Research On The Extraction Of Ecological Resource In Qilian County,Haibei Prefecture Based On GaoFen Satellite Images

Semantic segmentation is an important means of interpreting satellite images,but in practice,semantic segmentation has always been constrained by low accuracy and the large number of annotations.Improving accuracy and reducing the number of annotation samples are two key researches to promote intelligent interpretation of satellite images in ecological environment monitoring.To solve the above two problems,this research adopts supervised deep learning and semi-supervised learning respectively,and takes Qilian County,Haibei Prefecture,Qinghai Province as research area,finally completes the following work:(1)We constructs a Qilian Ecological Resource Extraction Dataset including six categories of buildings,farmlands,forests,water bodies,roads and bare land.We use24 scenes satellite images produced by domestic high-resolution satellites and get Qilian Ecological Resource Extraction Dataset after preprocessing.The data set has a scale of 16,530 images and a spatial resolution of 2 meters.(2)To improve the performance of deep learning semantic segmentation network on satellite images,we propose a semantic segmentation network named Hrre Net for moderate-and high-resolution satellite images.The network uses high-resolution feature representation to compensate for the resolution loss,and uses a pyramid pooling module to extract multi-scale context information,and a module to refine boundary.Experiments show that Hrre Net not only performs well on Qilian Ecological Resource Extraction Dataset,and also achieves the best results on moderate-resolution dataset CCFD(73.36% m Io U)and high-resolution dataset Vaihingen(80.53% m Io U).(3)To reduce the amount of labeled images required for training,we proposes a semi-supervised semantic segmentation network named Sere Net for moderate-and high-resolution satellite images.The network uses an adversarial semi-supervised semantic segmentation network as the base architecture,and adopts a special data augumentation method to introduces a mix consistency loss,and creates a rotation task to introduce a rotation consistency loss.Experiments show that Sere Net not only performs well on category water and road of Qilian Ecological Resource Extraction Dataset,and also achieves the best results on moderate-resolution dataset GID5(72.35%m Io U)and high-resolution dataset Vaihingen(76.19% mIoU).