Research On Classification Method Of High Spatial Resolution Remote Sensing Image Based On CNN

Currently,remote-sensing image is widely used in many fields such as land use and planning,environmental monitoring and protection,transportation planning and navigation,and national security.In high spatial resolution remote sensing images,it presents more clearly the details of surface features,more diverse types of surface features more complex backgrounds,and larger image scales,etc.This makes the traditional remote sensing image classification methods unable to meet the application requirements of high spatial resolution remote sensing images.With the great success of deep learning in natural scene images,deep learning methods have also been extended to the field of remote sensing.This thesis applies convolutional neural network to the field of remote sensing image classification.Mainly studied the three aspects of large-scale image prediction splicing algorithm,data enhancement and Dense ASPP model optimization:（1）Predictive splicing algorithm.Because of the large scale of remote sensing images,the performance of most computers used for deep learning training and prediction cannot meet the requirements.It is necessary to crop large-scale images into small-scale images to train and predict.In the model prediction stage,how to stitch the small-scale images output back to the large-scale images? Currently,research on this is relatively scarce.This thesis implements the boundary filling prediction splicing algorithm,and proposes the boundary regression prediction splicing algorithm,the probability overlay prediction splicing algorithm and the test-time data enhancement prediction splicing algorithm.Experiments show that the algorithm proposed in this thesis effectively improves the prediction effect of the model,especially the probability overlay prediction,which can improve large-scale image boundary prediction and splicing traces after prediction.During the test,the data enhancement prediction effect is better than the probability overlay prediction.But the improvement is limited and the consumption of computing resources is several times or even dozens of times that of the probability overlay prediction,which is suitable for application scenarios that pursue high precision.（2）Data enhancement.Aiming at the data enhancement stage of large-scale images,random cropping will change the distribution of data in large-scale images.This thesis proposes an online area random cropping data enhancement algorithm.First,the amount of training data that needs to be generated is equally divided into each large-scale image.Secondly,each picture is divided into an area.Finally,the area is randomly cropped,so that the generated training data tends to the original data distribution.Experiments show that the online area random cropping data enhancement algorithm is superior to online enhancement and offline enhancement.（3）Dense ASPP model optimization.This thesis combines features the lightweight of Mobile Net V2’s use of deep separable convolution and inverse residual modules and the Dense ASPP network’s large receptive field,multi-scale and dense sampling,perform network reconstruction on Mobile Net V2 as an encoder and Dense ASPP as the decoder.According to this method,experiments show that our proposed network has better performance.