Invasive species identification from high-resolution 4-band multispectral imagery

Invasive tree species pose major threats to various ecosystems, and their accurate identification and mapping are vital to the development of effective management strategies. Many invasive species demonstrate unique phenological characteristics that are visually identifiable through remote sensing. Previous species identification research relies heavily on the fusion of multiple remote sensing data sources, and are highly constricted in spatial scale. Here, we used high resolution, single instance, 4-band aerial imagery acquired during the blooming season to identify and map Callery pear (Pyrus calleryana) across all five New York City Boroughs (~ 1.188 km²) in urban, suburban, and non-developed environments. We compared traditional pixel-based filtering models against U-Net based convolutional neural networks (CNN). The U-Net CNN greatly outperformed the traditional pixel-based models, achieving a precision, recall, and F1 score of 86.9%, 89.5%, and 88.2% respectively compared to a performance of 47.2%, 52.7%, and 49.8% for the best of the pixel-based models. We also greatly improved CNN performance through the introduction of negative training data, specifically in non-urban areas. We show an effective deep learning strategy for identifying and mapping canopy coverage of Callery pear, which provides a base map for monitoring and management of Callery pear in the Greater New York City Metropolitan Area. More importantly, the method can be readily applicable to the mapping of Callery pear in other regions or other invasive species with unique phenological characteristics given the availability of punctual, high-resolution, multispectral imagery.