Identification and Counting of betel nut tree from UAV Imagery by Deep Learning

Keywords: Deep learning,UAV,Object detection,Betel nut
Abstract Type: Poster Abstract

Abstract

Betel nut was one of the largest agricultural cash crops in Taiwan, second only to rice. Because betel nut is mainly planted on hillsides, its shallow root system is considered not conducive to soil and water conservation by the Taiwanese government. Furthermore, the International Cancer Research Center has listed betel nut as a first-class carcinogen, and the Ministry of Health and Welfare (Taiwan) also pointed out that oral cancer is highly correlated to betel nut chewing among men in Taiwan. As a result, the government has launched several measures to actively reduce the total number of betel nut trees by helping farmers switch to other occupations and subsidizing farmers who give up betel nut planting based on the number of trees. However, because the subsidies are limited by traditional manual calculation methods which the number of betel nut trees could be uncertain and doubtful. This study proposes an automatic solution to collect high-resolution georeferenced images by drones, and to apply object detection based on deep learning methods in order to calculate the number correctly in a short time. Models from three state-of-the-art object detection network architectures (one-level, two-level, and anchor-free) are conducted and evaluated by comparing their confusion matrices. The trained models could be used in real-time UAV missions to assist in-field Betel nut detection.