APPLIED RESEARCH ON DEEP LEARNING METHODS FOR URURBAN GREEN SPACE GREEN CONSERVATION MANAGEMENT -IMPROVEMENT OF PLANT IMAGE LEARNING MODEL AND APPLICATION TO PLANT DISTRIBUTION ANALYSIS-

Abstract

 This study aims to create a highly accurate deep learning-based model to classify plant species efficiently. To develop our model, we use the plant images of conserved green areas in urban areas that we collected and surveyed for the past three years. Using the learning model (Shiraishi et al., 2021), we constructed a method to obtain a plant distribution map from a bird's-eye view image. The 103 classification tests investigated by the spring of 2020 revealed that the correct rate of mugwort was lower than that of the previous model. We investigated image elements to find the cause of difference in correct rate between the leaf images and the flower images. Furthermore, we used an Explainable AI technology to analyze the decision made by our model. As a result, the correct rate of identifying the 121 species surveyed by the fall of 2020 was 98.3% just after our model was trained. Finally, we divided the bird’s eye view image into super pixels and classified them according to the plant type. We concatenated adjacent ones with the same type of plant and obtained the distribution of plants.