Land-use Mapping Method Based On Multi-scale Learning And Deep Convolution Neural Network For High Resolution Remote Sensing Images

Land use data is the important fundamental information for national management on land resources,along with the fast development of economy and the rapid evolution of urbanization in China,land resource utilization problem has emerged and land resources investigation has great realistic significance.Land use survey needs a wide range and large amount of information,remote sensing data which has the advantages of rapid,accurate and convenient access,has been the most efficient data source of land use survey.Since the 1990 s,the high resolution(HR)remote sensing technology has developed rapidly,and higher spatial resolution brought more abundant spatial information,geometry feature,texture information and other details informations.At the same time of promoting land use survey,the complexity of HR image also brings automation land use mapping method to unprecedented challenges.Traditional visual interpretation and pixel-oriented classification method are no longer suitable for HR remote sensing image.The object-oriented classification which has been widely applied at present also have some problems such as low utilization degree of spatial features and limited feature selection,cannot adapt to the increasing demand of land use mapping..In recent years,the Deep learning developing rapidly in the research field of computer vision and pattern recognition.The Deep Convolutional Neural Network(DCNN)has strong ability of image feature abstraction and automatic feature extraction,which has become a very effective method in image processing.In this background,this paper introduced the DCNN into land use mapping after summarizes domestic and foreign related research achievements of theory and application.The research work has the following four aspects:(1)Through the analysis of HR remote sensing image features and the characteristics of land use information,found the problems in implementation of land use in HR image scene classification.By combing with DCNN's receptive field in feature extraction,put forward the learning of multi-scale features is the crucial point in application of DCNN.(2)Studying the fully convolution structure and residual network structure,proposed the Resnet-FCN network based on the encoder and decoder architecture.Designed 100 layer encoder network based on the residual module,integrated feature maps of different scale by skip architecture,and designed an end-to-end dense prediction network with deconvolution.(3)Proposed the MSNet model by further integration of multi-scale learning.Designed several parallel stream with input images of different scales,and use different dilated convolution after the last feature map in network.This architecture could extracts multi-scale features and effectivity improve the classification performance.(4)Experiments were conducted on the optical aerial remote sensing images with 0.5m resolution acquired in Zhejiang province,the training and test data sets consists of 1170 image with manual labeled.Verified the effectiveness of the proposed two DCNN network respectively,and compared with FCN network and the object-oriented method based on SVM.Through experimental analysis,the proposed DCNN based on multi-scale learning has a higher precision of calssification and can get more integrity scene.Compared with the traditional method,proposed network in this paper has improved a lot,and is more suitable for land mapping in remote sensing big data.