基于全卷积网络模型的高分遥感影像内陆网箱养殖区提取

Extracting inland cage aquacultural areas from high-resolution remote sensing images using fully convolutional networks model

为了研究高分遥感影像的内陆网箱养殖区自动快速提取,利用福建省北部内陆水域的GF-1影像和GF-2影像,并对影像中的网箱养殖区进行人工标注,经过旋转、缩放和镜像翻转等数据增强处理后构建了2种影像的内陆网箱养殖区样本库;利用样本库训练内陆网箱养殖区提取的深度学习全卷积网络(fully convolutional networks,FCN)模型并开展精度验证。结果显示,GF-1影像提取结果的F值达到83.37%,GF-2影像提取结果的F值达到92.56%。表明基于FCN的高分影像内陆网箱养殖区提取具有较高的精度,能够进行大规模内陆网箱养殖区提取应用,为内陆水产养殖区的监测提供重要依据。

The extraction of cage aquacultural areas was investigated using high-resolution GF-1 and GF-2 remote sensing images from northern Fujian Province. Image enhancement was performed by correction, fusion, and cropping. The sample database of inland cage culture areas of two kinds of images was constructed;The sample bank is used to train the in-depth learning fully convolutional networks(FCN) model extracted from inland cage culture area and verify the accuracy. The results of the test experiment show that the F-measure of GF-1 and GF-2 reaches 83.37% and 92.56%,respectively. It shows that the inland cage culture area extraction based on FCN has high accuracy, and can be used for large-scale inland cage acquaculture area extraction, which provides an important basis for the monitoring of inland aquaculture area.