Research On Automatic Road Extraction Of High-resolution Remote Sensing Image Based On Deep Learning

In the 21 st century,together with genetic engineering and nanotechnology,artificial intelligence technology is considered as the top three cutting-edge technologies and has an immeasurable effect on social progress and economic development.With the development of cloud computing,big data and computing hardware in recent years,artificial intelligence technology has ushered in the spring of development.As the fastest growing and most promising technology of artificial intelligence technology in recent years,deep learning technology has reached its blowout status in recent years and has achieved very good application results in many fields,especially in speech recognition and The field of image recognition has made breakthroughs or even surpassed most of human recognition accuracy.Although there are about thirty years of research on the technology of automatic road extraction for high resolution remote sensing images and many road extraction methods proposed during this period,due to the complex environment of high resolution remote sensing images,the extraction accuracy and integrity of roads have yet to be improved at present.However,the research on the application of deep learning method which has super learning ability in remote sensing image road recognition has just begun.High-resolution remote sensing image road automatic extraction technology is also expected to make a breakthrough by deep learning.In this paper,deep learning method is used to model and train for high-resolution remote sensing imagery road automatic extraction.The specific research works are as follows:(1)A method of automatically extracting high-resolution remote sensing images using the full-convolution network structure is proposed.The end-to-end road extraction of remote sensing images is performed without manual designed features,nor with the traditional sliding windows to combine deep learning with road extraction.The average accuracy of the model after full supervision training of the experimental data is 86%.(2)The study proposed a small sample training method for deep learning road extraction algorithm.At present,the deep learning algorithm requires too much data on the training database,and the data amount of more than 100,000 is often enough for the deep learning model to have better performance.Therefore,this paper does a deep research on the influence of training database on the performance of deep learning model,and then proposes a small sample training method.This paper applies the small sample training method to train the road extraction model in(1).The accuracy of the model is improved from 0.86 to 0.96.The performance of the model is not reduced but greatly improved compared to the model performance after the training of the original large number of training samples in(1).(3)This studied weak/unsupervised learning methods of deep learning.These methods include self-encoding learning,similarity mapping learning,generational confrontation learning,cyclic learning,cyclic generation confrontation,noise learning and correspondence learning.Its purpose is to make full use of a large amount of unmarked data for learning.The depth learning model based on the similarity mapping proposed in this paper has the highest extraction accuracy,and the road average accuracy reaches 88%.The extraction accuracy of this method even surpasses the full-supervisory training method in(1)and is worthy of follow-up research.