Image Segmentation Research Based On Wetlands And Urban Buildings In Shandong

It is of great value to classify wetland and monitor its dynamic changes using spectral characteristics analysis and machine learning method.With the increase of population,urban and rural construction develops rapidly.Reclamation,burial and occupation of wetlands,the discharge of harmful substances into the wetland.Therefore,it is emergency to monitor wetland toward to preserve the wetland environment and develop and construct reasonably.The dynamic monitoring of wetland Expect high precision,and there is no systematic and effective method.For example,the ecological characteristics of wetlands have great similarity,and the accuracy of automatic classification is not high;the phenomenon of "same object spectrum and foreign matter spectrum" remains to be solved.Urban building segmentation is a hot research field of high resolution remote sensing.However,the appearance and complex background of high-resolution remote sensing images make the precise semantic segmentation of urban buildings a challenge in related applications.Traditional feature learning methods can work well,but there are still some problems in the application of these technologies,which limits their wide applicability.However,the existing network model has too many parameters,so the information loss is easily caused by the sampling process.For the sake of solving these problems,we put forward the following points:1.We proposed a new method combined with muti-features,analysis these combination to choice preferable composition.This method makes image more smoothness.We got higher image precision by collecting a variety of feature variables of multi-phase landsat8 oli remote sensing images.The method uses the multi-phase Landsat8 oli image as the basic data source,combined with the multi-feature variables,discusses the separability of various types of objects in the research area,selects the best feature combination to analyze the segmentation effect of object-oriented classification method under different segmentation scales,and finds the optimal segmentation scale by calculating variance variation rate.Finally,the optimal feature combination and optimal segmentation scale are applied to three machine learning algorithms for wetland classification.The results show that:The best feature combination has better separation of the local objects,among which non wetland category is the highest,shrub swamp and herb swamp are difficult to distinguish among wetland categories.The NIR and red band in spectral band,greenness and fitness feature in the ear cap transform feature are the effective bands for wetland information extraction.The segmentation scale is determined in the object-oriented method,and the shape and compactness factors are set to 152.5,0.1 and 0.5,which has the best segmentation effect.Among the three machine learning methods,RF classifier has the best effect on wetland information extraction(overall accuracy,OA)and kappa coefficients,respectively,0.9139 and 0.8917.2.We propose an multi-layer to get a higher classification precision.The model is better at decrease network parameters,which in favor of limit time and complete model efficiently.The model compete with two parts.One is used for receiving input images and get more features,the other is used to restore the image.We proposed a new resblock to limit decrease information.The results demonstrate that it can train more efficiently.The classification precision with new residual block is 2.08% higher than that of the general unit neural network.