Extraction of typical oyster pile columns in the Maowei Sea, Beibu Gulf, based on unmanned aerial vehicle laser point cloud orthophotos

Abstract

Pile culture is a breeding method commonly used for oyster seedlings in the intertidal zone of southern China. Artificial visual interpretation serves as the primary monitoring approach for oyster seedling cultivation in marine areas. Manual visual interpretation is often time-consuming, inefficient, and does not provide spatially continuous information about the structure. Consequently, obtaining data on oyster pile columns and oyster seedling culture areas presents certain limitations. This study focuses on Shajing Town, Qinzhou City, Guangxi Zhuang Autonomous Region, China, as its research area. It utilizes multi-spectral image data from unmanned aerial vehicles (UAVs), light detection and ranging (LiDAR) point cloud technology, and deep learning algorithms to extract representative oyster pile columns in Maowei Sea within Beibu Gulf. By employing band features and texture indices extracted from UAV’s multi-spectral images as data sources and combining them with a classification and prediction model based on deep learning convolutional neural networks (CNN), we successfully extract the desired oyster pile columns. The results demonstrate that: 1) By comparing three machine learning models and integrating the LiDAR point cloud oyster pile column height model (OPCHM) into the S3 scenario, the convolutional neural network (CNN) attains an impressive overall classification accuracy (OA) of 96.54% and a Kappa coefficient of 0.9593, significantly enhancing and optimizing the CNN’s predictive accuracy for classification tasks; 2) In comparison with conventional machine learning algorithms, deep learning exhibits remarkable feature extraction capability.