计算机应用

• 人工智能与仿真 •    下一篇

基于深度学习的无人机影像夜光藻赤潮提取方法

李敬虎1,邢前国2,郑向阳3,李琳3,王丽丽1   

  1. 1. 鲁东大学信息与电气工程学院
    2. 中国科学院烟台海岸带研究所海岸带环境过程与生态修复重点实验室
    3. 中国科学院 烟台海岸带研究所
  • 收稿日期:2021-07-09 修回日期:2021-09-14 发布日期:2021-09-30 出版日期:2021-09-30
  • 通讯作者: 李敬虎

Noctiluca scintillans red tide extraction method based on UAV images and deep learning

  • Received:2021-07-09 Revised:2021-09-14 Online:2021-09-30 Published:2021-09-30

摘要: 针对目前卫星遥感中夜光藻赤潮识别精度低、实时性差的问题,提出一种基于深度学习的无人机(UAV)影像夜光藻赤潮提取方法。首先,以无人机采集的高分辨率夜光藻赤潮RGB视频影像作为监测数据,在原有UNet++网络基础上,通过修改主干模型为VGG-16,引入空间dropout策略,增强特征提取能力并防止过拟合。然后,使用ImageNet数据集预先训练VGG-16网络进行迁移学习,以提高网络收敛速度。最后,为可靠地评估所提方法的性能,在自建的夜光藻赤潮数据集Redtide-DB上进行实验。所提方法的夜光藻赤潮提取总体精度(OA)为94.63%,F1评分为0.9552,Kappa为0.9496,优于K近邻(KNN),支持向量机(SVM)和随机森林(RF) 3种机器学习方法及另外3种语义分割网络(PSPNet, SegNet和U-Net)。在模型泛化能力测试中,所提方法对不同拍摄设备和拍摄环境的夜光藻赤潮影像表现出一定泛化能力,OA为97.41%,F1评分为0.9659,Kappa为0. 9382。实验结果表明,所提方法可以实现夜光藻赤潮自动化、高精度的提取,可为夜光藻赤潮监测和研究工作提供参考。

Abstract: Aiming at the problems of low accuracy and poor real-time performance of Noctiluca scintillans red tide extraction in the field of satellite remote sensing, a Noctiluca scintillans red tide extraction method based on Unmanned Aerial Vehicle (UAV) images and deep learning was proposed. Firstly, the high-resolution RGB (Red-Green-Blue) video data collected by UAV was used as the monitoring data, the VGG-16 (Visual Geometry Group-16) backbone and the spatial dropout method were used in the UNet++ network to enhance feature extraction and prevent overfitting respectively. Then the pre-trained VGG-16 backbone from ImageNet was applied to the process of training network to increase the convergence speed. Finally, in order to evaluate the performance of the proposed method, experiments were conducted on our red tide image dataset Redtide-DB. The Overall Accuracy (OA), F1 score, and Kappa of the red tide extraction are up to 94.63%, 0.9552, 0.9496 respectively in the proposed method, and the proposed method outperforms the traditional machine learning methods, such as K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest (RF) and other semantic segmentation networks (PSPNet, SegNet and U-Net). Meanwhile, the red tide image of different shooting equipment and shooting environments was used to test the generalization ability of the proposed method, the OA, F1 score and Kappa of the red tide extraction are 97.41%, 0.9659 and 0. 9382, respectively, and the result proves that the proposed method has a certain generalization ability for new data. The Noctiluca scintillans red tide extraction method provides a reference to the accurate red tide extraction in complex environments, and the work provides a reference for Noctiluca scintillans red tide monitoring.

中图分类号: