Machine Learning in Urban Tree Canopy Mapping: A Columbia, SC Case Study for Urban Heat Island Analysis

Abstract

As the world’s urban population increases to the predicted 70% of the total population, urban infrastructure and built-up land will continue to grow as well. This growth will continue to have an impact on the urban heat island effect in all of the world’s cities. The urban tree canopy has been found to be one of the few factors that can lessen the effects of the urban heat island effect. This study seeks to accomplish two objectives: first, we examine the use of a commonly used machine learning classifier (e.g., Support Vector Machine) for identifying the urban tree canopy using no-cost high resolution NAIP imagery. Second, we seek to use Land Surface Temperature (LST) maps derived from no-cost Landsat thermal imagery to identify correlations between canopy loss and temperature hot spot increases over a 14-year period in Columbia, SC, USA. We found the SVM imagery classifier was highly accurate in classifying both the 2005 imagery (94.3% OA) and the 2019 imagery (94.25% OA) into canopy and other classes. We found the color infrared image available in the 2019 NAIP imagery better for identifying canopy than the true color images available in 2005 (97.8% vs. 90.2%). Visual analysis based on the canopy maps and LST maps showed temperatures rose near areas where tree canopy was lost, and urban development continued. Future studies will seek to improve classification methods by including other classes, other ancillary data sets (e.g., LiDAR), new classification methods (e.g., deep learning), and analytical methods for change detection analysis.