Object-oriented Classification Based On Deep Features For High Resolution Remotely Sensed Imagery

High-resolution satellite remote sensing imagery can provide abundant surface information for detailed description of the spatial structures and texture of ground objects.Edges of features are clearly defined and provide the conditions and basis for effective interpretation and analysis of ground objects.This research on intelligent interpretation of high-resolution remote sensing data is of high academic values for the development of automated analysis of remote sensing data,transforming data to knowledge,and improving data utilization.It is also practically significant for strengthening the applications of high spatial resolution remote sensing data in environmental monitoring,urban planning,military investigation,precision agriculture,and smart cities.Although high-resolution remote sensing data have shown significant advantages and good applications,there is still a large gap between the processing level and actual demands of high-resolution data.There is a serious imbalance between remote sensing information processing techniques and data acquisition capabilities.Remote sensing image classification is one of the most basic issues in remote sensing image processing.Object-oriented image analysis based on multi-scale segmentation is an effective method for remote sensing image interpretation,especially for high spatial resolution remote sensing image classification.However,it faces two challenges since its development at the end of the 20 th century.First,remote sensing image segmentation is the key to object-oriented image analysis,and the rich spatial object and spatial semantic information in high spatial resolution remote sensing images should be fully expressed and described at multiple scales.Therefore,how to quantitatively evaluateimage segmentation quality and to determine an optimal segmentation parameter combination are of significance.Second,in traditional object-oriented high-resolution remote sensing image classification methods,feature description,feature learning,and classifier design can be successfully done,however only shallow features can be extracted from original images.For classification,it is insufficient for the deep object feature expression of the object targets.The artificial design and selection of features are yet not conductive to the automatic interpretation and information mining of multisource remote sensing big data.In order to resolve both challenges,this thesis focuses on the quality assessment of unsupervised multi-scale image segmentation and objectbased depth feature automatic learning.The main achievements are as follows:(1)A novel unsupervised evaluation method was proposed in this paper to quantitatively measure the quality of segmentation results.In this method,multiple spectral and spatial features of images are first extracted and then integrated into a feature set to improve the feature representation of ground objects,and two indices are employed to estimate spatial stratified heterogeneity and spatial autocorrelation of segments,and then combined into a global assessment index as the final quality score.The trade-offs of the combined index are geometrically accounted for using Mahalanobis distance.The proposed method was tested on two current segmentation algorithms and three test images and compared with two existing unsupervised methods and a supervised method.(2)A high-resolution remote sensing image feature learning method based on superpixel feature representation was proposed in this paper,in order to make full use of the rich semantic information of high-resolution remote sensing images.In this method,the SLIC(simple linear iterative clustering)algorithm is applied to the superpixel segmentation of remote sensing imagery to be classified,and the resulting compact and uniform superpixel objects are taken as the input of the network model.After that,training samples are used to train the model parameters of the convolutional neural network,in order to achieve an automatic learning object depth features.This method overcomethe problem of the complex operation and the lack of target background information in the use of pixel neighborhood matrix as a network input and improves the efficiency and depth of feature extraction.(3)An adaptive boundary-constrained multi-scale convolutional neural network classification method was proposed in this paper.Based on a feature learning method and superpixel feature expression,this method uses super-pixel objects at different scales as the network input.A multi-scale convolutional neural network model is constructed to input data of different scales into the corresponding convolution layer.The multi-scale convolutional layer extracts the corresponding feature map according to the scale weights of the ground objects.In this network,feature fusion and classification are carried out to obtain the superpixel classification results.After that,the image segmentation quality evaluation index proposed in this paper is used to yield optimal object boundary information for self-superpixel classification results.The initial classification results are constrained and optimized to obtain the final classification results.