基于改进FCN网络的GOCI绿潮分类研究.docxVIP

基于改进FCN网络的GOCI绿潮分类研究 赵尊强　陈鑫　刘慧芳　杨光 摘 要：为提高GOCI影像中绿潮分类的精度和速度，并克服传统方法中阈值难以确定的缺点，基于全卷积网络，提出一种新的网络结构用于绿潮和海水分类。首先降低全卷积网络深度，然后加强网络结构中高层特征与底层特征网络的链接，最后通过Softmax层生成高级语义特征进行最终分类。在GOCI数据集上的实验结果表明，改进的网络模型可以取得很好的竞争性，召回率、F1系数和Kappa系数分别达到83.0%、84.4%和83.7%。与传统方法相比，基于深度学习模型的GOCI影像绿潮分类，不仅提供了更加详细的绿潮分布信息，而且缩短了整个绿潮解译图制作流程。 关键词：GOCI;绿潮分类;全卷积网络;深度学习;绿潮解译图 DOI：10. 11907/rjdk. 192330 开放科学（资源服务）标识码（OSID）： 中图分类号：TP301 文献标识码：A 文章编号：1672-7800（2020）003-0069-05 Greentide Classification for GOCI Based on Improved FCN Network ZHAO Zun-qiang，CHEN Xin，LIU Hui-fang，YANG Guang （School of Computer Science Engineering，Shandong University of Science Technology，Qingdao 266590，China） Abstract： In order to improve the accuracy and speed of greentide classification in GOCI images， and to overcome the shortcomings of thresholds in traditional methods， we propose a new network structure for greentide and seawater classification based on a full convolutional network. Firstly， the depth of the full convolutional network is reduced. Secondly， the link between the high-level features of the network structure and the underlying feature network is strengthened. Finally， the generated high-level semantic features are finally classified by the Softmax layer. The experimental results on the GOCI dataset show that the improved network model can achieve good competition， and the recall rate， F1 coefficient and Kappa coefficient reach 83%，84.4% and 83.7%. Compared with the existing traditional methods， the GOCI image greentide classification based on the deep learning model not only provides more detailed greentide distribution information， but also shortens the production process of the entire greentide interpretation map. Key Words： GOCI; green tide classification; full convolutional network; deep learning; green tide interpretation map 0 引言 在過去的十几年中，高光谱图像（HSI）发挥了重要作用。绿潮（Green Tide）作为一种生存能力和繁殖能力较强的大型藻类[1]，其大规模暴发对我国近海岸生态环境、旅游业和海洋生产都会造成影响[2]。20XX奥运会期间黄海海域爆发的绿潮，覆盖