Machine Versus Human: A Comparison Between Artificial and Human Intelligence with Urban Tree Canopy Assessments in Wisconsin

Abstract

The urban forest makes up an integral portion of the landcover of urban areas. Currently 80% of the population in the United States lives in urban areas. Understanding landcovers within urban areas and specifically in this study the urban forest is important to a healthy living environment. The urban tree canopy (UTC) provides one way to assess the urban forest. There are multiple methods used to map the UTC with different levels of accuracy. Two approaches to estimate UTC include analysis of remote sensing data or aerial imagery with computer-based systems using artificial intelligence (AI) and human based interpretations using human intelligence (HI). This study compared these two methods (AI and HI). The AI and HI classifications for UTC did not differ significantly (p=0.723) for Dane county, Wisconsin where the AI method was trained. These two methods had a significant difference (p<0.001) in mean UTC overall for the state of Wisconsin. Thus, outside the training area the AI had a mean 26.0% UTC that produced a lower UTC estimate (5.5% UTC) than the HI (31.5% UTC). The change in UTC between 2013 and 2018 in 404 Wisconsin communities was also done with the HI system. There was a significant increase (p<0.000) in tree canopy from 2013 to 2018, approximately 1.7% from 31.6% to 33.3%. The HI UTC estimates were evaluated for agreement between two human assessors in each community estimate. There was over 90% accuracy for UTC agreement for both 2013 and 2018.