables targeted mitigation strategies, such as green infrastructure, improved urban design, and strategic placement of cooling systems. These efforts enhance urban livability, reduce energy consumption, and mitigate heat wave effects. This study underscores the importance of integrating advanced machine learning techniques with traditional geostatistical methods to address complex environmental challenges. The methodology and findings are relevant across scientific disciplines, offering a framework for the future.

Urban Heat Island UHI

regression

machine learning

topography

Spatial Modeling

Environmental Monitoring

Rapid urbanization and climate change have brought the issue of Urban Heat Islands (UHIs) to the forefront. These phenomena manifest as higher temperatures in urban areas compared to their rural surroundings due to anthropogenic activities and landscape modifications (Arnfield 2003; Kim and Baik 2002; Oke 1982; Voogt and Oke 2003). UHIs not only increase thermal discomfort for residents but also contribute to higher energy consumption, intensified air pollution, and amplified climate change effects (Rizwan et al. 2008; Santamouris 2014; Santamouris 2015).

Medium-sized cities with diverse topographies present unique challenges in mapping and managing UHIs (Oke 1982). Complex topography can significantly influence the spatial distribution of temperatures, creating variations that are difficult to capture using traditional methods (Santamouris 2015). In this context, Cluj-Napoca offers an ideal case study due to its topographic and urban variability.

Combined modeling techniques using machine learning algorithms (ML) and geostatistical interpolation methods (Regression Kriging) provide more accurate estimates for the prediction of variables of interest in space and time. In many recent studies, combined methods are applied to monitor and evaluate different types of environmental variables, such as: physical-chemical characteristics and digital soil mapping (Pouladi et al. 2019; Santra et al. 2017; Tunçay et al. 2023; Veronesi and Schillaci 2019; Zhang et al. 2020; Zhu et al. 2022;), mapping of aboveground biomass of forests (Chen et al. 2019), estimation of geological characteristics (Erdogan Erten et al. 2022), water levels of the shallow groundwater system (Koch et al. 2019), land price estimation and mapping (Derdouri and Murayama 2020).

Current methods for UHI mapping often involve traditional geostatistical techniques such as SPLINE function, IDW, ordinary and universal kriging, etc., which interpolate spatial data based on the assumption of spatial autocorrelation. However, these methods can fall short in accurately capturing the complex interactions between urban morphology, topography, and temperature variations. Machine learning algorithms, such as Random Forest (RF), have been increasingly employed to enhance prediction accuracy by handling nonlinear relationships and interactions between variables. Recent studies have demonstrated the potential of combining machine learning techniques with geostatistical methods to improve the spatial accuracy of UHI models. The RF model was successfully used by Chen et al. in 2021 to predict air temperature in Nanjing, China, using independent variables such as land cover, impervious surface, surface albedo, and anthropogenic heat flux. Gardes et al. in 2020, developed both an RF model and a regression-based model to forecast nocturnal urban heat island intensity (UHII), using six predictors, and concluded that the RF model is superior to the regression-based model for predicting UHII. Ding et al. (2023) compared eight different air temperature prediction models using the network of existing weather stations in the city of Guangzhou, China. The best results for hourly air temperature data over a month were obtained using the regression kriging (RK) and multiple linear regression (MLR) methods. This study aims to improve UHI modeling by integrating the Random Forest machine learning algorithm with regression kriging techniques. By focusing on Cluj-Napoca, we address the complexities of spatial temperature variability and enhance UHI mapping accuracy.

The findings reveal that the Random Forest algorithm significantly improves the explanation of temperature variability and reduces prediction errors. The results highlight notable spatial trends, with high-temperature sensors concentrated in the central eastern part of Cluj-Napoca. Improved UHI modeling has substantial implications for urban planning and smart city initiatives. Accurate temperature mapping enables targeted mitigation strategies, such as green infrastructure, improved urban design, and strategic placement of cooling systems. These efforts enhance urban livability, reduce energy consumption, and mitigate heat wave effects.

Integrating advanced machine learning techniques with traditional geostatistical methods is essential for addressing complex environmental challenges. The methodology and findings of this study provide a framework for future research in urban climatology, environmental science, and smart city planning. By leveraging spatial analysis and machine learning, cities can develop more resilient and sustainable environments in the face of rapid urbanization and climate change.

2.1 Study area

The city of Cluj-Napoca is located in the Transylvania region, in northwestern Romania. The city is primarily placed on the terraces of the Someș Valley, emerging from the mountainous area. It is enclosed between the head of the "cuesta" and a hill that rises to about 900 meters, extending towards the valley in an east-west direction (Fig. 1).

The climate is temperate with moderate continental influences (Dfb climate type according to Koeppen). This results in distinct seasonal variations, with cold winters and warm summers. The local air dynamics are characterized by mountain breezes and the wind channeling phenomenon. The local air dynamics play a crucial role in shaping the urban climate. Mountain breezes, which occur due to the temperature differences between the mountain slopes and the valley, help in cooling the area during the summer nights. Additionally, the wind channeling phenomenon, where winds are funneled through narrow valleys or between buildings, influences the city's microclimate, contributing to variations in temperature and airflow patterns across different urban zones.

Cluj-Napoca is one of the largest cities in Romania, with a population of around 320,000 inhabitants. The city features a mix of residential, commercial, and industrial zones. Cluj-Napoca is experiencing rapid urban growth, with a significant increase in construction and infrastructure development. The city has parks and green areas, but their distribution is uneven.

The choice of Cluj-Napoca as a study area is based on several crucial factors. First, the city is experiencing rapid urban growth and significant urbanization, making this region a relevant site for studying the Urban Heat Island (UHI). Second, the unique topography, the diversity of economic activities, and urban characteristics such as residential, industrial, and commercial zones offer a unique opportunity to analyze thermal variations in various urban contexts. Finally, the availability of MICCRO (Holobâcă et al. 2022) network data and topographical surveys facilitates an in-depth analysis of the influence of geographic and climatic parameters on the formation of the UHI in Cluj-Napoca.

2.2 Regression methods

This study employs the regression kriging method, a geostatistical approach that combines regression techniques with spatial interpolation (Hengl et al. 2007). We integrate temperature data collected in the field with explanatory variables such as land use, building density, and other urban parameters. The regression kriging model accounts for the spatial structure of the data while incorporating regression relationships to estimate thermal variations. The deterministic part of regression kriging is based on the use of two types of regression: multiple linear regression (MLR) (Draper and Smith 1998) and Random Forest, a machine learning algorithm (Breiman 2001).

Multiple linear regression is a statistical method that models the relationship between a dependent variable and multiple independent variables by fitting a linear equation to the observed data. On the other hand, Random Forest is a robust machine learning technique that creates a multitude of decision trees during training and outputs the mode of the classes for classification or mean prediction for regression, enhancing model accuracy and reducing overfitting.

In the final stage, the stochastic part is added, which consists of the ordinary kriging of the residuals (Oliver and Webster 2014). This step ensures that the spatial autocorrelation of the residuals is appropriately modeled, further refining the accuracy of the temperature variation estimates.

3.1 MICCRO network

The design of the automatic observation network for the Urban Heat Island (UHI) started with the local climate zones (LCZ) defined by Stewart and Oke (2012). LCZs are urban areas with uniform surface cover, structure, materials, and human activity, extending horizontally from hundreds of meters to several kilometers. The delineation and use of LCZs offer two major advantages in the study of UHI: objective observation of UHI and comparability of results due to widespread international use.

Bechtel et al. (2015) highlighted that the progress of urban climate science is severely limited by the lack of high-resolution information describing the form and function of cities. As part of the World Urban Database and Access Portal Tools (WUDAPT) project, an international effort was launched to develop a global urban database and access portal to collect and disseminate this information coherently for urban areas worldwide (http://www.wudapt.org/). The portal provides a consistent methodology and access to the necessary software to map LCZs using supervised classification of satellite images. This methodology was used to identify LCZs in Cluj-Napoca (Fig. 2).

After determining the LCZs, it became possible to plan seasonal observations of the UHIs. The seasonal UHI maps were reclassified using quantiles to ensure the comparability of spatial data with unequal total variation intervals. The optimal placement of temperature sensors for the automatic monitoring system required several observation campaigns. The main concern was ensuring spatiotemporal representativeness by using LCZs to select observation points.

Observations were carried out during cloudless and windless nights (anticyclonic conditions) in winter, autumn and summer seasons. All manual observations have been made during the stable period (between 11 PM and 1 AM), using time correction. To limit the number of classes, three quantiles were used. The ArcGIS "combine" tool was employed to identify elementary thermal zones (with the same thermal behavior, specifically having the same combination of quantiles). The resulting polygons were simplified by eliminating areas that were too small (less than 0.1 km² in our case). The "critical points'' are the centroids of the elementary thermal zones (Holobâcă 2017). Finally, 40 USB HOBO DATALOGGER U23 PRO V2 TEMP/RH sensors were installed across an area of 179.5 km² (Fig. 3a), mounted on electrical poles at a height of 3 meters in Onset M-RSA – Solar Radiation Shield near the "critical points” (Fig. 3b).

3.2 Temperature data

The temperature data used in the regression analyses were obtained by averaging temperature values from the period with permanent MICCRO network observations (2020–2023). August was chosen for this study because it is the most stable summer month in our climate. Additionally, to detect the maximum influence of local conditions on air temperature, we used the average air temperature at 1:00 AM local time during the stability period.

For quality control, we filtered out values above two standard deviations from the mean to ensure the accuracy and reliability of the data. The input data were then verified using the Cluster and Outlier analysis available in ArcMap, which identifies hot spots, cold spots, and local spatial outliers using the Anselin Local Moran's I statistic (Anselin 1995). More precisely, each spatial entity (in this case, a point with a temperature attribute) is compared with the values of the neighboring points and a z-score value is calculated, which decides the placement of the respective point in one of the 4 categories: High-High cluster (the points with high temperatures are surrounded by points also with high temperatures); Low-Low (points with low temperatures are surrounded by points with low temperatures); High-Low outlier (a point with a high temperature is surrounded by other points with lower temperatures); Low-High (the point with low temperatures is surrounded by other points with higher temperatures). If the diversity of temperature values of a point is noted and it cannot be included in any of the 4 categories, the qualification "Not significant" is assigned. Points falling into either the cluster or outlier category indicate that the values are statistically significant at the 95% confidence level.

It is important that the analysis includes at least 30 spatial entities and that there is a high diversity of attribute values. In our case, 36 points on the territory of Cluj-Napoca were taken into account, each with a certain temperature value. As a way of conceptualizing the relationships between points, "Inverse distance" was selected, which assumes, based on Waldo Tobler's principle, that the closer points have a higher influence on a point than the more distant ones. There was no need to select a specific distance for analysis, considering the relatively small territory, but the tool automatically calculates the Euclidean distance that ensures that each entity has at least one neighbor. Considering the fact that a high number of permutations ensures a high precision of the results, the option of 999 permutations was selected. This means that the smallest possible pseudo-value is 0.001 and all other pseudo-p-values will be even multiples of this value. This verification process not only checked the spatial structure of the data but also identified any anomalous sensor readings in a spatial context.

3.3 Independent variable data

The independent variables analyzed in this study are related to location (altitude, X, Y coordinates, distance from the center, distances from the north-south and east-west axes) and the urban environment (building density, tree density, NDVI, and building height).

Altitude was derived from the EU-DEM dataset with a 30-meter resolution, created by combining SRTM and ASTER GDEM data. Building density, tree density, and building height were extracted from the Copernicus database (https://land.copernicus.eu/en) for the year 2018, with a spatial resolution of 10 meters. Tree density at 10 m resolution was obtained from Tree Cover Density data (https://land.copernicus.eu/en/products/high-resolution-layer-tree-cover-density), which provides information on proportional crown coverage per pixel, with values ranging from 0% (areas without trees) to 100% (areas completely covered by trees). Building density at 10 m resolution was retrieved from the Imperviousness Density raster (https://land.copernicus.eu/en/products/high-resolution-layer-imperviousness), which estimates the degree of imperviousness, ranging from 0% (completely permeable) to 100% (completely impervious). Building height at 10 m was extracted from the Urban Atlas Building Height data (https://sdi.eea.europa.eu/catalogue/idp/api/records/14c85db6-ac28-4a91-ad9c-e4356733976d), which contains information about building heights for 870 cities in Europe.

The NDVI index at 10 m resolution was calculated for the vegetation period from May to August 2018, using Sentinel-2 Level-2A satellite images via Google Earth Engine (GEE). The NDVI calculation utilized GEE's built-in normalized difference function, selecting bands B4 (RED) and B8 (NIR) as parameters. Only images with a cloud cover percentage of less than 10% were included in the analysis. To ensure consistent aggregation, all obtained images were resized to a 30-meter resolution (the coarser data resolution) in ArcGIS using the resampling tool.

3.4 The regression algorithm

Our analysis followed several key steps to ensure accurate and reliable results integrated in python scripts (Fig. 4):

Preparation of Environmental Raster Data: We ensured that the environmental raster data had the same extension as the point data (temperature) and the same resolution for accurate overlay.

Data Extraction and Tabulation: We extracted the variables and tabulated the data to prepare it for regression analysis.

Building Regression Models: We constructed the regression models using Bayesian optimization for multiple linear regression (MLR) and hyperparameter tuning for random forest regression (RFR).

Spatializing the Model: We applied the results of the regression models to spatialize the predictions across all pixels in the study area, ensuring comprehensive temperature mapping.

4.1 Anselin Local Moran's I statistic

Following the Cluster and Outlier analysis, 16 out of the 36 points were identified as statistically significant and were classified into one of four categories (Fig. 5). Specifically, 4 points in the eastern area were grouped into the "High-High Clusters'' category. The "Low-Low Clusters'' category included 7 points located in the western area. The "High-Low'' outliers consist of only two points located in the west. In contrast, the "Low-High'' outliers are composed of three points positioned in the eastern part. The remaining 20 points were classified as "Not Significant" because of the diversity of temperature values which cannot integrate the points into clusters or outliers. Spatially, this classification into the four categories reveals a clustering of higher temperature values in the eastern part of the study area, as well as the concentration of points with lower temperature values in the western part.

4.2 MLR results

In our study, we employed the "forward" method in multiple linear regression, an iterative approach which allows us to progressively select the most significant independent variables for inclusion in the regression model. Furthermore, the best model was selected using the Bayesian Information Criterion (BIC), which penalizes models with a larger number of variables to favor more parsimonious models and minimize the risk of overfitting.

The final model includes four variables. The positional variables selected were longitude (X), distance to the north-south axis and the distance to the east-west axis. Only one variable characterizing the urban environment, namely NDVI, has been retained. Based on the standardized beta coefficient (Table 1), which indicates the variable's importance in the regression, we identified the longitude (X) as the most crucial parameter followed by NDVI. In the last step, the MLR beta coefficients have been used to generate a raster that contains the predicted temperature values (Fig. 6).

Table 1

MLR Beta standardized parameters
Variable	Beta std
X	0.3481
NDVI	-0.3370
Distances axis nord south	-0.3315
Distances axis east west	-0.2398

The spatial distribution of air temperature values predicted by the Multiple Linear Regression (MLR) model (Fig. 6a) reflects the retained variables for regression. A clear contrast is observed between the eastern and western parts of the city, with longitude being the most important variable for the regression. The highest temperature values are also found in areas with reduced vegetation and near the east-west axis of the city. The most significant issue is the underestimation of temperature values in the hotspot located in the central-eastern part of the city. This underestimation is clearly highlighted by the kriging of residuals (Fig. 6b), with positive maxima situated precisely in this part of the city. Conversely, the cold spot in the northwestern part of the city is correctly identified, with temperature values being 3 degrees Celsius lower than those in the central area.

Figure 6a. The spatial distribution of air temperature values predicted by the Multiple Linear Regression (MLR); b. Ordinary Kriging of Residuals

4.3 RFR results

In this study, Random Forest regression was employed to model complex relationships between various spatial predictors and a target variable. The hyperparameter tuning has been used by GridSearchCV to optimize the Random Forest model by varying parameters such as maximum depth, number of estimators, and maximum features sampled per tree. This process aimed to enhance model performance, measured primarily through the coefficient of determination (R²), which quantifies how well the model fits the data.

Following training with the optimal parameters, the model's performance was assessed using R², mean squared error (MSE), and root mean squared error (RMSE). These metrics provided insights into the model's ability to explain variation and predict target values accurately across the dataset.

Moreover, spatial predictors stored as raster files were integrated using the pyspatialml library. The Random Forest model was applied to predict the target variable across the study area, generating a raster output depicting predicted values.

The Random Forest Regression (RFR) variables' importance is calculated based on the feature importance scores provided by the trained Random Forest model. Random Forest calculates feature importance by measuring how much each feature (predictor variable) contributes to decreasing the impurity or "node purity" when making decisions in each tree of the forest. Specifically, the importance of a feature is computed as the average (across all trees) of the decrease in node impurity (Gini impurity or entropy) weighted by the number of samples processed by each tree node that splits on that feature. Feature importance values have been standardized to be scaled between 0 and 1 for an easier interpretation. Standardization helps in comparing the relative importance of different features within the same model (Table 2).

Table 2

RFR variables importance
Variable	Importance
X	0.219
Distance axis nord south	0.147
Distance to center	0.119
Altitude	0.111
NDVI	0.105
Distance axis east west	0.096
Building density	0.074
Y	0.058
Building	0.057
Tree density	0.009

The longitude (X) has the highest importance score of 0.219, indicating it has the most significant influence on the model's predictions. With importance scores ranging from 0.111 to 0.147, Distance axis north-south, Distance to center and Altitude have a significant role in regression. A medium role is played by NDVI and Distance axis east-west with importance score near to 1. Building density, latitude (Y) and Tree density have importance scores ranging from 0.074 to 0.009. They contribute to the model predictions but have relatively lower importance compared to the top predictors listed above.

Compared to the results of the Multiple Linear Regression (MLR) model, the machine learning algorithm (Random Forest Regression, RFR) more accurately predicts both the hotspot in the central-eastern part of the city and the cold spot in the northwestern part (Fig. 7a). The RFR model also provides a more logical spatial distribution of temperature values, with temperatures decreasing towards the outskirts of the city, including the eastern limits. The spatial distribution of residuals (Fig. 7b) also shows an overestimation in the central part of the city, but both the magnitude and spatial extent of this overestimation are much smaller than in the MLR model.

4.4 RK results

The final step involved combining the deterministic output from Multiple linear and Random Forest regression with the stochastic component from Kriging (Fig. 8). This integration provided a comprehensive prediction map, highlighting areas of interest based on both modeled relationships and spatial autocorrelation of residuals.

The analysis of regression results highlights a notable spatial temperature disparity across the city. Both regression models indicate that the eastern part of the city consistently exhibits higher temperatures compared to the western part. This thermal pattern is evident not only in the overall analysis but also in the monthly averaged temperature values, which show a clear thermal inversion. Specifically, the higher altitudes in the southern part of the city situated at 250 m higher altitude tend to have 2 degree Celsius higher air temperatures than the lower elevations near the Someș Valley in the west.

A closer examination of the spatial distribution of temperature values reveals that the Random Forest Regression (RFR) model outperforms in identifying and predicting hot spots. The RFR model effectively captures a prominent hot spot in the central-eastern part of the city, aligning well with the findings from the Anselin Moran analysis. This consistency between the RFR predictions and the Anselin Moran analysis underscores the robustness of the RFR model in identifying significant temperature variations and spatial clusters.

In summary, the regression analysis and spatial data examination reveal a clear thermal pattern with the eastern city region being warmer. The RFR model, in particular, provides superior performance in pinpointing central-eastern hot spots, corroborating spatial statistical analyses and enhancing our understanding of urban temperature dynamics.

The robust performance of the Random Forest Regression model, as evidenced by the higher R² and lower error metrics (MSE and RMSE) (Table 3), demonstrates its superior ability to predict spatial temperature variations. This makes it a valuable tool for urban temperature analysis and helps in identifying critical hot spots that are consistent with spatial statistical analysis. The improved performance of the RFR model over the MLR model not only enhances the accuracy of temperature predictions but also provides deeper insights into the spatial distribution of temperature anomalies within the city.

Table 3

MLR vs. RFR
Model	R²	MSE	RMSE
Multiple Linear Regression	0.653	0.129	0.359
Random Forest Regression	0.844	0.058	0.241

In this study, we effectively utilized regression kriging to map nighttime temperatures during August in Cluj-Napoca, Romania. This technique, which combines spatial and meteorological data, provided a detailed assessment of local thermal variations, effectively highlighting the urban heat island effect.

In Cluj-Napoca, there is a deviation from the classic urban heat island (UHI) structure, where the city's hottest area is typically observed in the downtown area and decreases towards the outskirts, with local variations due to different physical properties of the urban surface. This spatial pattern is most commonly observed in newer cities in the United States.

In our city/case study, we can observe the significant role of local topographic conditions in both the horizontal and vertical thermal structure. Horizontally, a notable thermal contrast can be seen between the western and eastern parts of the city, with the heat island core extending eastward. The difference between the northwestern outskirts of the city and the hottest area in the central-eastern part exceeds three degrees Celsius in average monthly values (Fig. 3). This situation appears to be caused by the penetration of the nocturnal mountain breeze along the Someș Valley from the west, pushing warmer air towards the eastern part of the city. Similar cases have been reported by Stewart and Oke (2012) during significant nocturnal advection.

While the horizontal thermal structure of the heat island was also observed during previous observation campaigns, we can now identify the characteristics of vertical stratification. A second warm core is visible in the southern part of the city. This core corresponds to neighborhoods located on the southern slope of the Feleac massif. These neighborhoods are situated at an altitude approximately 250 meters higher than the central area near the Someș River. The difference of about two degrees Celsius suggests the presence of a nocturnal thermal inversion. In our opinion, this combination of nocturnal thermal brise and channeling through the river valley and urban canyons shape this very particular thermal structure in Cluj-Napoca, Romania.

Our analysis demonstrated that the random forest machine learning model outperformed traditional multiple linear regression in both explaining temperature variability and reducing prediction error. This indicates the model's robustness in handling complex spatial-temporal data.

The findings revealed notable spatial patterns, with higher temperatures concentrated in the central eastern areas of Cluj-Napoca. This insight is crucial for urban planning and public health, as it identifies regions that may require targeted cooling strategies to mitigate the impacts of urban heat.

This study successfully employed an advanced regression kriging approach to enhance the modeling of nighttime temperatures in Cluj-Napoca, Romania, during August. By integrating the Random Forest machine learning algorithm with traditional geostatistical methods, we provided a more nuanced understanding of local thermal variations and the urban heat island (UHI) effect.

The findings from this study reveal a significant departure from the classic UHI structure typically observed in newer cities, particularly those in the United States. In Cluj-Napoca, the hottest areas are not concentrated in the downtown core but are influenced significantly by the city's topographic conditions. This results in a distinct horizontal thermal contrast between the western and eastern parts of the city, with a prominent heat island core extending eastward. Similar cases have been reported by Stewart and Oke (2012) during significant nocturnal advection.

The influence of local topography is further evident in the vertical thermal structure of the city. A secondary warm core was identified in the southern neighborhoods, located on the southern slope of the Feleac massif at a higher altitude than the central areas near the Someș River. This finding suggests the presence of a nocturnal thermal inversion, driven by the combined effects of nocturnal mountain breezes and the channeling of airflow through the river valley and urban canyons. This complex interaction results in significant thermal stratification, highlighting the unique thermal dynamics of Cluj-Napoca.

The analysis demonstrated that the Random Forest machine learning model significantly outperformed traditional multiple linear regression in both explaining temperature variability and reducing prediction error. This underscores the robustness of machine learning techniques in handling complex spatial-temporal data, providing a more accurate and detailed temperature mapping.

The study's results have substantial implications for urban planning and public health in Cluj-Napoca. The identification of higher temperature concentrations in the central eastern areas emphasizes the need for targeted cooling strategies, such as the implementation of green infrastructure, improved urban design, and the strategic placement of cooling systems. These measures are essential for enhancing urban livability, reducing energy consumption, and mitigating the adverse effects of heat waves.

Moreover, the integration of advanced machine learning techniques with traditional geostatistical methods presents a powerful framework for future research in urban climatology and environmental science. This approach is not only applicable to Cluj-Napoca but can also be adapted to other medium-sized cities with diverse topographies, offering a comprehensive tool for urban heat island mitigation and sustainable urban development.

In conclusion, the study highlights the importance of leveraging modern data analysis techniques to address complex environmental challenges. By providing a detailed and accurate model of urban heat distribution, this research contributes to the development of more resilient and sustainable urban environments in the face of rapid urbanization and climate change.

Conflict of Interest statement

The authors declare that there are no conflicts of interest regarding the publication of this article.

Authors’ contributions I.H.H. -

Conceived and designed the study, conducted experiments, collected and analyzed data, and wrote the manuscript; M.A. - collected data, writing - review and editing; K.T.-I - collected data, data visualisation, writing - review and editing C.-D.U. - contributed to the methodology, data visualisation, writing - review and editing.

Funding statement:

Internal grants for supporting the strategic infrastructure of UBB 2019: Implementation of an Urban Heat Island measurement network in Cluj-Napoca.

Author Contribution

"I.H.H. - Conceived and designed the study, conducted experiments, collected and analyzed data, and wrote the manuscriptM.A. - collected data, writing - review and editingK.T.-I - collected data, data visualisation, writing - review and editing C.-D.U. - contributed to the methodology, data visualisation, writing - review and editingAll authors reviewed the manuscript"

Acknowledgement

"We would like to extend our heartfelt gratitude to all the students who participated in the field observations. Your dedication and hard work have been invaluable to this research. We are especially grateful to the volunteers from the Applied Climatology Study Group. Your expertise, enthusiasm, and commitment have significantly contributed to the success of this project. Without your tireless efforts, this study would not have been possible. Thank you all for your exceptional contributions."

Anselin L (1995) Local Indicators of Spatial Association—LISA”. Geographical Analysis 27(2):93-115. https://doi.org/10.1111/j.1538-4632.1995.tb00338.x
Arnfield AJ (2003) Two decades of urban climate research: A review of turbulence, exchanges of energy and water, and the urban heat island. International Journal of Climatology 23(1):1-26
Bechtel B, Alexander PJ, Böhner J, Ching J, Conrad O, Feddema J et al (2015) Mapping local climate zones for a worldwide database of the form and function of cities. ISPRS International Journal of Geo-Information 4(1):199-219
Breiman L (2001) Random forests. Machine Learning 45(1):5-32. https://doi.org/10.1023/A:1010933404324
Chen L, Wang Y, Ren C, Zhang B, Wang Z (2019) Assessment of multi-wavelength SAR and multispectral instrument data for forest aboveground biomass mapping using random forest kriging. Forest Ecology and Management 447:12-25. https://doi.org/10.1016/j.foreco.2019.05.057
Chen S, Yang Y, Deng F, Zhang Y, Liu D, Liu C, Gao Z (2021) A High-Resolution Monitoring Approach of Canopy Urban Heat Island using Random Forest Model and Multi-platform Observations. Atmospheric Measurement Techniques. https://doi.org/10.5194/amt-2021-301
Derdouri A, Murayama YA (2020) Comparative study of land price estimation and mapping using regression kriging and machine learning algorithms across Fukushima prefecture, Japan. J. Geogr. Sci. 30:794–822. https://doi.org/10.1007/s11442-020-1756-1
Ding X, Zhao Y, Fan Y, Li Y, Ge J (2023) Machine learning-assisted mapping of city-scale air temperature: Using sparse meteorological data for urban climate modeling and adaptation. Building and Environment, 234:110211
Draper NR, Smith H (1998) Applied Regression Analysis (3rd ed.). John Wiley & Sons
Erdogan Erten G, Yavuz M, Deutsch CV (2022) Combination of Machine Learning and Kriging for Spatial Estimation of Geological Attributes. Nat Resour Res. 31:191–213. https://doi.org/10.1007/s11053-021-10003-w
Gardes T, Schoetter R, Hidalgo J, Long N, Marquès E, Masson V (2020) Statistical prediction of the nocturnal urban heat island intensity based on urban morphology and geographical factors - An investigation based on numerical model results for a large ensemble of French cities. The Science of the total environment 737:139253. https://doi.org/10.1016/j.scitotenv.2020.139253
Hengl T, Heuvelink GBM, Rossiter DG (2007) About regression-kriging: From equations to case studies. Computers & Geosciences 33(10):1301-1315. https://doi.org/10.1016/j.cageo.2007.05.001
Holobâcă IH (2017) Le monitoring de l'ilot de chaleur urbain de Cluj-Napoca, Roumanie, Actes du XXX^e Colloque de l’Association Internationale de Climatologie, Climat, ville et environnement, Sfax, Tunisie
Holobâcâ IH, Alexe M, Temerdek-Ivan K (2022) Les premiers résultats de la surveillance de l’îlot de chaleur à Cluj-Napoca à l’aide du réseau automatique MICCRO (Monitorizarea Insulei de Căldura în Cluj - Romania). 35^eme Colloque Annuel de l’AIC, Toulouse, France
Kim Y, Baik JJ (2002) Maximum urban heat island intensity in Seoul. Journal of Applied Meteorology 41(6):651-659. https://doi.org/10.1175/1520-0450(2002)041<0651:MUHIII>2.0.CO;2
Koch J, Berger H, Henriksen HJ, Sonnenborg TO (2019) Modelling of the shallow water table at high spatial resolution using random forests. Hydrol. Earth Syst. Sci. 23:4603-4619. https://doi.org/10.5194/hess-23-4603-2019
Oke TR (1982) The energetic basis of the urban heat island. Quarterly Journal of the Royal Meteorological Society 108(455):1-24
Oliver MA, Webster R (2014) A tutorial guide to geostatistics: Computing and modelling variograms and kriging. Catena 113:56-69. https://doi.org/10.1016/j.catena.2013.09.006
Pouladi N, Møller AB, Tabatabai S, Greve MH (2019) Mapping soil organic matter contents at field level with Cubist, Random Forest and kriging. Geoderma 342:85-92. https://doi.org/10.1016/j.geoderma.2019.02.019
Rizwan AM, Dennis LYC, Liu C (2008) A review on the generation, determination and mitigation of Urban Heat Island. Journal of Environmental Sciences 20(1):120-128
Santamouris M (2014) Cooling the cities – A review of reflective and green roof mitigation technologies to fight heat island and improve comfort in urban environments. Solar Energy 103:682-703
Santamouris M (2015) An analysis of the heat island magnitude and characteristics in one hundred Asian and Australian cities and regions. Science of The Total Environment 512-513:582-598
Santra P, Kumar M, Panwar N (2017) Digital soil mapping of sand content in arid western India through geostatistical approaches. Geoderma Regional 9:56-72. https://doi.org/10.1016/j.geodrs.2017.03.003
Stewart ID, Oke TR (2012) Local climate zones for urban temperature studies. Bulletin of the American Meteorological Society 93(12):1879-1900
Tunçay T, Alaboz P, Dengiz O, Başkan O (2023) Application of regression kriging and machine learning methods to estimate soil moisture constants in a semi-arid terrestrial area. Computers and Electronics in Agriculture 212. https://doi.org/10.1016/j.compag.2023.108118
Veronesi F, Schillaci C (2019) Comparison between geostatistical and machine learning models as predictors of topsoil organic carbon with a focus on local uncertainty estimation. Ecological Indicators 101:1032-1044. https://doi.org/10.1016/j.ecolind.2019.02.026
Voogt JA, Oke TR (2003) Thermal remote sensing of urban climates. Remote Sensing of Environment 86(3):370-384
WUDAPT. (2022) World Urban Database and Access Portal Tools. Retrieved from http://www.wudapt.org/
Zhang Y, Ji W, Saurette DD, Easher TH, Li H, Shi Z. et al (2020) Three-dimensional digital soil mapping of multiple soil properties at a field-scale using regression kriging. Geoderma 366. https://doi.org/10.1016/j.geoderma.2020.114253
Zhu C, Wei Y, Zhu F, Lu W, Fang Z, Li Z, Pan J (2022) Digital Mapping of Soil Organic Carbon Based on Machine Learning and Regression Kriging. Sensors 22:8997. https://doi.org/10.3390/s22228997

No competing interests reported.

Download PDF

Reviewers agreed at journal
09 Nov, 2024
Reviewers invited by journal
03 Oct, 2024
Editor assigned by journal
01 Sep, 2024
Submission checks completed at journal
01 Sep, 2024
First submitted to journal
31 Aug, 2024

You are reading this latest preprint version

Enhanced Urban Heat Island Modeling with Machine Learning and Regression Kriging in a Topographically Diverse Medium-Sized City

Status:

Version 1

Abstract

Figures

1 Introduction

2 Methods

2.1 Study area

2.2 Regression methods

3 Data

3.1 MICCRO network

3.2 Temperature data

3.3 Independent variable data

3.4 The regression algorithm

4 Results

4.1 Anselin Local Moran's I statistic

4.2 MLR results

4.3 RFR results

4.4 RK results

5 Discussion and Conclusion

Declarations

Conflict of Interest statement

Authors’ contributions I.H.H. -

Funding statement:

Author Contribution

Acknowledgement

References

Additional Declarations

Status:

Version 1