LB 41-350 - Estimating statewide urban tree canopy cover with Google Earth Engine

Many communities and organizations quantify the extent of urban tree canopy in cities and towns to calculate urban forest benefits, document change over time, and inform management decisions. However, manually classifying urban tree canopy from aerial imagery is tedious and time consuming. Machine learning algorithms can expedite the image classification process, though these approaches vary in their accuracy and complexity. Researchers have developed multiple types of urban tree canopy image classification systems for many large cities and metropolitan areas, though less attention has been paid to smaller cities and suburban towns. Our goal was to develop and test a machine learning urban tree canopy image classification approach that is applicable to both large and small communities using Google Earth Engine. We chose Google Earth Engine because it is free for non-commercial use, accessible via a web browser, requires minimal programming competency, and has an extensive collection of National Agricultural Imagery Program (NAIP) data. Our study focused on U.S. Census defined urban areas in Wisconsin, USA where we had ready access to a large collection of georeferenced tree data. We used a pixel classification approach with Google Earth Engine’s built-in random forest algorithm applied to 2019 NAIP imagery. We classified four types of land cover: trees, grass, impervious surfaces plus soil, and water. Training and test points were randomly selected from several pre-existing data sources. The final model used eight predictors, which included imagery characteristics such as the normalized difference vegetation index and entropy. Our model had an overall classification accuracy of 92%. In general, model classification was better in Low Population Density areas compared to Medium and High Density, though differences were modest. One performance metric demonstrating this was sensitivity, the percent of correct classifications made on points in the land cover of interest. For example tree classification sensitivity was 95% in Low compared to 91% in Medium and High while grass sensitivity was 93% in Low and 86% in High. These results demonstrate how our model can create reasonably accurate estimations of urban tree canopy using Google Earth Engine. The model can be used by state agencies and other community groups to estimate urban tree canopy and guide policies for the management of the urban forest.