基于卷积神经网络和数据融合的筏式养殖区提取

Extraction of the raft aquaculture area based on convolutional neural networks and data fusion

准确提取海水筏式养殖区信息对于海洋资源管理和环境监测具有重要意义,但是筏式养殖区养殖筏因淹没于水中常出现数据弱信号区域的现象,导致仅凭光学影像提取精度较低。因此,本文以威海荣成湾为研究区域,通过添加通道注意力机制改进U-Net神经网络并结合高分2号光学影像光谱信息以及高分3号雷达影像纹理信息,尝试提高筏式养殖区提取精度。结果表明:(1)无论是对于单一的光学影像还是光学和雷达影像融合影像,添加通道注意力机制的U-Net神经网络预测结果总体精度都会提高,提高幅度在2.21%~4.12%之间。(2)利用改进后的U-Net神经网络处理融合数据,总体精度达到95.75%,相对于仅用高分2号影像的精度高4.3%;(3)对于弱信号区域,利用改进网络以及融合数据提取的总体精度和Kappa系数分别为91.61%和0.8277。该方法可以对海洋筏式养殖区弱信号区域进行有效提取,能够为海洋养殖面积统计以及海洋环境检测提供技术支持。

Accurate extraction of marine raft aquaculture area information is of great significance for marine re-source management and environmental monitoring.But the raft culture area is often submerged in water with weak data signal areas,resulting in low extraction accuracy based on optical images alone.Therefore,this paper takes Weihai Rongcheng Bay as the research area,and improves the U-Net neural network by adding channel attention mechanism to combine the spectral information of GF-2 optical image and the texture information of GF-3 radar im-age,trying to improve the extraction accuracy of raft aquaculture area.The results show that:(1)Whether it is a single optical image or a fusion image of optical and radar images,the overall accuracy of the prediction results of the U-Net neural network with channel attention mechanism will be improved,with an increase of 2.21%-4.12%.(2)Using the improved U-Net neural network to process the fusion data,the overall accuracy is 95.75%,which is 4.3%higher than that of only using GF-2 image.(3)For weak signal region,the overall accuracy and Kappa coeffi-cient of extraction based on improved network and data fusion are 91.61%and 0.8277,respectively.This method can effectively extract the weak signal area of marine raft aquaculture area,and can provide technical support for marine aquaculture area statistics and marine environment detection.