Reaearch On Remote Sensing Image Classification Based On Deep Learning

Remote sensing images contain abundant information of ground objects,and it is the most intuitive manifestation of the whole feature of the ground objects.The feature extraction of remote sensing image is the most basic operation for studying target objects and acquiring information.The classification and recognition of the remote sensing image through the extracted remote sensing image feature information is a further study of obtaining the information of target objects,which is of great significance to urban planning,disaster monitoring and ecological environment evaluation.The commonly used methods of remote sensing image classification and recognition mainly include artificial visual interpretation,supervised classification and unsupervised classification,expert system classification,neural network classification and so on.These methods can solve some classification problems.However,with the development of science and technology,high resolution remote sensing images and hyperspectral remote sensing images appear successively.The spatial resolution of the remote sensing images is more higher and the number of spectral channels is more.Also the dimensions of information are higher.Therefore,the information contained in remote sensing images is richer,and the difficulty of classification and recognition is also increasing.Deep learning is a machine learning method that emerged in recent years.It can automatically learn the intrinsic characteristics and laws of a large amount of historical data,so as to identify,judge new data or make a forecast for the future.To some extent,it also can improve the accuracy of classification and recognition.Therefore,it is of great significance for the feature extraction and classification research of remote sensing images through deep learning methods.Although deep learning has been applied in the field of remote sensing image classification and has achieved good results,there are also some problems at the same time,such as easy to fall into local optimum and manpower and resources consuming leaded by manual debugging parameter.In addition,when classified some datasets in deep learning method in previous studies,they usually use the model of convolutional neural network and seldom use other deep learning models,such as the model of auto-encoder or its variant.This makes the classification performance of this type of deep learning model on this datasets unclear.Taking Aerial Image Dataset(AID)and UC Merced_LandUse as the study of data and Matlab(R2016a)as the experimental platform,the paper applies the method of deep learning to classification of remote sensing images,and studies the problems of local optimum in the classification process,manual debugging of parameters,the classification performance of auto-encoder's variant model on UC Merced_LandUse datasets and the improvement of classification accuracy.Firstly,the paper proposes an adaptive parameter(learning rate)setting method by constructing the stacked denoising auto-encoder model with simulated annealing to avoid local optimum and achieve the goals of automatically adjusting learning rate parameter.Secondly,the paper first tests the classification effect of the stacked denoising auto-encoder model that is the variant of auto-encoder model on the UC Merced_LandUse dataset,and proposes a multi-feature stacked denoising auto-encoder method to improve the classification accuracy.The main findings are as follows:(1)Aiming at the problems of easily trapping into local optimal solution and manpower and resources consuming,slow convergence,low efficiency leaded by manual debugging parameter,the paper proposed a learning rate adaptive strategy with simulated annealing.By giving a certain proportion to the learning rate parameter,the model can adjust the learning rate according to the specific situation of learning process so as to acquire better results.In the case of setting optimal parameters and hidden layer nodes,a stacked denoising auto-encoder model with this strategy is constructed to classify images.The experimental results show that this strategy can effectively avoid local optimum and solve the problems brought by manual adjustment of parameters as it not only saves time,manpower and material resources,but also improves the classification accuracy.(2)Aiming at classifying the UC Merced_LandUse dataset through the method of deep learning,we are more likely to use the convolutional neural network than auto-encoder or its variant models in the previous studies.This paper selected the stacked denoising auto-encoder,which is one of the variant model of auto-encoder,to classify the UC Merced_LandUse dataset and verified its applicability;and proposed a multi-feature stacked denoising auto-encoder classification method,which extracts and integrates the feature of color,texture,shape and scale invariant feature transformation of UC Merced_LandUse images as the criteria of studying and classifying,to improve the accuracy.The experimental results show that the method can improve the classification accuracy,and the stacked denoising auto-encoder model based on this method also has certain classification ability for the UC Merced_LandUse dataset.