Research And Application Of Deep Learning In Remote Sensing Field

Remote sensing image is the image taken by unmanned aerial vehicles,satellites,etc.Remote sensing images are getting more and more easy in modern society,but the problem of land planning,disaster prevention and resource monitoring still needs human beings,but remote sensing images have not been used rationally.To this end,it is necessary to develop a algorithm that can identify,cut up remote sensing image,let the computer instead of human to solve these problems.The traditional pixel-based algorithm can not well eliminate the interference of light,weather and so on,and can not detecte well in the state of dense objects.Therefore,we try to apply the deep learning convolution neural network to the remote sensing image.we divide the algorithm into two steps,first distinguish the different types of features,and mark the exact edge;and then the results can be used to comparison analysis.Two steps results can be applied in different situations.In order to detecte and cut up the multi-classification,multi-scale,dense ground objects remote sensing images,we will improve the traditional convolution network structure,adopt the "end to end" full convolution network,skip structure,sparse convolution kernel,etc.At the same time,using a variety of data enhancement methods,it can not only to expand the data set,but also to increase the object state to avoid over-fitting.Through a large number of experiments,our network had the most appropriate network parameters.The improved convolution neural network has better fault tolerance for atmospheric and seasonal disturbances,and has higher recognition rate of dense objects and can adapt to remote sensing images of different scales.After the algorithm is completed,a lot of test experiments were carried out on the performance and function to ensure the stability and practicability of the algorithm.The test results show that the accuracy of segmentation is 80% or more on the specific test set of remote sensing data.