Study On Water Body Recognition Based On Deep Learning In Remote Sensing Image Of Cold And Dry Areas

It is of great significance to extract water information from remote sensing images of cold and arid areas in terms of water resources scheduling,disaster warning and ecological environment management.The water extraction technology based on deep learning can extract water information efficiently and timely,provide a large range of water semantic information in real time,and greatly reduce the cost of human and material resources.However,the terrain of cold and dry areas is complex,and there are many mountains and valleys.The remote sensing images are easily affected by mountain shadow,dry river bed,vegetation and so on.At the same time,there is a lack of training set for deep learning network training.Therefore,how to accurately extract the water features and refine the semantic segmentation is an urgent problem to be solved.This paper focuses on the difficulties of water extraction in cold and dry areas,fully considers the features of cold and dry areas,combines with the latest advantages of semantic segmentation network technology,and makes innovation on this basis.The main work of this paper is as follows(1)The data set of water extraction in cold and dry areas was constructed.Network training needs a large number of labeled data to fully extract the effective features in the training data.However,the existing data sets can not meet the requirements of water extraction in cold and dry areas,and can not collect the unique features of cold and dry areas.Therefore,this paper collects data from remote sensing images of cold and dry areas,and uses Labelme tool to label them artificially.In addition,in the actual training,in order to learn the semantic features of the data as much as possible,the data is amplified,and various situations are simulated by reducing and filling in zeros,clipping and enlarging,multi angle rotation,and various chroma transformations.(2)This paper proposes a R-Linknet network which combines transfer learning with hole convolution.In order to improve the efficiency of network training,speed up the training process and integrate migration learning,resnet50 is migrated to Linknet network as encoder.Dense ASPP is added in the middle of network codec to expand the network receptive field and extract more detailed features.At the same time,the core degradation problem caused by hole convolution combination of traditional pyramid pooling structure is avoided.(3)This paper proposes a non-uniform sampling method with multi sampling in high frequency region and less sampling in low frequency region.In the traditional image semantic segmentation,all regions are uniformly sampled,which will result in massive increase of data volume in the face of high-precision remote sensing image semantic segmentation.However,in semantic segmentation,most of the low-frequency regions are slow changing regions with little semantic change,so there is no need to sample this region too much.Therefore,the sampling method of non-uniform sampling can reduce the amount of calculation and improve the training efficiency without reducing the density of semantic feature extraction.(4)To explore the effect of semi supervised learning on network performance.Most of the existing network training methods use labeled data for training,but the production of labeled data hinders the development of deep learning.Unlabeled data also has a lot of semantic features.At the same time,the collection of unlabeled data is relatively simple.In order to make full use of the easily available unlabeled data and achieve the purpose of improving the accuracy of the network.The effectiveness of the method is verified by adding different proportions of data to the training set.Through the construction of data sets,the application of transfer learning,the construction of efficient network,the proposal of targeted modules and the upgrading and optimization of loss function,the network and method proposed in this paper have achieved excellent results in water extraction from remote sensing images in cold and arid areas.