Classification of Green Tide at Coastal Area Using Lightweight UAV and only RGB Images

Abstract

Remote sensing has attracted much attention as a realistic solution to monitor green tide outbreaks quickly and efficiently, and the scope of its utilization is also gradually increasing. The use of remote sensing from satellites and airborne platforms has many advantages when monitoring large areas, but its utilization is very limited considering the cost and spatio-temporal resolution. In this work, the availability of high-resolution unmanned aerial vehicle (UAV) image detection for the detection of green tides on the Jeju coast of Korea is presented, and the classification accuracy was evaluated through the application of various classification algorithms. The UAV survey area was 1.0 km(2), and the spatial resolution of the UAV images taken at an altitude of 250 m was 4.86 cm. In this study, it was divided the survey area into four classes: Ulva, sand, seawater, and submerged Ulva. classifying them was attempted using only the RGB bands of a lightweight UAV. The high-resolution UAV images were classified as the Mahalanobis distance (MHD), maximum likelihood (MLH), minimum distance (MID), and artificial neural network (ANN) algorithms, with the highest accuracy being for the MLH and ANN methods. The green tide was calculated by counting only pixels classified as Ulva and submerged Ulva in the UAV image to determine the area of the green tide in the research area. The green tide areas estimated by the MHD, MLH, MID, and ANN classification algorithms were 0.29, 0.38, 0.30, and 0.37 km(2), respectively. Given that many unclassified pixels have been found using the MLH method and that this phenomenon has not been found in the ANN method, the ANN algorithm can be considered to be the most effective for coastal green tide classification. This study showed that high-resolution RGB images of lightweight UAVs could produce highly satisfactory classification results. The methods presented in this study can be considered as a very effective approach in terms of their ability to quickly acquire high-resolution images at low cost and detect vegetation with high accuracy. In the future, more diverse targets and areas will be available for identification if an accuracy assessment of classification results is made based on the spatial resolution of UAV images.