Remote Sensing Image Classification Based On Deep Learning

Remote sensing images contain extremely rich and complete information in spatial distribution and spectral characteristics of surface features,which is an objective reflection of surface information.The remote sensing digital images obtained by using various sensors can be easily and quickly used to obtain relevant ground feature distribution information,which is of great significance for the establishment and maintenance of basic geographic information databases,ecological protection,urban construction planning,and disaster monitoring.However,due to the complex information and scenes contained in the images,classification of remote sensing images is difficult to realized.Traditional remote sensing image classification methods need to manually extract the corresponding spectrums and texture features,and require too expert experiences and knowledge.Therefore,they are not good at discriminating the nuances of images,slowly to be processed,and do not have consistent classification standards.Deep learning allows computers to automatically learn relevant high-level features and semantic information in the image field.Compared with traditional image classification methods,using the features and semantics extracted by deep learning to classify and distinguish images can improve classification accuracy greatly.Due to the similarities between remote sensing images and traditional images,it is of high feasibility and practical values to migrate processing methods based on deep learning to remote-sensing images.At present,the methods based on deep learning are widely used in the process of the classification of remote sensing images,and have achieved satisfactory classification results,but there are also corresponding deficiencies,such as the need for further improvement in accuracy and the limited improvement in classification accuracy of a single model.Aiming at solving the above problems,this paper applies the methods in multi-model integration,attention mechanism,asymmetric convolution and WASP on remote sensing image classification on the basis of deep learning,And the following studies have been carried out:(1)On the basis of attention mechanism,a remote sensing image classification method is constructed based on the attention mechanism and asymmetric convolution.This article takes U-Net as the main network architecture,and increases its attention mechanism to obtain its important feature information for correlation classification after performing asymmetric convolution calculations through using asymmetric convolution blocks instead of basic convolution calculations.(2)Aiming at the multi-scale characteristics of the ground features in the remote sensing images,WASP(Waterfall-like Cavity Spatial Pooling Architecture)is adopted based on the attention mechanism and asymmetric convolution to further enhance the feature extraction capability of its network structures and improve classification accuracy.(3)Aiming at the problem of limited space for improving the accuracy of a single model,this paper uses a multi-model integration method to make each model complement learning from each other to further improve classification accuracy,and builds a multi-GPU parallel processing environment to improve classification efficiency while improving classification accuracy.Research shows that embedding attention mechanism,asymmetric convolution,and WASP into the U-Net network structure can improve the accuracy of remote sensing image classification.At the same time,integrating remote sensing image classification models can further improve the classification accuracy.The research results have certain reference value for remote sensing image classification research,and have certain practical value in production.