Remote Sensing Image Classification Algorithm Based On Convolutional Neural Network

With the development of hyperspectral remote sensing technology,traditional classification algorithms can not meet the classification requirements of hyperspectral remote sensing images in big data environments.How to classify remote sensing images more accurately and rapidly has become a hot issue in the field of remote sensing.With the rapid development of artificial intelligence,the local connectivity,shared weights,pooling and hierarchical structure of convolutional neural network make it to achieve some degree of displacement and scale invariance and deformation.Its powerful learning ability and fault tolerance make it achieve excellent results in many fields such as image processing and natural language processing.In this paper,convolutional neural network model is applied to remote sensing image classification tasks.First of all,we review and sum up the current research results and existing problems of the remote sensing image classification methods at home and abroad,then we analyze not only the related technologies including model structure,working principle,parameter training and optimization but also the advantages and disadvantages of the convolutional neural network,and build a convolutional neural network model which is suitable for remote sensing images classification.Finally,we has done the following three parts based on this model:(1)Aiming at the problem that the single feature can not provide effective information for convolutional neural networks,a multi-source and multi-feature fusion method is proposed.The spectral features,texture features,spatial structure features of remote sensing images are combined in the form of vectors or matrices according to spatial dimensions by this method,and the convolutional neural network model is trained by the fused features.The experiments show that the fusion method can make the model learn more abstract and representative high-level features and improve the classification accuracy effectively.(2)In order to further optimize the structure of convolutional neural network model,a hybrid activation function PReLU-Swish is proposed.This function not only retains the non-linear characteristics of Swish function but also can effectively alleviate the problem of gradient dissipation.The experiments show that PReLU-Swish improves the classification performance of convolutional neural networks to a certain extent.(3)Aiming at the problem that the generalization ability of a single convolutional neural network is hard to be improved,an ensemble convolutional neural network classification model is proposed based on the works of(1)and(2).Firstly,the model uses the bagging algorithm to sample the original sample,and then three convolutional neural networks are trained separately.Finally,the classification results of the three networks is integrated with the weighted voting method.The experiments show that the ensemble convolutional neural network model can further improve the classification accuracy of remote sensing images and get the best classification results compared with other comparison algorithms.