Research On The Application Of Deep Learning In Remote Sensing Image Target Detection And Land Cover Classification

With the development of aerospace technology and sensor technology,remote sensing technology is entering a new stage that can quickly and accurately provide a variety of earth observation data and extensive application research.The demand for high resolution remote sensing images is increasing,High-resolution remote sensing image application of the breadth and accuracy of data processing put forward higher requirements.Target detection and land cover classification as the remote sensing data extraction of the two key tasks,is the basis of remote sensing applications.However,the diversification of remote sensing data sources and the deepening and integrated development of remote sensing applications have brought unprecedented challenges to both tasks,and the prior art theory requires a big breakthrough to meet the needs of development.The accuracy and efficiency of current high-resolution remote sensing image target detection and land cover classification method is low and.Traditional target detection methods can not provide a good solution when faced with complex environments and multi-source data.The common land cover classification method proposes an object-oriented classification method for high-resolution image synthesis classification,but the technology is not mature and the classification result can not meet the needs of high accuracy.These problems lead to high-resolution remote sensing image application process blocked,its advantages can not be fully reflected.In order to improve the application value of high-resolution remote sensing image,this paper has carried out the research of target detection and land cover classification based on deep learning.The key factor that influences the accuracy and efficiency of the target detection is the construction of the feature extraction and the detection model,that is,the depth of learning and the accuracy of the feature matching.In order to improve the accuracy and efficiency of the target detection,the text uses the depth learning method to construct the target detection architecture based on the two-level cascade convolution depth network.The first level is used to filter the scene and the second level is used to detect the target.By using the transfer learning method,the training time is shortened effectively,and the geometric features of the target in the remote sensing image are used to improve the false detection rate on the basis of improving the detection efficiency.The difficulty of high-resolution remote sensing image classification is the need to consider object-oriented classification methods,rather than pixel-level classification.In order to solve the problem of serious commission and omission in high-resolution remote sensing image classification,an object-oriented semantic class classification method based on deep learning is designed in this paper.Based on the fully convolutional neural network the classification framework is constructed,and the transfer learning method is used to optimize the classification model with a small number of samples to realize high accuracy object-oriented classification.In addition,aiming at the problem of low accuracy of single-polarized SAR image classification,the method of image fusion is used to increase the spectral information and improve the classification information.At the same time,the conditional random field method is introduced,and the classification of the classified boundary is optimized,and the classification result is obtained.The main work of the paper can be summarized as follows:(1)This paper summarizes the research status and development trend of remote sensing image target detection and land cover classification.Combining the problems existing in traditional research methods and the powerful driving force of deep learning in image understanding,this paper puts forward the use of deep learning method for high resolution remote sensing image target Detection and surface coverage classification of the program.(2)The principle of deep learning is described in detail,and the principles,methods and models are introduced,which provides the basis and technical basis for remote sensing image target detection and optical remote sensing image land cover classification,SAR image classification.Which focuses on the CNN network image classification function,Fast R-CNN and Faster R-CNN target detection function,and FCN image segmentation function.(3)This paper analyzes the existing problems of the existing target detection method and puts forward the drawbacks of using the existing deep learning model to carry out the target detection.Aiming at these problems,a rapid detection method of rapid large area remote sensing image based on deep learning is designed.Based on the principle of migration learning,a sample library of high-resolution optical remote sensing image aircraft is established,and a high accuracy remote sensing image target detection model is trained with a small amount of sample data.Considering the geometric characteristics of aircraft on remote sensing images,a region proposal generation algorithm based on geometric feature constraints is proposed,which effectively improves the efficiency of region proposal processing.Finally,a fast remote sensing image target detection method based on two cascade convolution neural networks is proposed.This method overcomes the shortcomings of the existing deep learning methods in remote sensing image target detection,and solves the problem that the weak supervisor learning can not extract the sufficient features of the target and improve the detection efficiency and precision.(4)The FCN classification method based on the deep learning framework is proposed for the classification of high spatial resolution optical remote sensing images.Firstly,the commonly used method of remote sensing image supervision is introduced,and the shortcomings of various methods are pointed out.Aiming at the problems existing in object-oriented classification method of high resolution image,a high resolution remote sensing image classification method based on FCN is proposed.Secondly,the overall architecture based on FCN classification method is constructed,and the classification process of the method is introduced in detail.The production of the sample,the training of the model and the setting of the network parameters are introduced.Finally,ZY-3,Worldview2 and Google Earth three groups of high-resolution remote sensing images were selected for classification experiments.(5)Aiming at the problem of imaging mechanism and image characteristics of SAR images,there are completely different forms of expression and understanding of optical remote sensing images.From the perspective of information fusion,SAR and multi-spectral images based on Shearlet transform and PCNN are proposed Fusion method,and effectively improve the spatial information and spectral information retention of the fusion image.The image classification framework and classification method of SAR fusion based on FCN are designed.Through the study of a small number of training samples,a rich deep image feature map is extracted to improve the classification accuracy of unipolar SAR images.The conditional random field model is used to adjust the classification results,which effectively optimize the land boundary.The HH monopolarization images with TerraSAR-X 3 meters resolution were fused and classified with Landsat5-TM multi-spectral images and Worldview2 multispectral images.