| Flood disaster is one of the ten natural disasters that threaten human existence.Rapid and accurate acquisition of disaster information plays an important role in formulating rescue strategies and improving rescue efficiency and quality.With the continuous development of satellite remote sensing and UAV low altitude remote sensing technology,remote sensing technology has been widely used in flood water identification,flood dynamic monitoring,disaster analysis,emergency rescue and other work,has become an efficient and quick technical means.The traditional methods of remote sensing water interpretation include manual visual interpretation unsupervised classification based on statistics and supervised classification.The manual visual interpretation method has the advantages of less visual interpretation equipment,simplicity and convenience,but this method requires a large number of interpretation signs and rich interpretation experience,too much reliance on manual experience,and low efficiency.In the process of unsupervised classification,the phenomenon of the same object with different spectrum and the same object with the same spectrum leads to the difficulty of classification matching and low accuracy.The supervised classification method is greatly influenced by human subjective factors,and the selection and evaluation of training samples are highly dependent on prior knowledge.Therefore,the traditional water extraction method has some shortcomings such as relying on expert experience,low automation degree and low precision,and it has poor effect in the rapid extraction of water information for flood disaster emergency.In recent years,with the continuous development of artificial intelligence and machine learning technology,the intelligent analysis,automatic extraction and visual display of remote sensing data show the advantages of high timeliness,low cost,more convenient and more accurate with the help of and reliance on artificial intelligence machine learning algorithm based on remote sensing images.In the light of the flood water flooding UAV remote sensing image extraction,analysis the Segnet,Unet,Deeplabv3 + three flood water extraction effect of convolution neural network,and optimize the structure of Deeplabv3 + neural network,after replace the lightweight backbone network,the introduction of the migration study thought,improving measures such as increasing dense connectivity features fusion structure,to speed up the convergence speed of the neural network,reduce the size of training data neural network requirements,reduce the spatial hierarchical information and the purpose of small objects information loss.In the process of model training,the ability of neural network to extract remote sensing image features is improved,and better flood water extraction effect is realized.The research emphases and main contents of this paper are as follows:(1)Data set production.Aiming at the training data set of UAV remote sensing image,this paper proposes a unique cutting and selection method: "field" shape cutting and empty selection.The image samples selected by this method can not only highly reflect the characteristics of the whole flood water body in the study area,but also effectively reduce the number of repeated samples.(2)The improvement of Deeplab V3+ neural network.In this paper,Deeplabv3 + neural network is improved as follows:<1>Change the trunk network to Xception with a lightweight Mobile Net V2 network.<2> Network-based transfer learning which introduces the idea of transfer learning.<3> The parallel void convolution pyramid module ASPP of Deeplabv3 + neural network is reconstructed.(3)experimental analysis.In this paper,two experiments are set,namely Experiment 1 and Experiment2.Experiment 1 is used to select the neural network with the best flood water recognition effect,and Experiment 2 is used to verify the generalization of the neural network with the best flood water recognition effect.(4)Disaster thematic map and disaster information output.After the selection of the optimal model and the completion of generalization verification,the optimal model is used to identify and extract the flood water body in UAV remote sensing image.By superposition of vectorized data such as cultivated land,housing,road and water system,the remote sensing thematic map of flood disaster is output,and the disaster information is analyzed. |
