Research On Object-Oriented Convolutional Neural Network For Landsat8 Image Classification

Land cover is an important basic data for the study of the earth's ecological environment,land management and sustainable development.Whether in the field of human economy or the earth's ecological environment,it makes great sence to study the change of land cover.Remote sensing is an important way for us to obtain the information of the surface coverage.Through remote sensing images,we can obtain the information of the surface at that time,and obtain the knowledge we need through interpretation of remote sensing images.Traditional remote sensing image interpretation methods include visual interpretation,pattern classification and image segmentation.Traditionally,the classification of objects is based on their own spectral,texture and shape information.This classification standard can not solve the problems of "dfferentbody with same spectrum" and mixed pixels.This phenomenon is more obvious in low and medium resolution images.Therefore,how to select and extract deeper and effective features of remote sensing images through classifiers is the focus and hotspot of current research on remote sensing image classification.In recent years,convolutional neural networks have made great achievements in the field of computer vision.Convolutional neural network can automatically extract low features such as spectral shape by using convolution layer and pooling layer,combine low features into high semantic features by using full connection layer,and optimize neuron parameters by back propagation algorithm to obtain a complete model.Compared with traditional classification methods,convolution has stronger feature extraction and expression ability,and higher tolerance rate.Considering these advantages of convolution neural network,this paper introduces convolution neural network model into the classification of medium-resolution remote sensing images,and an object-oriented convolution neural network model(OCNN)is proposed to classify Landsat images.Generally speaking,this paper studies the following points:(1)The input channel of convolutional neural network for high resolution image classification is usually RGB three-band data.Considering the seven-band image of Landsat,this paper introduces the method of remote sensing image processing.Firstly,the original seven-band data of Landsat are transformed by principal component analysis(PCA),and three characteristic bands are obtained.Then,the normalized vegetation index(NDVI),the normalized building index(NDBI),and the normalized water index(NDWI)are calculated respectively,and these bands are combined withthe true color image of Landsat.In order to enrich the feature space of image and explore the impact of different input channels on classification results,band combination is used to obtain input data of multiple channels.(2)At present,the commonly used scene classification models are VGGNet,GoogleNet and ResNet.The size of the input window the model used is 224 x 224.Such a large range of windows obviously can not be used as the input of the medium-resolution image.In order to apply the convolution neural network model to Landsat image classification,the size of the convolution window is reduced.Four different neighborhood windows,21*21,15*15,11*11 and 7*7,are used as input windows to study the effect of different neighborhood windows on the classification accuracy.(3)Object-oriented is a commonly used method for high-resolution image classification.Object-based classification takes into account the spectral,shape and texture characteristics of objects.Firstly,this paper explores the influence of different segmentation parameters on Landsat image segmentation results,and combines object-oriented method with convolution neural network,proposes an object-oriented convolution neural network classification method(OCNN),calculates confusion matrix,and compares with the original convolution neural network classification results,solves the problem of "salt and pepper" and achieves higher classification accuracy.