DOM-based Water Body Pattern Recognition Method

As the data put into the database,DLG is very important to maintain timeliness.How to use images to obtain the latest DLG data and reduce field surveys is of great significance to speed up the entire data update cycle.Water body elements occupy an important position in the terrain database data.The most important thing to update the water body is to obtain the latest water body distribution position.At present,the DLG data update mostly uses the old data and image manual update method,which takes time.At present,most commonly used image automatic water pattern recognition methods are mostly based on remote sensing satellite images.Because of the easy access to remote sensing satellite images,most of the resolutions are low.The ground features are usually small and the spectral information is rich.The method of spectral extraction is often used.However,there are two problems with this type of method.The first point is that traditional recognition methods require artificial selection of features,resulting in inconsistent classification results.The second point is that because of the remote sensing image,the spectral information of the image is richer,and the feature information such as the texture and shape of the image is lacking,resulting in the dark objects such as building shadows,asphalt roads,and dense vegetation being mistakenly divided into water bodies.Satellite images are also very expensive,so a new and cheap method is needed to improve the extraction of water information.The paper proposes a water body pattern recognition method for the current problem.This method combines common convolutional neural network algorithms with semantic segmentation,uses convolutional neural network to extract features,and uses image semantic segmentation to identify water bodies.The experimental data is DOM.Train the existing high-resolution DOM data,use the trained network to optimize the water image to be recognized,and finally use the optimized water body information to update the original water body vector data.Specific work includes:First,crop and enhance the existing DOM data,configure the relevant development environment,label the created training set data,mark the water body in red,and the rest of the background display in black.Second,the existing semantic segmentation network and convolutional neural network are algorithmically fused,and the training set is input into the combined network.Using MIo U(Mean Intersection over Union)as the basis of the experimental effect,the network-related parameters are adjusted until Until the satisfactory segmentation result is reached.Third,optimize the fully connected conditional random field based on the probabilistic model design of the recognition result.After the optimization,the results are vectorized.The original water body vector data and the latest recognized water body data are calculated to determine the type of recognized water body,such as a river Or lakes,etc.,to complete the DGL data update of the water surface.This article provides a new idea for extracting water image information and updating DLG database data,and proposes a method for identifying water bodies through deep semantic learning.It fully uses the texture,shape,and color information of water elements to improve the accuracy of classification.Moreover,the feature extraction by machine avoids the manual selection of the feature set by the user in use,and ensures the unity.