Student Research Reports

View Research Report

Rapid urban expansion and the proliferation of heat-retaining surfaces such as asphalt and concrete contribute to elevated temperatures in cities, also characterized as surface urban heat islands (SUHI). To increase the accuracy of surface temperature machine learning models, key for urban planning and disaster management, this study harnessed land cover use data collected by citizen scientists. The data comprised downward photos collected by 2024 NASA SEES Earth System Explorers interns using the GLOBE Observer app at 958 sites across the U.S. After labeling for land cover use via Zooniverse, each site was associated with labeled Sentinel-2 satellite data from Collect Earth Online and a mean land surface temperature (LST) Landsat-8 satellite reading for June 2024. Random Forest and XG Boost models were trained in Python on three distinct datasets – coordinates only, coordinates with GLOBE data, and coordinates with GLOBE and Collect Earth Online data – to develop predictive models for LST. Following Bayesian optimization with 10-fold cross-validation, Random Forest displayed R2 accuracies of 0.84, 0.78, and 0.79, respectively. XGBoost displayed R2 accuracies of 0.82, 0.80, and 0.82, respectively. While incorporating land cover data failed to improve predictive accuracy, improving data collection methods and ensuring higher quality data could reveal the true value of citizen-sourced land cover data. This research supports a deeper understanding of the complex relationship between land cover and LST, potentially aiding urban planners in mitigating SUHI effects and fostering community engagement in scientific research. Keywords: citizen science, land surface temperature, machine learning, urban heat islands