The object detection of remote sensing imagery has lower detection accuracy caused by complexity of background, target appearance variety and arbitrary orientation. In order to solve the problem, a method based on strongly-supervised deformable part models was proposed. Then multiple sub-models in each direction range of the object were trained. In addition, the object bounding rectangle, position and semantic information of every part were labeled. In the model training stage, firstly, multi-scale Histogram of Oriented Gradients (HOG) feature pyramid for every training image was constructed, and the model structure was initialized according to object-part label information and Minimum Span Tree (MST). Secondly, the sub-models corresponding to every direction region were trained using Latent Support Vector Machine (LSVM). Every sub-model was consisted of a object filter, multiple twice resolution part filters, and the position relation model. Finally, the mixture model was merged from all sub-models to detect object. In the object detection stage, the multi-scale feature pyramid was also firstly constructed, then matching response score in feature pyramid was computed using mixture training filter model and sliding window. Optimized detection results could be obtained by setting threshold for the response score and adopting Non-Maximum Suppression (NMS) algorithm. The object detection accuracy of the proposed method is 89.4% on self-built remote sensing data sets, compared to the highest accuracy among weakly-supervised Deformable Part Model (DPM), Exemplar Support Vector Machines (Exemplar-SVMs) and Histogram of Oriented Gradients-Support Vector Machine (HOG-SVM), the proposed algorithm has an improvement about 4 percentage points in detection behavior. The experimental results show that the proposed algorithm could has some improvement in solving above mentioned direction and background complex problems, and also can be applied in object detection of the airport military airplane.

Since the accuracy of single feature in scene classification is low, inspired by information fusion, a classification method combined feature-level and decision-level fusion was proposed. Firstly, Scale Invariant Feature Transform-Bag of Words (SIFT-BoW), Gist, Local Binary Patterns (LBP), Laws texture and color histogram features of image were extracted. Then, the classification results of every single feature were fused in the way of Dezert-Smarandache Theory (DSmT) to obtain the decision-level fusion result; at the same time, the five features were serially connected to generate a new feature, the new feature was used to classification to obtain the feature-level fusion result. Finally, the feature-level and decision-level fusion results were adaptively fused to finish classification. To solve the Basic Belief Assignment (BBA) problem of DSmT, a method based on posterior probability matrix was proposed. The accuracy of the proposed method on 21 classes of remote sensing images is 88.61% when training and testing samples are both 50, which is 12.27 percentage points higher than the highest accuracy of single feature. The accuracy of proposed method is also higher than that of the feature-level fusion serial connection or DSmT reasoning decision-level fusion.

For the classification of high-resolution remote sensing images, inspired by human vision system which extracts summary statistics information for scene perception, a feature extraction method based on summary features was proposed. In the method average orientation information and visual clutter were extracted and combined to form a representation based on summary statistics, in which average orientation information was summarized by using Gabor filters and visual clutter was measured based on visual crowding.The experimental results on the classification of 21 classes of remote sensing image set reveal that the classification accuracy of the proposed method is 6.5% higher than Gist and 3.22% higher than Bag-Of-Words (BOW), when the number of training images and testing images are both 50. It also has lower calculation burden. While compared with Gist, the proposed method doesn't need any human intervention.

Object-oriented analysis of polarimetric Synthetic Aperture Radar (SAR) has been used commonly, while the polarimetric decomposition is still based on pixel, which is inefficient to extract polarimetric information. A object-based method was proposed for polarimetric decomposition. The coherent matrix of object was constructed by weighted iteration of scattering coefficient of similarity, and the convergence of coherent matrix was analyzed, therefore polarimetric information could be obtained through the coherent matrix of object instead of pixel， which can improve the efficiency of obtaining polarimetric features. To more fully reflect the terrain target, spatial features of object were extracted. After feature selection, polarimetric SAR image classification experiments using Support Vector Machine (SVM) demonstrate the effectiveness of the proposed method.