Multi-Model Machine Learning Analysis of Urban Temperature Trends: A Comparative Study on Climate Change Impacts in U.S. Cities of Midwest KANI Region

Urban temperature prediction is critical for regional climate planning, environmental monitoring, and thermal hazard mitigation. This study employs a multi-model supervised machine learning framework to predict and forecast daily urban air temperatures and evaluate model performance across key counties in the U.S. Midwest KANI region: Polk (IA), Pulaski (AR), Lancaster (NE), and Johnson (KS), encompassing 38 urban centres. Using ERA5-Land reanalysis data (2000-2024) from the cloud-based Google Earth Engine platform, this study compares six regression-based ML models: Linear Regression, Random Forest, XGBoost, Support Vector Regression (SVR), K-Nearest Neighbors (KNN), and Decision Tree to evaluate their predictive efficacy in forecasting urban temperature changes. Using 288 months of temperature data across 4 major economically important counties, we trained and evaluated each model using R², RMSE, and MAE metrics. Ensemble tree-based models XGBoost and Random Forest achieved the strongest performance in daily temperature forecasting across all counties from 2020 to 2024, with R² values around 0.91, RMSE between 2.60°C and 3.54°C, and MAE as low as 1.88°C. These models successfully captured seasonal dynamics, with forecasted daily temperatures ranging from –15°C during winter extremes to over 31°C in summer peaks. A Friedman test followed by Nemenyi post-hoc analysis confirmed that Decision Tree significantly underperformed compared to XGBoost (p = 0.04) and SVR (p = 0.03), while XGBoost, RF, SVR, and KNN formed a statistically indistinguishable high-performance cluster (p > 0.05). Linear Regression and Decision Tree were both outside this group, exhibiting poorer accuracy and greater bias, particularly in extreme conditions. These findings emphasize the superior reliability of ensemble methods for operational climate forecasting and highlight the practical forecast range of –15°C to +32°C, enabling precise early warning systems for climate adaptation and heat risk planning across vulnerable counties in the U.S. Midwest.