High-resolution Remote Sensing Image Classification Based On Convolutional Neural Networks

With the increase of the spatial resolution of remote sensing images,the ground-based information in the images can be clearly and accurately represented.This has greatly promoted the development of remote sensing applications.However,most of the current remote sensing image classification methods are based on low-middle-range images.Which cannot efficiently obtain effective feature information from high-resolution remote sensing images.This also poses a great challenge for the application of high-resolution remote sensing images.In recent years,the deep learning algorithm that has emerged provides a theoretical basis and an effective algorithm for the efficient and intelligent classification of high-resolution remote sensing images.Convolutional Neural Networks(CNN)is an important structural model for deep learning.Its main feature is to reduce the sensitivity of translation,rotation,and distortion during image classification through local connections,weight sharing,and space sub-sampling.Therefore,CNN is applied widely in the field of image recognition and classification.By imitating the information processing mechanism of human visual system,this paper introduces the sparse constraint of Lorentz function into CNN,and construct a sparse constraint convolution neural network model to achieve the high resolution remote sensing images classification.The main research contents of this paper are as follows:(1)Based on the traditional CNN model——Le Net-5,through changing the number and size of convolutional kernels,the pooling methods,the size of the pooling domain,the activation function,and the number of network layers to analyze these changes for classification results on the MNIIST data set.Finally,according to the experimental results,a nine-layer convolutional neural network model with better classification performance is constructed—CNN-9.(2)This paper imitating human visual processing sparse encoding system introduces the function of Lorentz sparse constraint into the CNN-9 to construct sparse convolutional neural network based on Lorentz function(LCNN).Through the MNIST classification,this paper will analyze the performance of LCNN from robustness,sparsity andclassification rate.The experimental results show that LCNN can ensure the accuracy of image classification while reducing the classification time of images.(3)A classification algorithm is proposed,which combines LCNN and Support Vector Machine(LCNN-SVM),and this algorithm is applied to high-resolution remote sensing image classification.The high-resolution remote sensing image data set is used as data source.LCNN is a feature extractor and SVM is a classifier.The accuracy of image classification results is evaluated from confusion matrix,overall classification accuracy,Kappa coefficient and classification duration.The experimental results show that compared with the traditional classification method,LCNN-SVM can obtain more accurate and reliable classification accuracy.Also,it can improve the classification efficiency.