Research on a novel extraction method using Deep Learning based on GF-2 images for aquaculture areas
Bo Cheng, Chenbin Liang, Liu Xunan, Yueming Liu, Xiaoxiao Ma, Guizhou Wang
Abstract
Recently, moderate-resolution remote sensing images have been widely used in extracting coastal aquaculture areas that are usually used to obtain marine fishery resources. However, in moderate-resolution remote sensing images, the aquaculture areas destroyed by storm-tide are hard to be observed. In order to get information of coastal aquaculture areas quickly and accurately, to help the scientific management and the planning of aquaculture resources and to meet the needs of disaster emergency response, it is imperative to extract aquaculture areas based on high-resolution remote sensing images. However, as the spatial resolution increases, there are many disadvantages that traditional pixel-level classification methods bring about, for example, the sediments and some floatage on the water surface are easy to be mistaken as aquaculture areas. In view of these problems, we proposed a novel Semantic Segmentation Network, named Hybrid Dilated Convolution U-Net (HDCUNet), which combines U-Net and Hybrid Dilated Convolution (HDC) to further expend receptive field and prevent the ‘gridding’ problem simultaneously. In this paper, we chose Gaofen-2 images as experimental data to extract aquaculture areas, and the Precisions and Recalls of two types of aquaculture areas are both above 95% and the overall accuracy reaches up to 99.16%. Besides, we compared the extraction accuracy with other four methods: FCN-8s, SegNet, U-Net and Threshold Segmentation (TS) to analyse the extraction effect and verify the validity of HDCUNet.