Research On Classification Of High Resolution Urban Sensing Image Method Based On Deep Learning

Remote sensing image has been used for a long time in the study of surface feature classification.The appearance of high-resolution remote sensing image can reflect the features of urban surface feature more accurately.With the help of deep learning technology,accurate extraction of all kinds of ground object information in the city can be completed,as well as extraction and analysis of urban buildings,vegetation,roads and other information,which can further understand the status of the city and has certain significance for urban planning and development.Based on the 0.5m remote sensing images of Autonavi satellite,this thesis explores the classification method of deep learning.According to the features of ground objects in the remote sensing images,cities are divided into 9 categories,including vegetation,water,roads,general buildings,high-rises,urban villages,shadows,playgrounds and others.The specific research content is as follows:(1)Four deep learning models with high accuracy,such as Unet++,Deeplabv3+,Unet and ICnet,were studied and compared,and the classification model was trained on the self-made high-resolution remote sensing data set.Finally,the Unet++ model had the highest accuracy among the four models.The average accuracy rate reached 91.88% and the average crossover ratio reached 87.98%.The results show that the UN ++ model is more suitable for urban high-resolution remote sensing image classification.(2)For the original Unet++ model,local information is prone to noise,boundary regions between some classes are confused,and the convergence process is slow.Firstly,a triple attention model is introduced from the perspective of dimensional interaction,which reduces the attention of some unimportant parameters and improves the training efficiency and accuracy of the model.Secondly,a deep supervision mechanism of multilevel output fusion is introduced to combine the characteristic information of each level in detail,which effectively improves the accuracy of the model.Finally,the cosine annealing learning rate reduction method is used to make the model quickly escape from the local optimal solution and accelerate the convergence.In order to verify the effect of the improved model,the model effect comparison experiment before and after the improvement was designed.The results show that: The classification effect of the improved Unet++ model has been significantly improved.The overall classification accuracy is 95.04%,the average crossover ratio is 92.50%,and the Kappa coefficient is 0.9394,which are respectively2.68%,4.38% and 0.0325 higher than before the improvement.(3)The improved Unet++ model was used to extract the ground feature information of Baohe District,Shushan District,Yaohai District and Luyang District of Hefei,and the number of pixel points of various ground features in the four areas were counted.The results showed that the road ratio of Shushan district was the highest,reaching 6.22%,and the traffic was developed.The proportion of general buildings and tall buildings in Yaohai District is the highest,reaching 14.46% and 4.66% respectively,indicating a high degree of urbanization.The proportion of vegetation in Luyang District was the highest,reaching52.38%.The water proportion of Baohe District is the highest,reaching 24.68%,and the water resources are abundant.