Object Recognition Method Of Urban Typical Features In High Spatial Resolution Remote Sensing Imagery

The recognition of urban surface elements(coverage)is of great significance for urban planning changes,land use analysis and geographical situation monitoring.However,the rich,detailed and complex data characteristics of High Spatial Resolution Remote-Sensing Imagery(HSRRSI),and the complex structure of the urban surface,pose many challenges for image recognition.At present,a large number of methods for extracting and recognition feature information have been proposed,but the effect of object recognition and recognition of objects from high-resolution images is still poor.At the same time,the increasing demand for surface information acquisition has far surpassed the types and speeds that artificial visual interpretation can provide.Therefore,the research on identifying feature information from high-resolution images has broad theoretical significance and practical value.Based on the deep convolutional neural network and remote sensing image processing technology,this paper establishes a method and technical flow of high-resolution remote sensing urban typical objects recognition based on improved Mask R-CNN target recognition algorithm for three typical types: buildings,water and playgrounds.Firstly,image preprocessing methods including image fusion,framing,linear stretching and filtering are used to enhance the image features of the image,then build an suitable image set for Mask R-CNN algorithm.The features of the feature samples are extracted based on the pre-processed image set to produce training data,verification data,and test data.Secondly,the influence of convolutional neural networks(CNN)with different structures on HSRRSI feature recognition is studied.According to the characteristics of three typical features,a suitable network structure is designed to improve the Mask R-CNN algorithm,and the improved Mask R-CNN is used to extracts the depth features of HSRRSI from the training data,and trains and generates the feature recognition model.For the evaluation of the performance of the recognition algorithm,this paper uses three representative object-oriented classification methods in eCognition: decision tree algorithm,K nearest neighbor algorithm,the random forest algorithm.Those results are compared with the improved Mask R-CNN algorithm.Finally,the HSRRSI data set of different time series is tested using the trained recognition model to verify the generalization ability of the improved Mask R-CNN algorithm.The experimental results show that:(1)The improved Mask R-CNN algorithm is better than the decision tree,K nearest neighbor and random forest algorithm.The improved Mask RCNN algorithm has the highest recognition accuracy for the three types of features.The accuracy rate is 0.9133 and the recall rate is 0.9238,and the integrity of the feature recognition is better.The individual object objects can be segmented.(2)CNN with different structures has an influence on the identification of different types of features and shows certain regularity.For complex feature categories(such as buildings)with obvious spectral,texture,and shape heterogeneity and no fixed spatial patterns,it is more beneficial to improve the recognition accuracy by focusing on the network to extract the deep features of complex features.The heterogeneous features of the features(such as water bodies,playgrounds),focus on the network to extract the effective features of the middle and high-level,the recognition accuracy is better.(3)The training model generated by the improved Mask R-CNN algorithm is further tested on different data sets with different time series and obvious spectral changes in the same region,and the overall accuracy is 0.8794,0.896.The recall rate has a good recognition effect.