The Visual Representation Of High Resolution Remote Sensing Imagery And Its Application In Classification

From the perspective of philosophy, object recognition is to recognize the essence through the appearance. The object classification, recognition and extraction of remote sensing imagery are all conducted based on the object features, thus feature plays a very important role in object interpretation of remote sensing. However, the high resolution remote sensing imagery usually cannot present the feature of object directly. It needs analysis, mining, transformation, mathematical modelling and so on to obtain effective features. Thus, the feature representation of high resolution remote sensing imagery becomes an important research.In this paper, the overall strategy of visual representation of high resolution remote sensing image is discussed. Two research routes of visual bionic and expert knowledge transfer are proposed. The main contents are as follows:1) This paper proposed a visual saliency representation algorithm for high-resolution remote sensing images by improve Itti model, and presents a set of visual representations of sporadic vegetation suitable for both RGB images and near infrared images. It solves the problem of feature extraction of sporadic vegetation because of its small area, fuzzy and mixed pixels. Finally, a method of automatic extraction of sporadic vegetation is proposed using Fuzzy ART. Experiments on RGB images and near infrared images show that the accuracy of sporadic vegetation extraction is more than 90%.2) This paper proposed a Multi-scale and Multi-direction Structure Index (MMI). Compared with traditional representation algorithms of structural texture, which emphasizes the regularity of texture and the consistency of scales and directions, the MMI can be suitable for near-regular textures and more complex textures with multi-scale and multi-direction. The application of MMI in the classification of high-resolution remote sensing images is also studied, and an automatic plantation extraction algorithm based on MMI is proposed. It can solve the problems that traditional plantation extraction algorithms are usually less suitable and has high data dependency. And several MMI-based automatic classification algorithms for artificial objects are proposed to further improve the refinement of remote sensing information extraction. Finally, the main parameters involved in the MMI are experimented and analyzed.3) This paper proposed a Boundary Parallel-like Index (BPI). Compared with traditional shape features and texture features by several experiments, it is proved that BPI is more effective in characterizing shape information, especially for building shape characterization. Compared with traditional shape features and building morphological index, gray level co-occurrence matrix and wavelet texture, it can obtain higher precision for building extraction. Therefore, combining BPI with spectral features and many texture features, a high-precision automatic method of building extraction is proposed by constructing multi-layer classification knowledge network.4) Based on object-oriented idea, this paper solves the basic unit problem of deep learning applied to the representation and classification of remote sensing images. Based on this, the self-representation of high resolution remote sensing image and its application in classification are studied. An algorithm of automatic classification of remote sensing imagery based on CNN is proposed. An algorithm of remote sensing imagery classification combining CNN and SVM is proposed. An algorithm of remote sensing imagery classification combining CNN and RF is proposed. And the effectiveness of these algorithms is tested and compared through related experiments. This paper provides a preliminary solution for the automatic acquisition of remote sensing image features.