Paper by Adnan Firoze, et al: “Historically, only resource-rich U.S. cities have collected data about where their public trees are, usually through labor-intensive manual surveys or via coarse canopy-cover estimation. However, a significant portion of city trees are on private property, making them difficult to quantify with surveys, yet they contribute uniquely to species diversity and ecosystem service distribution. Further, canopy-cover estimation cannot provide information about tree density, locations of trees across different land types, or changes in tree counts. Cities are under continual change, and the mean mortality rate of urban trees is twice that of rural trees.Thus, frequent updating of tree analytics is critical for sustainable, habitable cities.
Method. Recent advances in computing—in particular, generative artificial intelligence (AI)—have enabled our multidisciplinary team, spanning computer science, engineering, and forestry, to develop a first-of-its-kind computational method that can individually locate and maintain an inventory of trees in at least 330 U.S. cities (Figure 1). Using satellite data, this approach can complete the inventory process in less than a day of automated computing. Individual trees are challenging to discern in satellite images due to occlusion and resolution limitations, which in turn limits traditional segmentation-based approaches. Our approach leverages several key insights to enable a scalable generative AI solution. First, a frequent capture rate of satellite imagery (e.g., daily, monthly, etc.) provides spatiotemporal vegetation footprints, yielding richer information than single images. Our method includes a deep spatiotemporal vegetation cover classification using satellite images that classifies a city into tree, grass, and background, followed by a cluster-creation process and then individual tree localization using a set of conditional generative adversarial networks (cGANs). Further, our method can be applied to current or archived satellite imagery, allowing for change detection and historical analysis…(More)”.
