Classification Method For Improving Deeplabv3+high Score Remote Sensing Images

At present,my country is in a period of rapid development of remote sensing techniques and high-speed growth,but the corresponding semantic segmentation of remote sensing images is relatively slow.In the past,various threshold methods that have been used in the past have not been increased.The amount is classified,and the simple depth learning model has a serious missed missing point,as well as edge segmentation,etc.,which has not been able to meet the need to accurately divide the application needs.The remote sensing spectrum characteristics of the water body are easily influenced by the atmospheric,water quality,water environment,background land coverage,and the large area and the small river water sensing monitoring face features a variety of problems.Deep neural network because of the ability to fit the complex nonlinear function,and the semantic division of the image is one of the popular research contents of computer vision,the rich data of remote sensing images provides a massive source of data for its research,which can The characteristics of all the pixels of the target are widely used.The author is based on the research foundation of the existing semantic segmentation on the remote sensing image.It is not accurate in terms of depth learning semantics in remote sensing image classification,and the edge segmentation is poor,the small river is not coherent,and the results of river morphology.Two architectures,the main tasks of the paper are as follows:First,the author uses the water body of the remote sensing as a research object,collects the data set of Guangdong Province,the high-score,the data set passes the geometric refinement,image registration,image of the image,and cutting,using Labelme to make manual marking,production After the image is mirrored and revised,the data is unbalanced,and the classic semantic split network UNet,SegNet,PSPNet,Deeplabv3+,and then performs the classification results,the optimal network Deeplabv3+It is proposed two methods of Deeplabv3+and attention mechanisms.Method 1:Fertilizes the original picture to the empty convolution,first pass the channel information through the channel attention mechanism,then the ASPP structure of the encoder is connected in series,with the maximum poolization,and according to the decoder prediction according to the original network Split the picture;method 2:Take the original picture through the characteristics of the empty volume,one feed into the first time the space is the double-focus mechanism module,one feed to the ASPP structure,and then two The feature fusion is characterized by feature,using average poolization,and still predicts in accordance with the decoder of the original network.Finally,the effectiveness of the side wall annotation method is performed from algorithm time complexity and different convolution steps.Through experiments,the attention mechanisms taken in this paper are preferably classified in parallel with the ASPP structure,so that the morphology of the fine river is stored,and the interrupts present when the fine river classification is reduced.In addition,from the evaluation index,the MIOU of the method reached 97.81%,and the MPA reached 98.85%.Improve the distinction between the block lake water extraction,the distinction between the mountain shadow,towns,etc.In addition,the classification method of the improved Deeplabv3+high-score image proposed by the study facilitates the full use of remote sensing data to accurately monitor the dynamic changes of land water in lakes,rivers,reservoirs,and promotes the close-related ecology that is closely related to terrestrial water.In-depth understanding and solving environmental problems.