Remote Sensing Scene Classification On Optimized Convolutional Neural Network

With the rapid development of satellite sensing technology,the resolution of acquired remote sensing images is getting higher and higher.How to extract valuable information from high-resolution remote sensing images has become a huge challenge.Accurately classifying remote sensing scenes has become the focus and difficulty of current research on high-resolution remote sensing images.In recent years,a large number of domestic and foreign remote sensing launched a scene classification method is proposed series of remote sensing scene classification,including those based on low,medium and feature convolution neural network feature extraction method,effectively extract the remote sensing scene feature information.However,the existing classification method still has the following problems:(1)the current classification does not take into account the characteristics of remote sensing images,limits the performance of classification;difficulty(2)high-resolution remote sensing data sets for small training problems,leading to the use of the depth of the network model training effect is limited.For the problems,this paper's research work can be summarized as follows:(1)proposes a method for remote sensing scenario based on the random sub-image classification model.For remote sensing images have similar characteristics adjacent cell,so that the characteristics of the image sensing local information as the information can expression of the entire image.Remote sensing image adjacent similar cell characteristics of the different distributed randomly sheared by the size of the sub-picture Cauchy,predicted distribution of these sub-images classified as the basis for discriminating the original image,to improve the data set on the target classification accuracy.(2)the training dataset difficult for small high-resolution remote sensing image problems,optimizing neural network convolution scene classification method.For nerveinduced activation function appears RELU element disappeared,instead of a method using RELU function Leaky Re LU,and a natural image with large data sets pre-trained network model is used as a large depth feature extractor,the final fine-tuning parameter.By repeatedly comparing experimental results show that the effectiveness and prove the depth of the impact of network feature extraction.