Research About Remote Sensing Image Classification Based On Deep Learning

With the development and application of remote sensing technology in practice,related technology products also play a very important role in so many fields.Among them,it is the key factor about how to obtain effective and accurate remote sensing information.And remote sensing image classification is one of the important methods for extracting remote sensing information,and it is also the hot research field in the current research of remote sensing technology.Therefore,it is our main content about how to classify the remote sensing images accurately and effectively in this article.At present,the traditional classification algorithms which are usually used in remote sensing image classification have insufficient classification performance and accuracy to meet our needs in practical.And the traditional pixel-based classification methods are prone to misclassification and missing points,which can also lead to poor classification performance.This thesis proposes a double input convolutional neural network based on object classification.This structure is different from the traditional convolutional neural network.It corresponds to two convolutional neural network inputs.The network model is input with two kinds of data sources respectively.Each of them is classified by a same convolutional neural network,and the predicted classification results of the two input sources are finally summarized.The final predicted classification result is obtained by the argmax method.The input data of the network is the super-pixel image which is cropped from the original image as well as the super-pixel image's mask image.At the same time,this thesis proposes a remote sensing image feature extraction method.Before the convolution through which the features were extracted from the original CNN,we performed the PCA dimension reduction on the input data at the first.And then the feature after dimension reduction is extracted through Gabor filtering.Then,the features are extracted.The filtered feature data is then input to CNN to complete the final extraction and classification training.Experiments show that the improved algorithm and model are superior to the common convolutional neural network model in the prediction and classification accuracy of ground features.In order to verify the effectiveness of the optimized convolutional neural network model and the improved feature extraction method for classifying images.The thesis conducts a classification experiment which was performed on hyperspectral remote sensing images of Heilongjiang Province with high resolution of 16 m.Then the classification effect of the proposed method is analyzed through the four indicators of confusion matrix,overall accuracy,user accuracy and Kappa coefficient.And it was compared with the classification results of traditional image classification algorithms.Experimental results show that the accuracy of the proposed method is better than some existing traditional classification methods.