Urban Thermal Dynamics at Pixel Resolution: Neighborhood-Specific Analysis Using Machine Learning and Multi-source Geospatial Data in Guadalajara, Mexico

Abstract. This research examines the relationship between urban physical characteristics and land surface temperature within discrete thermal pixels obtained from Landsat 8 Collection 2 Level-2 products in Guadalajara, Mexico. The city’s extensive collection of high-quality geospatial data enables urban thermal analysis with high granularity. We integrate multiple datasets: a LiDAR-derived urban tree inventory with individual tree metrics; building footprints with roof material classifications; green space polygons; transportation infrastructure including roads and sidewalks; and water body delineations. Our methodology focuses on the pixel level—for each 30-meter thermal pixel (900 square meters), we precisely quantify all urban features within its boundaries, creating a comprehensive dataset where each pixel contains measurements of all elements present. This spatial integration enables multivariate regression modeling using machine learning, where predictor variables represent the quantity of each urban element and the target variable is pixel temperature. Using interpretable machine learning techniques, we identify the relative importance of urban elements in predicting thermal patterns, achieving R² values ranging from 0.69 to 0.83 across different urban contexts. This pixel-level approach provides granular understanding of urban thermal dynamics, supporting evidence-based urban planning decisions for thermal management.