DeeplabV3+中引入注意力机制的小面积溢油识别方法
A Method for Small-area Oil Spill Detection by Introducing Attention Mechanism into DeeplabV3+ Network
李翔 1王增利 1芮小平 1邹亚荣2
作者信息
- 1. 河海大学地球科学与工程学院,南京 211100
- 2. 自然资源部国家卫星海洋应用中心,北京 100081;自然资源部空间海洋遥感与应用研究重点实验室,北京 100081
- 折叠
摘要
海面溢油危害大、分布广、时间空间不确定性大,对海洋溢油进行高精度识别与监测对全球生态环境的保护有着重要意义.针对传统卷积神经网络模型对小面积溢油分类不准确的问题,提出一种基于深度学习的SAR影像海上小面积溢油的高精度提取方法.该方法以DeeplabV3+网络模型为基础,引入SE注意力机制来提高网络对海面上小面积溢油的分类精度.基于欧空局开源数据集Oil Spill Detection Dataset建立了海面溢油提取模型,训练得到的平均交并比为77.70％,平均像素精度为98.16％.通过目视对比模型预测图和对比两类精度指标,发现经过注意力机制优化后的DeeplabV3+网络模型明显优于原始网络模型,对小面积溢油监测的效果明显提升.
Abstract
The dangers of oil spills at sea are extensive and widespread,and the time and spatial uncertainties are significant.Therefore,high-precision identification and monitoring of marine oil spills are of great importance to the protection of the global ecological environment.This paper addresses the problem of inaccurate classification of small-scale oil spills by traditional convolutional neural network models,proposing a high-precision extraction method for small-scale oil spills from SAR images based on deep learning.This method is based on the DeeplabV3+network model and improves the network's accuracy in classifying small-scale oil spills on the sea surface by introducing the SE attention mechanism.The paper establishes a sea surface oil spill extraction model based on the European Space Agency's open-source"Oil Spill Detection Dataset".The model achieves an MIoU of 77.70％and an MPA of 98.16％in training.By visually comparing the model prediction images and the accuracy metrics of the two classes,it is found that the DeeplabV3+network model,optimized with the attention mechanism,significantly outperforms the original network model.The addition of the attention mechanism has significantly improved the network's effectiveness in monitoring small-scale oil spills.
关键词
海洋溢油/深度学习/注意力机制/DeeplabV3+/SAR影像Key words
oil spill in the ocean/deep learning/attention mechanism/DeeplabV3+/SAR image引用本文复制引用
出版年
2024
NETL
NSTL
万方数据