OBIA Classification Of Remote Sensing Image Based On Convolutional Neural Network

Currently,with the rapid development of remote sensing imagery technique,various high spatial resolution remote sensing images can be obtained conveniently with high efficiency.The improvement of spatial resolution enriches the details and features inside the remote sensing images,however,presents new challenges for classification tasks.Traditional classification methods,which use low-level features,have failed to gain satisfactory classification results.How to extract high-level features of high spatial resolution remote sensing images is the key to improve the accuracy of classification.Some scholars proposed the Per-pixel classification method based on convolutional neural network(Per-pixel CNN),which achieved higher accuracy with the help of high-level features.However,using pixel as the basic analysis unit still has weakness and limitation in practical applications.Therefore,proposing a novel theory to combine convolutional neural network with the classification task in remote sensing field has significant meaning.Considering the characteristics of high spatial resoution remote sensing image,this paper analyses the difficult points of the classification task of high spatial remote sensing image.Pros and cons of existing convolutional neural network based classification method in the remote sensing field has been surveyed and analyzed.Based on the theory of Objected-based Image Analysis(OBIA),a novel OBIA classification framework based on convolutional neural network(CNN)has been built.By using segmentation objects as the basic analysis unit,and extracting objects' high-level features by CNN,this paper proposes the superpixel classification method based on single scale convolutional neural network(Per-superpixel SCNN).Besides,in order to overcome the flaws caused by scale effect,superpixel classification method based on multi scale convolutional neural network(Per-superpixel MCNN)has also been proposed.Experimental results from urban and suburban areas illustrates the applicability of proposed new methods,which have three key advantages:(1)Overcoming the shortcomings of Per-pixel CNN methods.Effectively avoiding the salt-pepper effect caused by mixed pixels and greatly reducing computation quantity by using segmentation object as basic analysis unit.(2)Dramatically improving the classification accuracy by using convolutional neural network to extract and utilize high-level features.(3)Solving the problem that multi scale features of remotesensing image are hard to extract.Further improving the classification accuracy.This paper bridge the gap between OBIA and CNN,combining CNN technique with remote sensing classification from new angle,which has great significance in theory and provide important guidance to future studies.