Research On Classification Of High-resolution Remote Sensing Image Combining Object-oriented And Deep Learning

Feature classification and recognition is one of the key research directions in the field of remote sensing.With the improvement of the spatial resolution of remote sensing images,the details of feature information are getting richer and richer.The rich details information make the differences in features of the features intensify and reduce the differences between the categories..In this case,the traditional pixel-level classification method can no longer meet the actual application requirements.At present,the focus of related research has changed from pixel-level classification to object-level classification,overcoming some of the difficulties in high-resolution remote sensing image classification and recognition.Among them,image segmentation and feature extraction are two important steps in the object-level classification method.There are still many problems to be solved,such as: irregular segmentation shape size,low boundary agreement,and insufficient feature extraction and expression.In response to the above problems,the main research of this paper includes the following two points.(1)In view of the irregular shape of the high-resolution remote sensing image segmentation results and the low boundary fit,this paper proposes a simple linear iterative clustering superpixel segmentation algorithm(RIULBP-SLIC)that fuse local texture features.First,bilateral filtering is used to denoise the image;the rotation-invariant equivalent local binary pattern coding value is calculated to extract the local texture features of the image;the simple linear iterative clustering algorithm(SLIC)is used to extract pixel color features and spatial location information;Combined with local texture features,color features and spatial location information,similarity measurement rules are re-established,and pixel points are classified into super-pixel blocks.(2)Aiming at the problems of inadequate classification feature extraction of highresolution remote sensing images and limited feature recognition capabilities,this paper proposes an object-oriented multi-scale residual contraction network classification method(MSRSNet).Based on the superpixel block obtained by the above method,the smallest circumscribed rectangle formed by the connection of the neighboring superpixel centers at different angles is selected as the input of the network model;the residual contraction network(RSNet)is used to extract multiple scale features of the object,and the The learning strategy fully trains the network model to achieve accurate classification of high-resolution remote sensing images and overcome the problem of overfitting.The experimental results show that the improved superpixel segmentation algorithm in this paper has higher segmentation accuracy on multiple remote sensing data sets,the resulting superpixel block shape distribution is more regular,and the edge fit is higher.The multi-scale residual contraction network classification model proposed by combining object-oriented theory and transfer learning strategy can effectively improve the feature expression ability,improve the classification effect,overcome the over-fitting problem caused by small samples,and improve the generalization ability of the model.