丁德秋,马丹,陈帆,樊妙,邢喆,唐秋华.基于蚁群算法优化极限学习机的声学底质分类方法[J].海洋通报,2024,(6):
基于蚁群算法优化极限学习机的声学底质分类方法
Optimization of acoustic substrate classification for extreme learning machine based on ant colony algorithm
投稿时间:2023-12-08  修订日期:2024-04-07
DOI:10.11840/j.issn.1001-6392.2024.06.006
中文关键词:  极限学习机  反向散射强度  底质分类  蚁群算法  图像滤波处理
英文关键词:extreme learning machine  backscattered intensity  sediments classification  ant colony algorithm  image filtering
基金项目:国家自然科学基金 (42206200);山东省自然科学基金 (ZR2023MD073)
作者单位E-mail
丁德秋 国家海洋信息中心天津 300171 2535743823@qq.com 
马丹 国家海洋信息中心天津 300171 md_nmdis@foxmail.com 
陈帆 国家海洋信息中心天津 300171 1518811678@qq.com 
樊妙 国家海洋信息中心天津 300171 fm_nmdis@163.com 
邢喆 国家海洋信息中心天津 300171 xz_nmdis@163.com 
唐秋华 自然资源部第一海洋研究所 山东 青岛 266075自然资源部海洋测绘重点实验室 山东 青岛 266590 tangqiuhua@fio.org.cn 
摘要点击次数: 57
全文下载次数: 11
中文摘要:
      海洋底质作为深海海底环境的重要组成部分,其类型及分布特征在深海资源开发利用和海洋工程建设中有着巨大的参考价值,是海洋测绘调查的一项重要内容,基于多波束测深系统采集强度数据的监督分类方法逐渐获得广泛应用。随着人工智能的迅速发展,神经网络在声学底质分类中得到广泛应用,极限学习机(extreme learning machine,ELM) 权值和偏重不再需要迭代优化,是一种学习速度较快的神经网络。针对ELM神经网络中由于初始权值和偏重矩阵随机确定而导致ELM分类器鲁棒性差的问题,本文选取蚁群算法优化ELM神经网络的初始参数,构建了ACO-ELM神经网络分类模型,经多次迭代后,由于信息素的累积,蚂蚁种群不断向最优路径偏移,训练精度逐渐增高,模型逐步达到平稳。通过底质分类实验验证表明,BM3D+ACO-ELM分类器处理的多波束声呐图像斑点噪声得到了有效控制,在西南印度洋脊龙旂热液钙质软泥和硫化物混合区域,BM3D+ACO-ELM分类器相比于其他三种分类器具有明显优势,底质分类精度得到较大提高,其中硫化物分类精度为93.23%,深海钙质软泥分类精度为93.78%。
英文摘要:
      As an important part of the deep-sea seabed environment, the types and distribution characteristics of marine sediments are of great reference value in the exploration and exploitation of deep-sea resources and the construction of ocean engineering, which are essential in marine surveying and mapping surveys. The supervised classification method based on the intensity data collected by the multi-beam sounding system has gradually been widely used. With the development of artificial intelligence, neural networks have been widely used in acoustic substrate classification. Extreme learning machine (ELM) weight and bias no longer need iterative optimization, which is a kind of fast-learning neural network. Given the ELM in the neural network as initial weights and bias caused by random matrix to determine ELM classifier problem of poor robustness, this article selects the initial parameters of ant colony algorithm to optimize ELM neural network, built the ACO - ELM neural network classification model. After many iterations, the training accuracy gradually increased, and the model gradually reached stability, due to the accumulation of pheromone ant colonies to the optimal path deviation. The results show that the multibeam underwater sonar image speck noise processed by BM3D+ACO-ELM classifier is effectively controlled by the substrate classification experiment. Compared with the other three classifiers, BM3D+ ACO-ELM classifier has obvious advantages in the hydrothermal calcinous oozies and sulfide mixing area in Longqi, southwest Indian Ocean. The accuracy of sediment classification is greatly improved, among which the accuracy of sulfide classification is 93.23% and that of deep-sea calcareous ooze is 93.78%.
查看全文    下载PDF阅读器
关闭