基于U-Net的国产高分卫星影像海水养殖区分类提取方法

Classification and extraction method of mariculture area from domestic high-resolution satellite images based on U-Net model

针对海水养殖区具有类型多样(筏式养殖、围海养殖、网箱养殖等)、差异较小和难以实现高精度分类提取的问题,本文将空洞金字塔卷积模块与U-Net神经网络模型融合,提出一种适用于海水养殖区的多分类提取方法。首先,采用变差函数分析方法发现不同类别的海水养殖区的变差函数波形与基台值的差异;其次,定义一种相仿于变差函数搜索域的不同扩张率空洞卷积并联的卷积结构,用其替换U-Net模型中普通卷积结构,构建ASP-U-Net模型;最后,为验证采用ASP-U-Net模型的海水养殖区分类提取能力,选取7景我国高分1/2号卫星影像为数据源,相较于经典FCN、SegNet、PspNet和经典UNet模型,ASP-U-Net模型对海洋养殖区的分类提取在多种指标下均最优,这说明使用提出的卷积结构能够有效扩大感受野,更适合于多类海水养殖区的特征表达。

Aiming at the problems of diverse types of marine mariculture areas(raft mariculture,reclamation mariculture,cage mariculture,etc.),small feature differences and the difficulty in achieving high-precision classification extraction,this paper integrates atrous spatial pyramid convolution module into the U-Net neural network model and proposes a multi-classification extraction method suitable for mariculture areas.Firstly,variogram analysis is used to find the difference between waveform and sill value of variogram of different types of mariculture areas.Secondly,define a parallel convolution structure of atrous convolutions with different atrous rates similar to the variogram search domain and replace the ordinary convolution structure in the U-Net model with it to construct the ASP-U-Net model.Then,in order to verify the ability to classify and extract mariculture areas using the ASP-U-Net model,7 scenes of domestic Gaofen 1/2 satellite images are selected as data sources.Compared with the classic FCN,SegNet,PspNet and classic U-Net models,the ASPU-Net model has the best performance among them in classification and extraction of marine mariculture areas under various indicators.It shows that using the proposed convolution structure can expand the receptive field effectively and is more suitable for the expression of characteristics in multi-type mariculture areas.