| With the improvement of accuracy and resolution of remote sensing satellite images,the high-definition remote sensing satellite images have become an important data source for monitoring changes in the ecological environment of lakes in different regions.For the lakes in cold and arid regions present complex texture feature information such as uneven distribution of water bodies,large differences in seasonal changes,complex surrounding vegetation and topography,and surface eutrophication on high resolution remote sensing satellite images,the existing image segmentation algorithms do not propose an optimization scheme for the extraction of large-scale refinement of lake water bodies in cold and arid regions.In recent years,the level of semantic segmentation techniques based on deep learning has increased dramatically,and the scale of image segmentation and the completeness of segmentation targets have been greatly improved.Therefore,this paper uses a semantic segmentation network model based on deep learning as the basis,and selects s even lakes distributed in the Inner Mongolia Autonomous Region of China that meet the typical cold and arid regions as the research objects.Two new water body extraction methods based on deep learning are proposed.The main work of this paper is as follows:(1)Build semantic segmentation dataset.The existing dataset of water body characteristics is relatively single and does not reflect the characteristics of water bodies in the study area of this paper.In this paper,we address different seasons of multiple lakes in the study area,remote sensing images from different satellites and a variety of specific representative disturbance images to enhance the professionalism of the dataset.And pre-processing operations were performed on the dataset to produce a semantic segmentation dataset suitable for the characteristics of the study area in this paper.(2)A SER34 AUnet water extraction method based on an improved Unet network model is proposed.First,the attention mechanism SENet module is added to the residuals module of the Resnet34 feature extraction network,highlight the feature information of the water body part in the output feature map during downsampling,And the above-mentioned improved SER34 network with stronger ability to extract water features in images is used as the feature extraction part of the Unet model.Secondly,the dual relationship-aware attention mechanism DANet is added to the improved Unet model to further improve the accuracy of the model in extracting the water body parts.Finally,the segmented image of the water body is obtained by upsampling and feature fusion operations on the output feature map through the decoding part of Unet.(3)A SER50 AFCN water extraction method based on an improved FCN network model is proposed.First,the attention mechanism SENet module is added to the residuals module of the Resnet50 feature extraction network,highlighting the feature information of the water body part of the output feature map during downsampling.And the above improved network with deeper layers and better ability to extract water features in images is used as the feature extraction part of the FCN model in the SER50 network.Secondly,the hybrid attention mechanism CBAM is added to the improved FCN network to make the model have global adaptive perception capability and pay more attention to the water body part of the image.Finally,the segmented image of the water body is obtained by the up-sampling operation of the FCN model.(4)The two network models are trained and tested on a homemade dataset,and the performance is compared with three baseline network models.The results show that the two methods proposed in this paper outperform the three baseline models in terms of pixel accuracy,category-average pixel accuracy,and average intersection ratio,and have higher accuracy in extracting water bodies on the homemade dataset.Subsequently,the proposed network model is trained and tested on a public dataset using the proposed network model in this paper.The results show that the proposed method outperforms the above three baseline models and has a high generalization capability. |
