Research For Remote Sensing Classification Of Convolution Neural Network Based On Region Information

The fast improvement in remote sensing technology and related fields makes it easy to obtain high resolution images.How to extract features and classify image data quickly and effectively is an important problem in remote sensing image research.The high resolution of remote sensing data makes the pixel-based classification method no longer applicable,and the object-oriented methods become the most important method in high resolution image classification.Deep learning shows excellent performance in image recognition and natural language processing.This paper presents a learning method which is based on convolution neural network to classify remote sensing images.Aiming at solving the problem of low efficiency caused by too many sampling windows in remote sensing image classification,this paper proposes a novel method based on region and convolution to improve the classification efficiency.There are two steps in the classification process.The first step of the proposed algorithm is to classify the input images by convolution neural networks.Then,S VM is used to classify the results of convolution neural networks.The proposed algorithm is evaluated by being compared with the results of other existing methods,experimental results show that proposed algorithm can achieve higher classification accuracy.The fully convolution network can directly output the classification map which has the same size of input image.Remote sensing data,unlike other natural image data,contain much richer edge information and area information.In this paper,we draw on the idea of HED(Holistically-Nested Edge Detection),and introduce deep supervision to the fully convolution network to realize the use of region information in remote sensing images,the features in network should response to the boundary.And we also propose a region-based loss function.The proposed algorithm shows that the model which integrated edge information and region information outperforms the base model.