The Classification Of High Resolution Remote Sensing Images Based On Convolutional Neural Network

The classification of remote sensing image objects is a basic research focus in the field of remote sensing,which can obtain geospatial information to provide services for related applications by classifying the ground objects contained in the image.Because of can present ground objects clearly and detailedly,high resolution remote sensing images are used more widely.However,the traditional image classification methods are not effective for the high-resolution images.Therefore,it is necessary to find a method to effectively classify high-resolution remote sensing images.In recent years,convolutional neural network has been widely used because of its excellent performance in image task.In addition,the convolution of the neural network algorithms and techniques used to study the classification of remote sensing image features on the target,high resolution image is extracted through its intelligent learning a variety of characteristics,to achieve high resolution remote sensing image based on convolution neural network features target classification task,solve the use of traditional method for high resolution image classification effect is difficult to be effectively increase.This paper carries out the following research contents and work:(1)Based on convolutional neural network,the Deeplabv3+ framework is constructed in this paper and the specific parameters of model are trained on the high resolution remote sensing image training datasets,which is suitable for classification of high resolution remote sensing images.Then,this model is used to classify remote sensing image object targets,and the predicted results of the model are highly consistent with the classification and distribution of the original image targets,it indicate that this method can classify image feature targets effectively.(2)This paper also proposes an improved Deeplabv3+ network architecture,and a new model is obtained by retraining parameters in the high resolution remote sensing image training datasets.Then this new model is used to classify the same remote sensing image object targets,and output the results of the improved network predicted.Compared with Deeplabv3+,the network improves the classification accuracy of small ground objects.(3)The high resolution remote sensing image benchmark datasets of Potsdam was acquired and preprocessed,and the high resolution remote sensing image datasets was produced which containing more diverse and complex ground object targets,and experiments were carried out on the data set using the network.(4)Qualitative and quantitative analysis are used for the predicted results of networks output.The support vector machine is also used to compare with the methods in this paper which is processing by end-to-end,the network is advanced.The results show that the classification results of this methods have better visual effect and each evaluation index is higher than the classification results of the support vector machine.