Research On Multi-target Segmentation And Detection Of Remote Sensing Maps Based On Deep Learning

The map has gone through thousands of years of research and development,and has long been a long-lasting practical tool.Even in the distant future,maps still have an irreplaceable position in industries such as navigation and geological research.The map was first hand-painted,then remotely transmitted remotely after the satellite was launched,and finally until the rise of the computer,the map gradually improved in response to people's needs.There are two forms of maps currently used by people,one is remote sensing map and the other is electronic map.The electronic map contains all the city's municipal areas,road traffic conditions and business information.Generally speaking,the information of roads,buildings,and waters in electronic maps are usually converted from the recognition of remote sensing satellite images.The main purpose of this article is to extract road,building,water and farmland information from remote sensing satellite map images.There are three difficulties in the identification of remote sensing satellite maps.The first difficulty is that the amount of information is too rich,the color diversity and the terrain information are too complicated.The second difficulty is that there are obstacles to feature extraction.Because the satellite is a bird's-eye view of the earth,there is a certain low-level occlusion phenomenon.The third difficulty is to recognize slow real-time update.As of 2019,the total mileage of highways constructed in China reached 4.846 million kilometers,ranking first in the world.While the rapid development of China's highways,the speed of updating China's infrastructure is accelerating,which undoubtedly increases the difficulty of real-time identification.Based on the above three issues,the use of deep learning to identify remote sensing satellites has become a mainstream research direction.This article will also conduct experiments and research based on this direction.In the 1970s and 1980s,research on information extraction methods for remote sensing satellites has begun.From the initial use of shallow networks to the fierce deep learning in the past decade,the research direction of remote sensing satellite information extraction has undergone many changes.The purpose is to improve the recognition accuracy and efficiency as much as possible.However,most of the research directions now focus on the identification of cities,ignoring the complex and diverse roads,variable-style buildings and waters of different colors in rural areas.The main research direction of this paper will focus on rural roads,buildings,waters and farmland,and focus on how to use multiple networks to extract remote sensing satellite map information under high resolution.The main task of this paper is to use deep learning to perform multi-target recognition on high-resolution remote sensing satellite maps.Before multi-target recognition,this paper will carry out binary classification of roads,buildings,waters,and farmland,and explore networks with better effects to improve.Before the experiment,a large number of label images will be obtained through specific labeling software,which will greatly reduce the training error and also ensure the diversity of the image.In this paper,a variety of networks are used for comparison in experiments to identify the one that is most suitable for multi-classification of remote sensing satellite images.This paper attempts to use ordinary convolutional neural networks to conduct experiments on remote sensing satellite images.In order to introduce D-LinkNet,other similar semantic segmentation networks will be used for comparison,including the classic FCN fully convolutional network and the UNet network commonly used for medical images.At the same time,in order to enrich the diversity of methods,Pix2 pix and Cycle GAN,which are derivative networks that generate adversarial networks,are used for comparison.After a lot of experiments,this article will focus on using DLinkNet for improvement.In order to maintain the recognition efficiency and recognition method of DLinkNet,this article will improve part of the structure of D-LinkNet network.In both binary classification recognition and multi-class recognition,the prediction method based on geometric transformation is used in the prediction,which improves the accuracy of prediction.The final structure of the network is only used for binary classification,which does not meet the needs of multi-target recognition in this paper.Therefore,after single-target recognition of D-LinkNet on roads,buildings,waters,and farmland,the experimental network was added to the One-Hot coding mechanism,and the last layer was output as a multi-channel form with categories of channels.The pixels are all in two categories,namely the background and the category of the channel,and the output probability value is between 0 and 1.In order to obtain better multiclassification experimental results,Softmax and Cross-entropy were added.It has been verified that it can achieve the effect of multi-classification,and the improved D-LinkNet network can well divide each area.