Research On Semantic Segmentation Algorithm Of Land-cover Remote Sensing Image Based On Improved Unet

With the rapid development of aerospace technology,the acquisition of high-resolution remote sensing images is becoming more and more convenient.It’s significant to obtain the earth’s surface information from a large number of data and extract automatically and identify various features in the images by using semantic segmentation technology for the planning and management of land resources.Traditional image semantic segmentation can’t satisfy the application requirements in the era of big data because of its low accuracy and poor segmentation timeliness.In recent years,with the rapid development of semantic segmentation technology based on deep learning,deeper networks can be applied to solve more complex problems in the field of image segmentation.This paper aims to improve the accuracy of remote sensing image semantic segmentation,and proposes an improved semantic segmentation model based on Unet.The specific research content is as follows:(1)Taking the vaihingen remote sensing image as data set,the image is cropped by setting a sliding window with a certain length to make the size more suitable for training.In view of the problems of unbalanced and small scale in the distribution of data set samples,geometric transformations such as rotation,scaling,and flipping,pixel transformations such as noise brightness adjustment and filtering are used to expand the data to enhance the diversity and completeness of the data set to improve the generalization ability of the model.(2)On the basis of applying Unet and Seg Net networks to study the semantic segmentation of land-cover remote sensing images.this paper proposes a D-Unet model based on the Unet model.This model deepens the network by introducing residual blocks and extracts high-level semantic features containing rich information to obtain More complex surface information from remote sensing images.At the same time,a deep separable convolution is introduced at the bottom of the network to alleviate the problem of a large increase in parameters caused by the deepening of the network.(3)In order to solve the problem of large-scale small targets such as low vegetation in remote sensing images,and the problem of different sizes of internal targets.In the bottom layer of D-Unet,the atrous spatial pyramid pooling is introduced,and D-AUnet model is proposed,which fully integrates semantic information and location information.In order to overcome the problem of multiple pooling of the network,resulting in a rapid drop in resolution and loss of too much detailed information,this paper uses a deep separable convolutions with stride of 2 to replace the maximum pooling,and proposes D-A-DUnet model to achieve a layer-by-layer reduction in resolution,to further improve the accuracy of segmentation.The experimental results show that the improved semantic segmentation model proposed in this paper has achieved good segmentation results on the Vaihingen dataset.