Study On Automatic Recognition Of High Resolution Optic Remotely Sensed Images Based On Visual Features

Remote sensing developed since 1960s is one of the most popular space-based Earth Observation techniques. After over 50 years'development, remote sensing has been a main method for spatial information acquiration and widely used in geology and relative fields. With the improvement in resolution and the intergrated system of space-aerial-ground based remote sensing comeing into being, remote sensing is not only used in traditional fileds, but developes into common service filed of our society. All these changes indicate the great potential of applications and the great market demands of remote sensing.Remote sensing borns for applications and is improved by applications, thus the spatial information should be interpreted from the images at first. For a long time, the interpretation mainly depends on visual interpretation. The visual interpretation can provides more reliable results as it integrates human intelligence and knowledge. However, this method is time-consuming, laborious and expensive, and also lower accuracy in quantitive analysis. The high resolution leads to data increase greatly and market demands increase greatly. Face of mass data, it is obvious that visual interpretation does not meet the requirements of application. But it is unfortunately that the state-of-the-art automatic interpretation technique is still being studied and is far from practical application. Under such conditions, to develop automatic and intelligent intergretation technique is much necessary, and now it is a popular and hot study in remote sensing. This problem is also the main concern of our study.As different image type often has different iamge properties, this study focuses on high resolution optical remotely sensed images. The main purpose of the study is to try to explore the general model and common guideline of automatic interpretation of remotely sensed image. And under this technique framework, we try to develop some models for serval popular ground objects'extraction.The main innovations in the dissertation as follows:1. We develop a layer-separated interpretation model for gereral model of automatic interpretation. The main idea of the model is thought that remotely sensed images is superimposed result of different object layers and the course of interpretation is to devide the image into different layers and measure the objets properties. To different object types, different features should be used in modeling. And the order should be from simple object type to complicatied objects, layer by layer, and reach to interpret all objects in the end. For fearture selection, we also propose a rule base on visual recognition feature and its cost and realizable in automatic extraction.2. We propose an original and image depend water body extraction method for automatic water body extraction from remotely sensed images. The model is selected visual recognition features for modeling and dose not require any other data also NIR image. In order to protect the properties of small samples, we develop an automic compute algorithm for clusters numner and thresholds which is based on minimum summation of distance. Experimental results indicate that the extracted results using this model is much fine in agreement with visual interpretion results.3. We have studied the color characteristics of shadow in remote sensing and have deeply studied the characteris of hue, saturation and intensity of shadow. Based on these studied results, we propose an orgininal multi-feature intergration (MFI) shadow detection method. Experiments indicate that the shadow detection result using this method is close to that of visual interpretation in general, and is better in applicability.4. We have studied the visual recognition features of shadowed regions. Combining with the study results of shadow color, we also propose another original shadow detect method---Self-Adaptive Feature Seletion method (SAFS). The SAFS model selects intensity and six bool relationships in red, green and blue components as modeling features. Firstly the intensity is translated for improving separability of shadowed area and non-shadowed area. And then use PM algorithm to make decision. The marked characteritc of SAFS model is that SAFS can automaticly select proper feature to build the model for shadow detection according the properties of image and can be used for both grey and color image. Experimental results indicate that the shadow detection results using SAFS models are much fine and agree with visual interpretion results. In general, the SAFS model provides a practical method for application.5. We propose an automatic vegetation extraction method from high resolution image which is built on visual recognition features and visual recognition results. The model selects hue, saturation and intensity as three modeling features, and use HSI transformation to obtain hue, saturation and intensity images. Then use the special property in hue of vegetation to coarsely segement the hue image. After that, the new hue image together with saturation and intensity images is translated by Gauss functions for improving separability between shadowed area and non-shadowed area. Experimental results indicate that the automatically extracted results using this model are much agreement to visual interpretation results. And the mothod provides an applicable method for larger scale vegetation data collection.6. We develop Injected Conversely Filling method (ICF) for completely automatic seed filling method. The ICF method can be used to any complicated regions and successfully resolve the problem of automatic selection of seed points.