Classification Of Remote Sensing Objects Based On Autoencoder And Convolutional Neural Network

Remote sensing classification is a necessary way to obtain useful data and information from remote sensing images.Especially with the continuous improvement of high-resolution remote sensing technology,the traditional pixel-based classification method has been unable to meet the current classification requirements.In this case,object-oriented classification method came into being.At present,the general objectoriented classification methods need to rely on manual feature selection for classification,which not only consumes manpower and material resources,but also has strong blindness and subjectivity.Deep learning is one of the most advanced learning algorithms in the field of artificial intelligence.It provides an effective framework for the automatic extraction of features,which makes it popular in various research fields.In this paper,two classical algorithms of deep learning including autoencoder and convolutional neural network are applied to object-oriented remote sensing classification.However,the two algorithms will face corresponding problems when they are directly used in object-oriented remote sensing classification.This paper analyzes the problems in detail and proposes corresponding solutions,aiming at improving the classification performance of autoencoder and convolutional neural network in object-oriented remote sensing classification and providing an exploratory idea of automatic learning characteristics for object-oriented remote sensing classification.Specific research results are as follows:(1)A hybrid model of stacked autoencoder and BP neural network is designed.In order to solve the problem that autoencoder is difficult to train when it is applied to object-oriented remote sensing classification because of the high input data dimension caused by the object composed of many pixels as the basic classification unit,a hybrid model of stacked autoencoder and BP neural network is proposed.Firstly,unsupervised feature extraction is performed on unlabeled samples using stacked autoencoder with locally sampled small samples,and weights are pre-trained at the same time.Then,labeled samples are fine-tuned by BP neural network,and the classification results of small samples are integrated by the idea of ensemble learning.The experimental results show that the classification accuracy of the model can reach 0.930,which is an effective and feasible classification method.(2)The multiscale convolutional neural network based on multiple classifiers is completed.Due to the complex structure of the convolutional neural network,which requires a lot of labeled sample training and the difficulty of acquiring labeled sample data due to the diversity of remote sensing data,the idea of multiple classifiers integration is proposed to improve the classification performance of convolution neural network.That is,the output from the network is improved,the classification task is decomposed,and then the improved Bayesian voting method is used for integration.The experimental results show that the classification accuracy can reach 0.916,which provides a new approach for object-oriented remote sensing classification.(3)A hybrid model based on convolutional auto-encoder and convolutional neural network is realized.In order to solve the problem of convolution neural network in classification,this paper proposes to improve the input of the network.The convolutional auto-encoder supports feature extraction and data dimension reduction of remote sensing data.The extracted features are input into the convolutional neural network and subsequently classified.Experimental results show that in the proposed model makes the classification accuracy reach 0.944,and the classification efficiency is improved.The number of labelled samples can be reduced by more than half,all while ensuring a classification accuracy of no less than 0.9,which suggests the effectiveness and feasibility of the proposed model.