Research On Deep Learning Semantic Segmentation Technology For Remote Sensing Image

Since the 20 th century,with the rapid progress of satellites and computer hardware equipment,China has gradually developed advanced high-resolution earth observation and inter satellite communication capabilities.Using advanced optical and electrical instruments,we can sense some characteristics of the target object from afar.This technology is widely used in various fields.Due to the developed remote sensing equipment,for some complex environments or specific terrain,through photography or scanning,we can timely penetrate and collect the electromagnetic wave band that cannot be seen by the naked eye,monitor the changing targets day and night in real time and analyze the collected data,so as to obtain diverse and comprehensive information.As an important image analysis technology in the field of remote sensing,semantic segmentation also has incomparable advantages for social development,technological evolution and natural exploration.Because the terrain is complex and changeable,and the hardware equipment is insufficient,it is difficult to detect long-distance,so the traditional remote sensing image segmentation often has the problem of inaccurate segmentation.How to accurately extract all pixels of each class in remote sensing images has become an important research topic in recent years.In this research,some progress has been made in the research of balancing accuracy and speed in the field of remote sensing image semantic segmentation.Two improved network models are studied and proposed.The specific contents and innovations are as follows:(1)In order to focus on learning the characteristic information of remote sensing images,attention mechanism is integrated.Different combinations of location attention module and channel attention module are introduced into different positions in the typical semantic segmentation model ICNet,and comparative experiments are carried out on Potsdam and Vaihingen in ISPRS data set.The experiments show that the improved network segmentation effect is better,and compared with the classical network model.Using the same evaluation index on the same data set,it is proved that the segmentation accuracy of the improved network model is improved.In order to further verify the effect of learning rate on segmentation effect,experiments were carried out under different learning rates,and the effects of their values on hardware requirements,model convergence speed,and segmentation accuracy were compared.(2)In order to solve the problems that the existing network models can not better capture feature information,the receptive field is small,and the required relationship does not match,based on Deep Lab v3+ network,the basic network is improved,and a multi-scale fusion semantic segmentation model for remote sensing images is constructed by introducing a multi-scale fusion strategy to capture context information.The comparative experiments on Potsdam and Vaihingen in ISPRS dataset show that the segmentation effect is better.In order to further verify the performance improvement of this model,the same evaluation criteria are used to compare with the classical network model on the same data set.Experiments show that the improved network model improves the segmentation accuracy.In this research,the above two network models are tested on large and public ISPRS remote sensing image datasets Potsdam and Vaihingen,which verify the feasibility and effectiveness of the design method in this research.