The Research Of Remote Sensing Scene Classification Based On Convolutional Neural Network

With the rapid development of aerospace technology,the available remote sensing images not only have a higher spatial resolution,but also have a sharp rise in quantity.These advances pose great challenges to the information extraction and interpretation of remote sensing images.The classification of remote sensing image scene classification plays an important role in a wide range of applications and has been widely concerned by scholars domestic and oversea.In this study,cooperated with Hebei meteorological research institute,we focused on the classification of resolution remote sensing images.In this field,traditional classification strategies which orient pixels and objects exhibited deficiencies in extracting high-level semantic features of scenes and failed to perform the classification of the remote sensing scene effectively.Therefore,how to extract the high-level semantic features of remote sensing scenes and classify them effectively has become a hot issue of remote sensing image interpretation.In recent years,many scholars have made great efforts in generating various data sets and proposing diverse remote sensing image scene classification methods.However,the existing scene classification methods still have several limitations:(1)The lack of labeled samples;(2)Middle resolution remote sensing interpretation payed more attention to the extraction of spectral information than spatial information;(3)The inability of artificial low-level features dealing with scene description limited the accuracy of scene classification;(4)The middle or high resolution remote sensing images have the diversity in ground objects and scales.Besides,the scene features are lack of robustness.Aiming at these problems in scene classification of remote sensing images,this study focuses on the scene classification methods of middle and high resolution remote sensing images respectively.The main work includes:(1)We composed middle resolution remote sensing data sets for OLI and TM sensors.(2)The classical classification methods based on low level,middle level features and convolutional neural network were used to evaluate OLI and TM data sets respectively.(3)We proposed a remote sensing image scene classification method based on multi-scale feature fusion.In order to solve the problem of different size of ground object inthe middle resolution remote sensing scene,we utilized multi-scale pooling method to extract different scale feature information and obtained more comprehensive feature expression by connecting features of multi-level pooling.(4)A remote sensing scene classification algorithm based on transfer learning and multi-scale feature fusion was proposed.In view of the problems such as overfitting due to the small size of the open high resolution remote sensing data set and the serious deformation caused by the height change of the sensor,we adopted the idea of transfer learning to reduce the risk of overfitting and used the method of multi-layer and multi-scale feature fusion to reduce the impact of surface deformation on the classification accuracy.Experiments were carried out on two open datasets,and the results proved the effectiveness of the multi-scale feature fusion method.