Classification Of Remote Sensing Scene Image Based On Convolutional Neural Network

With the development of remote sensing observation technology aimed at earth and increasingly perfect hardware facilities,remote sensing images with higher resolution can be acquired and collected.These high-quality images can provide more space,shape and texture information.These rich geological information have broad application prospects in fine agriculture,urban planning,cadastral survey,natural disaster detection and many other fields.Because remote sensing images contain a large amount of information,and the data application lags far behind its data acquisition ability,so how to quickly classify these massive images is the key to accelerate the utilization efficiency of remote sensing images.Traditional machine learning algorithms(such as SVM,decision tree,random forest,etc.)extract features from manually designed feature descriptors,and then use classifiers to classify these features.The key step of the algorithm is to extract the feature information contained in the image and then make further classification.However,due to the spatial complexity and diversity of remote sensing image data,it is difficult to extract accurate and valuable information from the image with the features of manual design,so these algorithms cannot meet the application requirements.Although sometimes some small sample data can also achieve good results,when it comes to large and complex batch image data,its generalization ability and classification effect is still difficult to meet the application requirements due to the limited computing power.With the rapid development of deep learning algorithm,the production efficiency of many fields,such as industry,medical treatment,society and security,has been greatly improved.The advantage of convolutional neural network is that it can extract the features of image data in a data-driven way,which stands out from many solutions of all remote sensing image scene classification problems,and this method has been widely used in image classification,semantic segmentation,target detection and other computer vision fields.But in some particular application scenario,the classification accuracy of network is still to be improved,and with the increase of network parameters,the calculation of the hardware of the cost and time cost is becoming more and more big,can not meet the real-time prediction ability,this also brings great difficulties to the ground deployment in the practical application of remote sensing image scene classification model.To solve the above problems,the CNN network structure needs to be optimized to achieve better results.The research of this paper mainly includes the following contents:(1)The important theoretical concepts of convolutional neural network,such as convolutional layer,pooling layer,fully connected layer,batch normalization,forward propagation,and back propagation,are introduced in detail.Finally,it provides some ideas for solving the problem of overfitting.In addition,the structure of the current mainstream convolutional neural network is studied and discussed in detail and make improvements to the current classification Dense Net network which has superior performance compared to many classic networks.The improvement is adding a way of supervision to the network structure,the third dense layer of original Dense Net network after pooling with all connections,add a new loss function,the new loss combines with the final loss of dense layer 4 as the ultimate loss and softmax classifier was used for output.Then the improved Dense Net network and the mainstream convolutional neural network were tested and compared on the nwpu-resisc45 data set and AID data set,and the classification accuracy of each network was statistically analyzed.It was found that the improved Dense Net network had the highest classification accuracy.This proved the superiority of the new network structure.(2)In view of the current large networks in engineering problems in the process of the ground,such as the model is too big that challenges the storage space of embedded devices,or the movement of the hardware device cannot meet the deep network model to calculate the force(FLOPS)demands to compress Dense Net model,this paper draws lessions on the depth of the Mobile Net separable convolution calculation to modify the Dense Net convolution module,and reduce network growthrate super parameter values and then a series of network depth comparison is tested.Finally,through two kinds of precision evaluation methods,namely confusion matrix and accuracy,it is verified that the compressed Dense Net model not only consumes less computing resources,but also achieves good accuracy.(3)Although the use of convolutional neural network to classify remote sensing scene images has achieved good results,in the process of practical application,we tend to operate under the fixed computing resources of hardware equipment.Even if the network performance is strong,if the computing cost is exceeded,the network model will not be available.How to design a better and more optimal model under the conditions of a given computing resources is particularly important.Based on this problem,this paper proposes a method for expanding model width and depth at the same time,and finally it was proved experimentally that adjusting network width and depth simultaneously than just one allows the model to achieve better classification effect.