Research On Remote Sensing Image Classification Based On Small Sample Data

In recent years,with the development of deep learning,the accuracy of image classification in large-scale data sets is improving rapidly.Image classification technology is more and more popular in practical application,providing great convenience for human life and work.But it takes a lot of manpower and financial recourses to obtain large-scale data set while the data set in remote sensing is very small,so the research of image classification based on small sample data is of great significance.But the current remote sensing image classification technology is still inadequate under the condition of small sample data set.The traditional machine learning scheme requires researchers rich experience of image processing and image feature extraction,also the prior knowledge in the related fields,which make the traditional machine learning scheme hard to apply to other fields.In view of shortcoming of traditional machine learning scheme,many researchers have proposed a new scheme which treat the deep learning model as a feature extractor and classify them using support vector machine.This scheme can improve the accuracy of image classification at the expense of the convenience of system design which split feature learning and classification task of deep learning.Compared with pure deep learning,it requires extra work.Therefore,in view of the shortcoming of the current image classification technology in small sample data set,this paper proposes two deep learning model optimization scheme and carry out experimental research on two remote-sensing data set.In view of the fact that deep learning is easy to be over-fitting,this paper proposes a model optimization method which is similar to residual network structure.Unlike residual network which its low-level feature map are added to its top-level feature map,the model optimization this paper proposed will concatenate the low level feature map with its top level feature map.The feature map fusion experiment shows that the top level convolutional layer will be to fit a mapping which trends to zero if the low-level feature map is capable to discriminate remote-sensing data set.So there are only a few feature map will be activated.Meanwhile,it will reduce the possibility of noise that the upper convolutional layer output.And it will alleviate the over-fitting phenomenon and improve accuracy of image classification using the model optimization that this paper proposed.Moreover,in view of the fact that deep learning model has so many parameters which make the model more complex,this paper proposes a model sparsity scheme which similar to singular value decomposition.By compressing the feature map that the upper convolutional layer output,it eases the burden of the current convolutional layer which reduces the number of parameter of the model.And it controls the number of similar or redundant feature map that deep learning model produces using over-complete method.Experiments show that it greatly improves the accuracy in remote-sensing data set in nonfine-tuning condition using model sparsity while it still has a better results than the origin model in the fine-tuning condition if hyper parameter chosen wisely.