Analysis of YOLOv5 and DeepLabv3+ Algorithms for Detecting Illegal Cultivation on Public Land: A Case Study of a Riverside in Korea

Abstract

Rivers are generally classified as either national or local rivers. Large-scale national rivers are maintained through systematic maintenance and management, whereas many difficulties can be encountered in the management of small-scale local rivers. Damage to embankments due to illegal farming along rivers has resulted in collapses during torrential rainfall. Various fertilizers and pesticides are applied along embankments, resulting in pollution of water and ecological spaces. Controlling such activities along riversides is challenging given the inconvenience of checking sites individually, the difficulty in checking the ease of site access, and the need to check a wide area. Furthermore, considerable time and effort is required for site investigation. Addressing such problems would require rapidly obtaining precise land data to understand the field status. This study aimed to monitor time series data by applying artificial intelligence technology that can read the cultivation status using drone-based images. With these images, the cultivated area along the river was annotated, and data were trained using the YOLOv5 and DeepLabv3+ algorithms. The performance index mAP@0.5 was used, targeting >85%. Both algorithms satisfied the target, confirming that the status of cultivated land along a river can be read using drone-based time series images.