Construction of a Novel Data Set for Pedestrian Tree Species Detection Using Google Street View Data

Cities of the future will carefully manage their ecological environment, including parks and trees, as critical resources to balance the effects of climate change. As such, tree health has become an integral part of well-managed cities and urban areas, where tree censuses provide a critical source of ecological data. This data provides information to improve the ecological status of these areas; however, gathering this data is laborious and requires expert knowledge to accurately register each tree species included in the census. With recent advances in object-detection methods, automating this type of census is now possible. However, these approaches require training data to be gathered, labelled, and validated. This study merged data from a tree census in the Auckland region (New Zealand) with Google Street View image data and used a pre-trained model on specialised datasets released by prior authors to create a training dataset for pedestrian-view tree species detection. This approach can be used as the basis for wider data collection and labelling of New Zealand urban tree species, crucial for inventorying the state and health of its urban forests. Here, we demonstrated that training and deploying a fine-grained object detection model to an edge device for real-time inference on a video stream can achieve speeds of 25 frames per second (fps).