Anhui Province is located in eastern China (114°54′–119°37′E, 29°41′–34°38′N), lying in the transition zone between the Yangtze River Delta and the Huaihe River Basin, with an area of approximately 140,100 square kilometers. Its terrain shows significant spatial variation: the southern part is dominated by mountainous areas, while the northern part is mainly composed of plains and hills. Major water systems in the province include the Yangtze River and the Huaihe River. Climatically, Anhui has a subtropical to warm-temperate monsoon climate, with an annual average temperature ranging from 14°C to 17°C and annual precipitation of 800–1800 mm.In terms of social economy, the population of Anhui Province was approximately 61.27 million in 2023, with an urbanization rate of 59.4%. As an important part of the western Yangtze River Delta, the province is experiencing rapid urban expansion. This geographical and socioeconomic background makes Anhui an ideal place for studying the driving effect of urban morphology on CUHI.

The surface air temperature (SAT) data utilized in this study is provided by the National Meteorological Information Center under the China Meteorological Administration. This dataset includes homogenized monthly and daily mean SAT, as well as maximum and minimum SAT records from 2419 national meteorological stations across China.

Over the past several decades, China’s land cover has undergone significant transformations amid economic development. The annual China Land Cover Dataset (CLCD) is a dynamic land use dataset released by Wuhan University. As developed by Yang and Huang (2021), this dataset was constructed with a 30-m spatial resolution using 335,709 Landsat images via Google Earth Engine. The most recent version of the dataset includes land cover information for China spanning from 1985 to 2021, with an overall land classification accuracy of 80%.

The population (POP) and gross domestic product (GDP) data used in this study, were obtained from the “China Statistical Yearbook” published annually by the National Bureau of Statistics. These data provide detailed information on the population distribution and economic development level over the years, which helps to analyze the relationship between POP and GDP factors and CUHII.

SAT measurements at relocated stations exhibited consistent decreases compared to pre-relocation baselines. These station-pair observational discrepancies during synchronous monitoring periods directly quantify CUHII in corresponding urban areas. Figure 2 delineates the spatial distribution of CUHII across Anhui Province, with recorded values spanning 0.06°C–1.12°C. Through natural breaks classification (Wen et al., 2019), seven distinct station groups were identified based on CUHII magnitude. The highest-intensity cluster (0.97°C–1.12°C) consists of five stations, dominated by Mengcheng Station in northern Anhui where pre-relocation built-up coverage reached 61.08%. Three regionally representative stations - Huainan (north Anhui), Hefei (central Anhui), and Dongzhi (south Anhui) - all exhibited CUHII exceeding 1°C, each surrounded by over 40% built-up areas prior to relocation. Conversely, the lowest-intensity group (0.06°C–0.21°C) contains four stations, including Fengtai Station in northern Anhui with only 15.2% pre-relocation built-up coverage. The remaining three stations in this category maintained built-up ratios below 20%, illustrating the direct correlation between urbanization levels and thermal environment alterations. This stratification systematically reveals how varying degrees of urban encroachment differentially influence microclimate conditions at meteorological observation sites.

In addition, this paper conducted a statistical analysis of the distribution characteristics of urban morphological parameters before station relocation. Taking built-up areas as an example, the key findings are as follows: for ARbt, the average coverage across all stations stood at 31.08%, with values ranging from 9.3% to 68.8%. The probability density showed a trend of first increasing and then decreasing as the built-up area expanded, approximating a left-skewed normal distribution, with the density peak occurring in the interval of 25%–30% built-up coverage; the average value of LPIbt was 24.37, with a maximum of 66.96 and a minimum of 2.49; the highest probability density was observed in the range of 20–25; fractal analysis results indicated that FRAbt had an average of 1.09, with values spanning 1.02–1.13, and the interval of 1.08–1.1 corresponded to the highest probability density; in terms of distance metrics, DISbt had an average of 0.76 km, ranging from 0.22 km to 1.75 km, with the density peak appearing in the interval of 0.71–0.82 km. These distribution patterns rigorously confirm that the station environments before relocation had already been significantly affected by urbanization. Specifically, the configuration of built-up land cover directly contributes to urban thermal bias in surface air temperature (SAT) records through altered energy balance mechanisms. The concentration of parameters within specific value ranges—especially the fact that ARbt exceeded 25% at most stations—further validates the existence of a critical threshold, beyond which urban encroachment begins to significantly distort microclimate observations.

This study quantified relationships between CUHII and urban morphology through correlation analysis of relocated meteorological stations. The heatmap (Figure 3) visualization employs color gradients (blue: negative, red: positive) and intensity scaling to represent correlation strength, with asterisks denoting statistical significance. LPIbt showed the strongest positive correlation (r = 0.68, p < 0.001), while ARc demonstrated maximal negative correlation (r = −0.59, p < 0.001). These patterns confirm urbanization-induced thermal impacts: expanding impervious surfaces reduce land thermal capacity and amplify anthropogenic heat, whereas agricultural lands enhance evapotranspiration cooling (Zeng et al., 2010; Zhao et al., 2018). Moderate correlations emerged for ARbt, LPIc, and POP, all p < 0.01. Contrary to previous findings (Knight et al., 2010), DISu showed non-significant correlation (r = 0.09), potentially reflecting regional sampling limitations - a methodological constraint to be addressed through expanded Yangtze River Delta research.

The predictive performance of two distinct modeling approaches—SVR and RF—was evaluated for CHUII estimation. SVR showed training performance (R ² = 0.91, RMSE = 0.05°C; Figure 4a) and maintained test phase constraints (R ² = 0.67, RMSE = 0.09°C). The RF model outperformed both approaches, achieving exceptional training accuracy (R ² = 0.95, RMSE = 0.04°C; Figure 4b) while maintaining robust test generalization (R ² = 0.76, RMSE = 0.08°C). Comparative metrics highlight RF’s superior capability in capturing nonlinear relationships. Training phase analysis showed 46% higher R ² than linear models and 4% improvement over SVR. Test phase performance demonstrated 124% R ² enhancement versus linear approaches and 13% superiority over SVR. The 50% RMSE reduction in RF’s test predictions compared to linear models further confirms its precision. Mechanistically, RF’s ensemble decision tree architecture effectively integrates multiple urban parameters through recursive partitioning. This methodological advantage positions RF as the optimal tool for urban climate modeling, particularly in complex urbanization scenarios requiring high spatial resolution predictions. Future implementations will integrate temporal urban growth data to enhance model generalizability across evolving metropolitan regions.

This study employed the RF model to evaluate the relative importance of urban morphological parameters in predicting CHUII (Figure 5). The analysis revealed distinct hierarchical contributions across four parameter categories. For land use parameters, ARbt emerged as the second most influential factor (importance = 0.18), consistent with prior findings that urban-rural surface contrast alters net radiation partitioning into sensible versus latent heat fluxes (Miao et al., 2012). ARc (importance = 0.15), ARw (0.11), and ARv (0.09) followed in descending order. Landscape parameters demonstrated greater dominance, with LPIbt (0.25) ranking first overall - a 39% higher importance than ARbt. This reflects how contiguous urban patches reduce surface heterogeneity, impeding energy-matter exchanges between adjacent land covers (Estoque et al., 2017). LPIc (0.17) showed comparable influence to ARbt, with urban vegetation’s cooling capacity attributed to synergistic evapotranspiration and shading effects (Givoni, 1994; Shashua-Bar and Hoffman, 2000). Location parameters exhibited moderate impacts, with distance to DISbt (0.09) surpassing POP (0.08) in predictive importance. Comparative analysis established land use and landscape parameters collectively account for 82% of model explanatory power, versus 17% from location/social factors. The critical parameter hierarchy was identified as: LPIbt > ARbt > LPIc > ARc > DISbt > POP. While these findings elucidate two-dimensional urban morphology effects, the study acknowledges a methodological gap in three-dimensional parameter integration. Previous research demonstrates building height variability coulud explains the street-level temperature variance (Chun and Guldmann, 2014), while tree canopy vertical structure contributes the cooling through multi-layer shading (Zheng et al., 2019; Guo et al., 2023). Future investigations will incorporate LiDAR-derived 3D metrics (e.g., sky view factor, frontal area index) through multisensor remote sensing fusion, enabling comprehensive characterization of urban form-thermal environment interactions across spatial scales.

The following section describes how the research methods and findings were adapted for a university course on Applied Data Science, demonstrating the practical utility of urban climate research in education.

The flipped classroom model restructures traditional pedagogy by reversing the sequence of knowledge transfer (pre-class self-study) and knowledge internalization (in-class interactions). Pioneered by Bergmann and Sams (2012), this approach shifts instructional focus to student-centered collaborative activities, where teachers facilitate discussions and problem-solving rather than deliver lectures. Key mechanisms include: 1) Pre-class knowledge acquisition through digital resources; 2) Classroom time dedicated to applying concepts through peer interactions and teacher guidance. This dual-phase design enhances student autonomy while strengthening socio-emotional connections through structured teacher-student dialogues. Empirical studies across disciplines confirm its effectiveness in fostering deeper cognitive engagement and meeting learners’ psychological needs (Schultz et al., 2014). This course achieves precise breakthroughs through a three-tier system reconstruction and optimizes the teaching process by leveraging the flipped classroom model: students engage in self-directed learning of basic theories of spatiotemporal data analysis before class, focus on technical practice and collaborative inquiry during class, and extend to industry application projects after class. The specific architecture is as follows (Figure 6).

During the knowledge input phase, students establish a spatiotemporal analysis framework through self-directed learning. They methodically study meteorological time series decomposition and spatial autocorrelation analysis using micro-lecture videos, perform sliding average filtering on temperature data in Jupyter Notebook, and generate daily urban heat island intensity curves. Teachers pre-release a library of typical urban heat island cases, requiring students to calculate heat island effects during heatwaves and conduct spatial overlay analysis of building density to form initial insights into spatiotemporal correlations. The knowledge internalization phase focuses on three core classroom components. First, students engage in case comparison discussions and group debates on the applicability of interpolation algorithms in complex urban areas. Second, they develop spatiotemporal density clustering algorithms, perform kernel density calculations on meteorological station data using Python, and optimize sliding time window parameters. Third, students collaborate to build 3D feature engineering modules (building height, sky visibility) and tune thermal environment feature matrices. For example, in feature engineering, they integrate street view images to extract aspect ratio indicators and construct multidimensional driver matrices with remote sensing ecological indices. During knowledge transfer, extended projects involve integrating multi-source data to construct random forest models for heat island prediction and ventilation corridor optimization. The final comprehensive solution includes a spatiotemporal data dashboard, Python API prediction model, and parameter sensitivity planning report, covering the full technical chain from data governance to planning decisions.

This study obtained ethical exemption from Tongling University Ethics Committee for analyzing anonymized classroom interaction data. Students were informed through institutional platforms about pedagogical data usage, with rights to voluntary participation and withdrawal preserved. All procedures complied with institutional guidelines and national regulations. Implemented in Spring 2023, the flipped classroom model enhanced students’ spatiotemporal analysis skills through a “pre-class construction, in-class internalization, post-class application” framework. A comparative study between 2022 (60-student experimental class: 32M/28F) and 2021 (58-student control class) cohorts showed significant improvements: experimental class mean score 93.7 vs. 74.8, with an 18.9-point difference confirmed by t-test (p < 0.001). Mastery of spatial interpolation rose from 58% to 93%, spatiotemporal autocorrelation accuracy from 47% to 85%, and machine learning modeling errors dropped from 72% to 37%—all metrics showing statistical significance (p < 0.001) (Table 1). Course enrollment increased, with higher proportions of high-scoring students. The research-teaching integration established an undergraduate practice mechanism: over 90% of students reported deepened machine learning modeling understanding, 80% noted improved research and writing skills. Fifteen undergraduates participated in urban thermal environment projects over three semesters, a 100% increase from pre-reform levels. This practice demonstrates that integrating real-world cases, innovative pedagogy, and comprehensive evaluations effectively enhances learning motivation, strengthens the understanding of CUHII physics mechanisms, and prepares students for academic and professional careers.