부산대학교 도서관

Precise and reliable identification of riparian vegetation along rivers is of paramount importance for managing bodies, enabling them to accurately plan key duties, such as the design of river maintenance interventions. Nonetheless, manual mapping is significantly expensive in terms of time and human costs, especially when authorities have to manage extensive river networks. Accordingly, in the present paper, we propose a methodology for detecting and automatically classifying the riparian vegetation of urban rivers. Specifically, the calibration of an unsupervised (Isodata Clustering) and a supervised (Random Forest) machine learning algorithm (MLA) is carried out for the classification of the riparian vegetation detected in aerial orthoimages with a resolution of 1 meter. Riparian vegetation is classified using Normalized Difference Vegetation Index (NDVI) features. In the framework of this research, the Isodata Clustering slightly outperforms the Random Forest, achieving a higher level of predictive performance and reliability throughout all the computed performance metrics. Moreover, being unsupervised, it does not require ground truth information, which makes it particularly competitive in terms of annotation costs when compared with supervised algorithms, and definitely appropriate in case of limited resources. We encourage river authorities to use MLA-based tools, such as the ones we propose in this work, for mapping riparian vegetation, since they can bring relevant benefits, such as limited implementation costs, easy calibration, fast training, and adequate reliability.