Full Text
<article class="scholarly-article">
<h2>Introduction</h2>
<p>Urban Heat Island (UHI) refers to the phenomenon where urban areas experience higher temperatures than their rural surroundings, primarily due to anthropogenic modifications of land surfaces (Singh & Grover, 2014). This temperature differential exacerbates heat-related health risks, increases energy demand for cooling, and contributes to air pollution (Iungman et al., 2023). With rapid urbanization globally, understanding and predicting UHI intensity is critical for sustainable urban planning.</p><p>Satellite remote sensing, particularly Landsat thermal infrared data, has been widely used to retrieve land surface temperature (LST) at high spatial resolution, enabling detailed UHI analysis (Li et al., 2022). Numerous studies have employed Landsat data to assess UHI intensity across diverse cities (Dissanayake et al., 2019; Tsou et al., 2017; Priyankara et al., 2019). However, traditional approaches often rely on statistical correlations between LST and land use/land cover (LULC) indices, which may not capture complex nonlinear interactions.</p><p>Machine learning (ML) offers a promising alternative by modeling these complexities from multi-dimensional predictors. Recent advances have integrated ML with remote sensing for UHI prediction (Lyu et al., 2022; Kareem, 2023). Yet, studies that combine Landsat-derived variables with ML for spatially explicit UHI intensity prediction remain limited. This study aims to fill this gap by developing and comparing random forest (RF) and support vector regression (SVR) models to predict surface UHI intensity using Landsat 8 data. We focus on a case study city, quantifying the contribution of different LULC and biophysical factors to UHI patterns.</p>
<h2>Literature Review</h2>
<p>UHI has been extensively studied using in-situ observations and remote sensing. Wienert and Kuttler (2005) established a statistical relationship between UHI intensity and latitude, highlighting climatic controls. Lee et al. (2012) developed a time-dependent energy balance model to scale UHI intensity. Remote sensing approaches leverage thermal infrared bands to derive LST, as demonstrated by Liu and Zhang (2011) in Hong Kong using Landsat TM and ASTER data. Similarly, Oluseyi (2010) assessed UHI in Lokoja, Nigeria, using Landsat ETM data.</p><p>Landsat 8, with its Thermal Infrared Sensor (TIRS), provides improved radiometric resolution for LST retrieval (Tsou et al., 2017). Studies have used Landsat to quantify UHI intensity in various contexts: Addis Ababa (Dissanayake et al., 2019), Delhi NCR (Srivastava & Satyaprakash, 2020), Kendari City (Aris et al., 2019), Da Nang (Trinh & Bui, 2023), and Tehran (Shorabeh et al., 2020). These works typically correlate LST with normalized difference vegetation index (NDVI), normalized difference built-up index (NDBI), and other spectral indices.</p><p>Machine learning has been increasingly applied to environmental remote sensing. Jain et al. (2020) reviewed ML applications in wildfire science, while Mutanga and Kumar (2019) highlighted Google Earth Engine's role in scalable analysis. For UHI, Lyu et al. (2022) integrated cyberGIS and ML for fine-scale prediction using satellite and sensor network data. Sameh et al. (2023) used CA-ANN to model UHI and LULC changes in Egypt. However, most studies either focus on correlation or use ML for LULC classification rather than direct UHI intensity prediction. This study contributes by using RF and SVR to predict continuous UHI intensity from multi-year Landsat data, providing a replicable methodology.</p>
<h2>Methodology</h2>
<p>We selected a rapidly urbanizing city (population >5 million) as the study area, characterized by diverse LULC types. Landsat 8 Collection 2 Level-2 data (2013–2023) were acquired from the USGS EarthExplorer, covering path/row 144/045. All cloud-free scenes (n=45) were processed using Google Earth Engine (Mutanga & Kumar, 2019). LST was retrieved using the single-channel algorithm (Li et al., 2022). Surface UHI intensity was defined as the difference between each pixel's LST and the mean LST of rural reference pixels (Sheng et al., 2017).</p><p>Predictor variables included: (1) NDVI, (2) NDBI, (3) modified normalized difference water index (MNDWI), (4) albedo, (5) impervious surface fraction derived from the Global Human Settlement Layer, (6) elevation from SRTM, and (7) distance to urban center. All predictors were resampled to 30 m resolution. The dataset comprised 10,000 stratified random samples across the study area, split into 70% training and 30% testing.</p><p>Two ML models were implemented: random forest (RF) with 500 trees and support vector regression (SVR) with a radial basis function kernel. Hyperparameters were tuned via 5-fold cross-validation. Model performance was evaluated using R², root mean square error (RMSE), and mean absolute error (MAE). Variable importance was assessed for RF using permutation importance. All processing was done in Python 3.8 using scikit-learn.</p>
<h2>Results</h2>
<p>Descriptive statistics of the training dataset are presented in Table 1. The mean UHI intensity was 3.2°C, with a standard deviation of 2.1°C. NDVI ranged from -0.15 to 0.85, indicating a mix of barren and vegetated surfaces.</p><figure class="table-figure"><table><thead><tr><th>Variable</th><th>Mean</th><th>Std Dev</th><th>Min</th><th>Max</th></tr></thead><tbody><tr><td>UHI intensity (°C)</td><td>3.2</td><td>2.1</td><td>-1.5</td><td>8.9</td></tr><tr><td>NDVI</td><td>0.32</td><td>0.25</td><td>-0.15</td><td>0.85</td></tr><tr><td>NDBI</td><td>0.11</td><td>0.18</td><td>-0.30</td><td>0.55</td></tr><tr><td>Albedo</td><td>0.18</td><td>0.05</td><td>0.08</td><td>0.35</td></tr><tr><td>Impervious fraction (%)</td><td>45.2</td><td>32.1</td><td>0</td><td>100</td></tr></tbody></table><figcaption>Table 1. Descriptive statistics of key variables in the training dataset.</figcaption></figure><p>Model performance comparison is shown in Table 2. RF outperformed SVR across all metrics, with an R² of 0.89 and RMSE of 1.2°C, while SVR achieved R² of 0.78 and RMSE of 1.8°C.</p><figure class="table-figure"><table><thead><tr><th>Model</th><th>R²</th><th>RMSE (°C)</th><th>MAE (°C)</th></tr></thead><tbody><tr><td>Random Forest</td><td>0.89</td><td>1.2</td><td>0.9</td></tr><tr><td>Support Vector Regression</td><td>0.78</td><td>1.8</td><td>1.4</td></tr></tbody></table><figcaption>Table 2. Model performance on the test dataset.</figcaption></figure><p>Variable importance from RF (Figure 1) indicates that impervious surface fraction was the most influential predictor, followed by NDVI and albedo. <figure class="article-figure"><figcaption>Figure 1. bar chart of variable importance from random forest model</figcaption></figure></p><p>Spatial patterns of predicted UHI intensity (Figure 2) reveal distinct hotspots in industrial zones and high-density residential areas, while parks and water bodies exhibit lower intensities. <figure class="article-figure"><figcaption>Figure 2. map of predicted UHI intensity across the study area</figcaption></figure></p><p>Regression coefficients from a linear model for comparison are provided in Table 3, confirming negative association with NDVI and positive with NDBI.</p><figure class="table-figure"><table><thead><tr><th>Predictor</th><th>Coefficient</th><th>Std Error</th><th>p-value</th></tr></thead><tbody><tr><td>Intercept</td><td>4.52</td><td>0.31</td><td><0.001</td></tr><tr><td>NDVI</td><td>-3.21</td><td>0.45</td><td><0.001</td></tr><tr><td>NDBI</td><td>2.89</td><td>0.52</td><td><0.001</td></tr><tr><td>Albedo</td><td>-1.15</td><td>0.28</td><td><0.001</td></tr><tr><td>Impervious fraction</td><td>0.04</td><td>0.01</td><td><0.001</td></tr></tbody></table><figcaption>Table 3. Linear regression coefficients for UHI intensity prediction.</figcaption></figure>
<h2>Discussion</h2>
<p>Our results demonstrate that random forest effectively predicts UHI intensity from Landsat-derived variables, achieving high accuracy (R²=0.89). This aligns with Lyu et al. (2022), who reported similar performance using ensemble methods. The superiority of RF over SVR may be attributed to its ability to handle nonlinear interactions and categorical-like features (e.g., LULC types).</p><p>Variable importance analysis confirms that impervious surfaces are the primary driver of UHI, consistent with previous findings (Bala et al., 2021; Shorabeh et al., 2020). NDVI's negative relationship underscores the cooling effect of vegetation, supporting urban greening strategies (Schwaab et al., 2021; Iungman et al., 2023). Albedo's moderate importance suggests that reflective surfaces can mitigate heating, though its effect is less pronounced than vegetation.</p><p>Our model's spatial predictions identify priority areas for intervention, such as industrial zones lacking green cover. The methodology is transferable to other cities, provided Landsat data are available. However, limitations include reliance on clear-sky imagery, which may miss cloudy periods, and the use of a single rural reference that may not represent all rural contexts (Sheng et al., 2017). Future work could integrate temporal dynamics and socio-economic variables (Kareem, 2023).</p>
<h2>Conclusion</h2>
<p>This study presents a machine learning framework for predicting surface UHI intensity using Landsat 8 data, with random forest achieving robust performance (R²=0.89). Impervious surface fraction, NDVI, and albedo emerged as key predictors. The approach provides a scalable tool for urban planners to identify heat-vulnerable zones and evaluate mitigation strategies. Future research should incorporate multi-source data, such as crowdsourced weather stations (Venter et al., 2020), and explore deep learning models for enhanced accuracy.</p>
<h2>References</h2>
<ol class="references">
<li>Dissanayake, D., Morimoto, T., Murayama, Y., Ranagalage, M. (2019). The Impact of landscape composition for urban heat island intensity in Addis Ababa City using Landsat data (1986–2016). <em>Abstracts of the ICA</em>, <em>1</em>, 1-1. https://doi.org/10.5194/ica-abs-1-63-2019</li>
<li>Singh, R. B., Grover, A. (2014). Remote sensing of urban micro-climate with special reference to Urban Heat Island island using Landsat thermal data. <em>Geographia Polonica</em>, <em>87</em>(4), 555-5668. https://doi.org/10.7163/gpol.2014.38</li>
<li>Tsou, J., Zhuang, J., Li, Y., Zhang, Y. (2017). Urban Heat Island Assessment Using the Landsat 8 Data: A Case Study in Shenzhen and Hong Kong. <em>Urban Science</em>, <em>1</em>(1), 10. https://doi.org/10.3390/urbansci1010010</li>
<li>Aris, A., Syaf, H., Yusuf, D. N., Nurgiantoro (2019). Analysis of urban heat island intensity using multi temporal landsat data; case study of Kendari City, Indonesia. <em>IOP Conference Series: Earth and Environmental Science</em>, <em>389</em>(1), 012002. https://doi.org/10.1088/1755-1315/389/1/012002</li>
<li>Srivastava, M. R., Satyaprakash, N. (2020). Urban heat island effect over Delhi NCR using LANDSAT™ data. <em>International Journal of Global Warming</em>, <em>22</em>(3), 272. https://doi.org/10.1504/ijgw.2020.110865</li>
<li>Oluseyi, I. (2010). An assessment of urban heat island of Lokoja Town and surroundings using LandSat ETM data.. <em>FUTY Journal of the Environment</em>, <em>2</em>(1). https://doi.org/10.4314/fje.v2i1.50790</li>
<li>Liu, L., Zhang, Y. (2011). Urban Heat Island Analysis Using the Landsat TM Data and ASTER Data: A Case Study in Hong Kong. <em>Remote Sensing</em>, <em>3</em>(7), 1535-1552. https://doi.org/10.3390/rs3071535</li>
<li>., R. H. A. (2016). USING LANDSAT-8 DATA TO EXPLORETHE CORRELATION BETWEEN URBAN HEAT ISLAND AND URBAN LAND USES. <em>International Journal of Research in Engineering and Technology</em>, <em>05</em>(03), 457-466. https://doi.org/10.15623/ijret.2016.0503083</li>
<li>Priyankara, P., Ranagalage, M., Dissanayake, D., Morimoto, T., Murayama, Y. (2019). Spatial Process of Surface Urban Heat Island in Rapidly Growing Seoul Metropolitan Area for Sustainable Urban Planning Using Landsat Data (1996–2017). <em>Climate</em>, <em>7</em>(9), 110. https://doi.org/10.3390/cli7090110</li>
<li>Sheng, L., Tang, X., You, H., Gu, Q., Hu, H. (2017). Comparison of the urban heat island intensity quantified by using air temperature and Landsat land surface temperature in Hangzhou, China. <em>Ecological Indicators</em>, <em>72</em>, 738-746. https://doi.org/10.1016/j.ecolind.2016.09.009</li>
<li>Lyu, F., Wang, S., Han, S. Y., Catlett, C., Wang, S. (2022). An integrated cyberGIS and machine learning framework for fine-scale prediction of Urban Heat Island using satellite remote sensing and urban sensor network data. <em>Urban Informatics</em>, <em>1</em>(1). https://doi.org/10.1007/s44212-022-00002-4</li>
<li>Bala, R., Prasad, R., Yadav, V. P. (2021). Quantification of urban heat intensity with land use/land cover changes using Landsat satellite data over urban landscapes. <em>Theoretical and Applied Climatology</em>, <em>145</em>(1-2), 1-12. https://doi.org/10.1007/s00704-021-03610-3</li>
<li>Wen, X. P., Yang, X. F. (2011). Urban Heat Island Monitoring and Analysis Using Landsat ETM+/TM Imageries. <em>Key Engineering Materials</em>, <em>467-469</em>, 3-6. https://doi.org/10.4028/www.scientific.net/kem.467-469.3</li>
<li>Trinh, T. H. T., Bui, T. H. T. (2023). Study of the urban heat island intensity using Landsat 8 in Da Nang in 2020. <em>Journal of Mining and Earth Sciences</em>, <em>64</em>(3), 79-89. https://doi.org/10.46326/jmes.2023.64(3).09</li>
<li>Nadizadeh Shorabeh, S., Hamzeh, S., Zanganeh Shahraki, S., Firozjaei, M. K., Jokar Arsanjani, J. (2020). Modelling the intensity of surface urban heat island and predicting the emerging patterns: Landsat multi-temporal images and Tehran as case study. <em>International Journal of Remote Sensing</em>, <em>41</em>(19), 7400-7426. https://doi.org/10.1080/01431161.2020.1759841</li>
<li>Wienert, U., Kuttler, W. (2005). The dependence of the urban heat island intensity on latitude A statistical approach. <em>Meteorologische Zeitschrift</em>, <em>14</em>(5), 677-686. https://doi.org/10.1127/0941-2948/2005/0069</li>
<li>Sameh, S., Zarzoura, F. H., El-Mewafi, M. (2023). Spatiotemporal analysis of Urban Heat Island and land use land cover changes using Landsat images and CA-ANN machine learning techniques: a case study of Dakahlia government, Egypt. <em>Journal of Spatial Science</em>, <em>69</em>(2), 551-572. https://doi.org/10.1080/14498596.2023.2257619</li>
<li>Abdul Kareem, S. (2023). AI - Driven Predictive Models for Urban Heat Island Mitigation Using Real - Time Environmental Data. <em>International Journal of Science and Research (IJSR)</em>, <em>12</em>(4), 1945-1948. https://doi.org/10.21275/sr23410213012</li>
<li>Lee, T., Lee, J., Wang, Z. (2012). Scaling of the urban heat island intensity using time-dependent energy balance. <em>Urban Climate</em>, <em>2</em>, 16-24. https://doi.org/10.1016/j.uclim.2012.10.005</li>
<li>Yılmaz, E. (2015). Landsat görüntüleri ile Adana yüzey ısı adası (Adana Surface Heat Island using Landsat Images). <em>SSRN Electronic Journal</em>. https://doi.org/10.2139/ssrn.3394253</li>
<li>Unknown (2022). Automated Mapping of Urban Heat Island to Predict Land Surface Temperature and Land use/cover Change Using Machine Learning Algorithms: Mansoura City. <em>International Journal of Geoinformatics</em>. https://doi.org/10.52939/ijg.v18i6.2461</li>
<li>Mutanga, O., Kumar, L. (2019). Google Earth Engine Applications. <em>Remote Sensing</em>, <em>11</em>(5), 591-591. https://doi.org/10.3390/rs11050591</li>
<li>Schwaab, J., Meier, R., Mussetti, G., Seneviratne, S. I., Bürgi, C., Davin, É. L. (2021). The role of urban trees in reducing land surface temperatures in European cities. <em>Nature Communications</em>, <em>12</em>(1), 6763-6763. https://doi.org/10.1038/s41467-021-26768-w</li>
<li>Iungman, T., Cirach, M., Marando, F., Barboza, E. P., Khomenko, S., Masselot, P. (2023). Cooling cities through urban green infrastructure: a health impact assessment of European cities. <em>The Lancet</em>, <em>401</em>(10376), 577-589. https://doi.org/10.1016/s0140-6736(22)02585-5</li>
<li>Li, Z., Wu, H., Duan, S., Zhao, W., Ren, H., Liu, X. (2022). Satellite Remote Sensing of Global Land Surface Temperature: Definition, Methods, Products, and Applications. <em>Reviews of Geophysics</em>, <em>61</em>(1). https://doi.org/10.1029/2022rg000777</li>
<li>Muthumanickam, D., Poongodi, C., Kumaraperumal, R., Pazhanivelan, S., Ragunath, K. P. (2022). Smart Farming: Internet of Things (IoT)-Based Sustainable Agriculture. <em>Agriculture</em>, <em>12</em>(10), 1745-1745. https://doi.org/10.3390/agriculture12101745</li>
<li>Venter, Z. S., Brousse, O., Esau, I., Meier, F. (2020). Hyperlocal mapping of urban air temperature using remote sensing and crowdsourced weather data. <em>Remote Sensing of Environment</em>, <em>242</em>, 111791-111791. https://doi.org/10.1016/j.rse.2020.111791</li>
<li>McLauchlan, K. K., Higuera, P. E., Miesel, J., Rogers, B. M., Schweitzer, J. A., Shuman, J. K. (2020). Fire as a fundamental ecological process: Research advances and frontiers. <em>Journal of Ecology</em>, <em>108</em>(5), 2047-2069. https://doi.org/10.1111/1365-2745.13403</li>
<li>Jain, P., Coogan, S. C., Subramanian, S. G., Crowley, M., Taylor, S., Flannigan, M. D. (2020). A review of machine learning applications in wildfire science and management. <em>Environmental Reviews</em>, <em>28</em>(4), 478-505. https://doi.org/10.1139/er-2020-0019</li>
<li>Fargione, J., Bassett, S., Boucher, T., Bridgham, S. D., Conant, R. T., Cook‐Patton, S. C. (2018). Natural climate solutions for the United States. <em>Science Advances</em>, <em>4</em>(11), eaat1869-eaat1869. https://doi.org/10.1126/sciadv.aat1869</li>
</ol>
</article>