urban heat riskexplore satellite imagery through deep learning
introduction
as cities grapple with climate change impacts, understanding urban heat distribution becomes increasingly critical. in this study, i investigated san francisco's heat vulnerability patterns using landsat 8 satellite imagery, focusing on the relationship between urban development and temperature variations.
methodology
my approach combined remote sensing techniques with machine learning methods. using google earth engine's python api, i processed 2020 landsat 8 imagery at 30-meter resolution, carefully selecting scenes with minimal cloud cover (< 5%) to ensure data quality. the analysis incorporated several environmental indices to capture different aspects of the urban landscape: vegetation coverage through ndvi, urban density through ndbi, and surface temperature through thermal infrared measurements.
conclusion
interestingly, the composite heat risk index i developed (combining lst, ndbi, and ndvi) showed remarkable variation across the city, ranging from -270.80 to 29,885.79. this wide spread reflects san francisco's diverse urban landscape, from densely built downtown areas to vegetated residential neighborhoods. to classify these patterns, i tested three machine learning approaches. while the svm model achieved nearly perfect accuracy, suggesting overfitting, the random forest classifier provided more reliable results with 91% accuracy and 86% precision in identifying high-risk areas.
perhaps most significantly, the analysis revealed clear spatial patterns of heat vulnerability corresponding to urban development intensity. these findings could prove valuable for urban planners and policymakers in developing targeted heat mitigation strategies.
copyright @liuhaobing