深度边缘光谱U-Net海水网箱养殖信息提取  

Marine cage aquaculture information extraction based on deep edge spectral U-Net

作　　者：柯丽娜[1] 翟宇宁 范剑超[2] Ke Li’na;Zhai Yuning;Fan Jianchao(School of Geography,Liaoning Normal University,Dalian 116029,China;Department of Marine Remote Sensing Technology,Na-tional Marine Environmental Monitoring Center,Dalian 116023,China)

机构地区：[1]辽宁师范大学地理科学学院,辽宁大连116029 [2]国家海洋环境监测中心海洋遥感技术室,辽宁大连116023

出　　处：《海洋学报》2022年第2期132-142,共11页

基　　金：国家自然科学基金(42076184,41876109,41806207,41706195);国家重点研发计划(2016YFC1401007,2017YFC1404902);国家高分重大科研专项(41-Y30F07-9001-20/22)。

摘　　要：网箱养殖是海水养殖的重要类型之一,传统网箱养殖目标光谱特征受近岸植被、水体影响较大,易出现噪声问题。新型深海网箱养殖目标离岸较远,但养殖目标海面框体部分较小,与自然水体光谱相似性较高,难以实现有效提取。本文提出深度边缘光谱U-Net模型对两种海水网箱养殖类型进行养殖信息提取。该模型通过Canny算子双边滤波算法去除波段运算后冗余光谱信息,提取边缘光谱特征并利用U-Net跳跃连接结构将其与深度卷积网络特征相融合,经softmax分类器逐像素分类实现网箱养殖信息提取。以海南近岸网箱养殖与深海网箱养殖为研究对象进行养殖信息提取,经实验对比所提方法在传统近岸网箱目标上精确度达到97.35%,新型深海网箱目标上提取精度达98.99%,其提取结果明显优于传统无监督算法和典型深度学习网络模型。Cage aquaculture is one of the most important types of mariculture. The spectral characteristics of offshore cage aquaculture are greatly affected by the coastal vegetation and water body, so it is easy to cause noise problems. The new type of deep-sea cage aquaculture target is far from the shore, but the sea surface frame part of the aquaculture target is small, which has high spectral similarity with the natural water, and is difficult to extract.In this paper, deep edge spectral U-Net(DES-Unet) model is proposed to extract aquaculture information of two types of cage aquaculture. In this model, Canny operator bilateral filtering algorithm is used to remove the redundant spectral information after band operation, and the edge spectral features are extracted. The U-Net jump connection structure is used to fuse the edge spectral features with the deep convolution network features, and the pixel by pixel classification of softmax is used to extract the cage aquaculture information. Taking the offshore cage aquaculture and deep-sea cage aquaculture in Hainan Island as the research objects, the aquaculture information is extracted. The experimental results show that the accuracy of the proposed method is 97.35% on the offshore cage target and 98.99% on the deep-sea cage target. The result is better than the classical unsupervised algorithms and traditional deep learning model.

关 键 词：网箱养殖 U-Net模型 深度特征 边缘光谱特征 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置] S967.3[自动化与计算机技术—控制科学与工程]