Study On Automatic Classification Method For Remotely Sensed Imagery By Incorporating Spatial-Spectral Features

Classification of image is one of the core content in remotely sensed image processing and important basis of many applications like land cover classification or investigation of resources and environment. With the development of the remote sensing technology and deepening of industrial application, the contradiction between the process automation and result accuracy of classification is becoming increasingly prominent, and how to reasonably solve this contradiction has gradually become one of the main objectives in the study of various classification methods.As almost all the current classification methods need require manually guidance, it's hard to adapt such application requirement as large volumes, quantitative, et al. We proposed the concept of automatic classification which is entirely without manual operation. By combining spatial-spectral coupled cognitive theory and methods of pattern recognition, and refining the remote sensing classification process into feature learning and pattern learning the premise of remote sensing images, preliminarily realized the automatic classification, we realized the automatic classification program preliminarily under the precondition of keeping the accuracy of classification. The main research work and achievement are as follower:1. Based on the study of existing classification methods, we established the system of automatic classification, which include per-pixel and per-parcel methods to correspond data with different spatial resolution.2. Automatic samples selection, the key process of automatic classification, has been realized with improved accuracy of spectral and spatial feature extraction method.3. Fuzzy technology is used to transform the SVM classification method, which combines samples with membership so it's possible to realize quantitative control in the process of pattern learning. As a result, the accuracy of classification is improved.4. The iteratively classification model is established to incorporate spatial and spectral features, and with this model it's easier to build a knowledge integration system to optimize the eventually classification results.5. According to the above automatic classification algorithm, a prototype system is developed which has been applied in the land cover classification in different regions, and good results were obtained.Although the concept of the automatic classification has been proposed and has made a lot of achievements in application, but there are still many deficiencies in current study. For example, current classification methods often don't consider the influence from scale on many image features, classify types and so on. And the science and adaptability of priori knowledge's conjunction need to improve. Furthermore, the iteratively classification model also can be improved through combining multiple classifiers.