Some Key Technology Research On Processing Of High Resolution Remote Sensing Image

With the development of remote sensing technology,the resolutions of remote sensing images have become higher and higher,and the obtained remote sensing datasets have become larger and larger.In high resolution remote sensing images,the geography objects are distributed interlacedly,the categories of geography objects are diverse and the details of them are more abundant,i.e.,colors,textures,shapes and edges,and so on.Therefore,there always exist three issues in the processing of remote sensing images: 1)Different categories of geography objects have different characteristics,so how to find the optimal features or their combinations for describing a particular category of geography object;2)With limit computing and storage resources,how to cluster the dynamic datasets or large datasets;3)The objects in remote sensing images have different sizes and directions,how to overcome the problems of large calculation amount and the slow speed when using sliding window method to detect the objects.According to the above three kinds of issues,this dissertation mainly focus on the study of some key technologies,i.e.,feature description analysis of typical geography objects of remote sensing images,the supervised clustering of data stream of chunks,fast object detection of remote sensing images,and the following innovative research work has been carried out:1.To address the feature description of typical geography objects of remote sensing images,an optimal feature selection and analysis framework based on multiple kernel learning is developed to find the optimal features and their combination for the description of objects of high resolution remote sensing images.In the proposed method,a total of 26 feature descriptors which belong to 16 kinds of features of 4 categories,i.e.,texture feature,color feature,local feature and global structure feature,are mapped into high dimensional feature space by kernel function,and the feature weight matrix is trained by using the multiple kernel learning algorithm.Then,the feature vectors of some typical geography objects are extracted from the feature weight matrix,and the feature vectors are ranked in order of descending weights,thereby the optimal features combination suitable to describe typical geography objects are found out.In addition,a comprehensive ranking method is also proposed to compute sort order of all the features on all geography objects categories.The experimental results on land-use dataset and 19 classes remote sensing dataset show that the optimal features combination are different for different types of geography objects,and using the multi kernel learning algorithm to configurate the features to describe images are superior to using one single feature to describe images.2.To cluster the dynamic or large datasets,a Supervised Adaptive Incremental Clustering(SAIC)algorithm is proposed to cluster the data streams of chunks.Unlike the supervised clustering for static datasets,the supervised clustering for data streams should have the incremental learning ability to process dynamic data,and can avoid or reduce the “local bias of input order” problem,and so on.SAIC includes both learning and post-processing phases.In the learning phase,a counter for each cluster is set to record the winning times of the cluster in the learning process,and a learning rate based on the winning times is calculated to realize the adaptive update of this cluster.The data points are shuffled and learned for multi passes to improve the robustness of clustering results.Whether the learning is terminated or not is determined by comparing the number of clusters before and after each learning round,and then the number of iterations is determinated automatically.In the post-processing phase,the clusters sizes are calculated by using the counter values and the number of iterations,and using clusters sizes to eliminate outliers or boundary points.Experimental results on 4 synthetic datasets,14 UCI datasets and 1 VHR sensing image dataset show that SAIC reaches to or outperforms some the other supervised clustering algorithms and unsupervised incremental clustering algorithms,and has the scalability and incremental learning ability for the clustering of data streams of chunks.3.For detecting the objects of remote sensing image,a Fast Selective Search approach is proposed for specific object detection in VHR remote sensing smages.In the proposed FSS approach,the edge detection results and the corner detection results are used to create corner metric matrix and then binarized.Then the connected components can be labeled.The labeled regions are selected and merged according to their bounding box area,rectangularity,the number of corners,and etc.Consequently some candidate windows or sub-regions which may contain the target object are pre-selected.Finally,these pre-selected windows are detected by the trained detectors.Experimental results on plane dataset and boat dataset show that,the proposed method can detect all kinds of aircrafts and vessels effectively and efficiently,compared with state-of-the-art SS(Selective Search)approach,the detection speed is significantly improved,and the false detection rate is also greatly reduced.