Lee, K. and Park, K.S., 2023. Deep learning model analysis of drone images for unauthorized occupancy detection of river site. In: Lee, J.L.; Lee, H.; Min, B.I.; Chang, J.-I.; Cho, G.T.; Yoon, J.-S., and Lee, J. (eds.), Multidisciplinary Approaches to Coastal and Marine Management. Journal of Coastal Research, Special Issue No. 116, pp. 284-288. Charlotte (North Carolina), ISSN 0749-0208.

Rivers in the Republic of Korea are classified into two types: national and local rivers. While large-scale national rivers are systematically managed and maintained by local governments, there are still limitations in managing small-scale local rivers. Various problems occur in these sites, such as illegal cultivation and farming, installation of shipping container houses and temporary buildings, garbage disposal, and installation of canopies, among others. Subsequently, disputes arise from unauthorized vehicle entries to collect aggregates. It also leads to additional temporal and manpower losses from supervising the large river sites. Thus, the necessity of utilizing drones and artificial intelligence in managing river sites has become increasingly evident. This study was conducted to find an AI algorithm that use drone images to decipher discrete objects existing in the river site. To accomplish this, 9 objects (water, automobile, road, farmland, green house, temporary building, round bale silage, canopy, and garbage) in the river site were annotated and processed by algorithms such as YOLOv5, YOLOv7, DeepLabv3+, HRNet, and U-NET to determine the most optimal algorithm. The mAP@0.5 evaluation metric was adopted for analyzing the results, with 85% set as the lower limit. As a result, it was confirmed that the YOLOv5 and DeepLabv3+ algorithms satisfied the index of 0.85, making them suitable for interpreting each property in the river site.